VENDORS | eDiscovery Team

Why Quantum Law, and Why Now?

June 17, 2026

Ralph Losey, June 17, 2026.

Privacy, Proof, and Judgment
in the Next Technology Shift to Quantum

A man in formal attire stands in front of a grand building labeled 'Justice - Reason - Evidence', facing a futuristic scene featuring digital elements and a quantum device, with the inscription 'Q DAY WILL CHANGE PRIVACY. PROOF. JUDGMENT.'

Most legal technology shifts do not announce themselves with trumpets. They slip into ordinary practice first. Email was once informal chatter, then became the core of discovery. Predictive coding was once treated as radical, then became accepted. Generative AI was once a toy for clever prompts, then became a sanctions risk, a productivity tool, and an ethics problem all at once.

Quantum computing may follow the same pattern, but it could also arrive with a loud bang and deeper consequences. The timing is uncertain. Some forecasts are aggressive; others stretch out for many years. But when cryptographically relevant quantum computing arrives, the legal effects will be hard to ignore. It will affect how long secrets stay secret, how courts evaluate machine-generated proof, and how lawyers exercise judgment when powerful systems produce answers no ordinary computer can easily check. That is why quantum law deserves attention now, before the first crisis forces good lawyers to fumble, struggle, and learn in public.

A visual representation of legal themes in a futuristic setting, featuring an hourglass, a quantum computer, and a map labeled 'The Legal Terrain Ahead', highlighting concepts like Cryptography, Privacy, Liability, and Post-Quantum Security.

Q-Day, Quantum Proof, and the Legal Terrain Ahead

This article is a clarion call and a first outline of the terrain ahead. It is designed for everyone who deals with confidential data, evidence, or dispute resolution. Quantum computing law will arrive through inventions built by corporate, university, and government teams of engineers and scientists, increasingly assisted by powerful AI systems. Some forecast its arrival next year; others stretch it out for many years.

We may not know the date when the quantum computer future becomes practical, but we do know that its arrival could prove to be very disruptive. Some semi-quantum segments have already begun to seep into legal practice through vendors, simulations, and hybrid AI systems. Full quantum computing may arrive suddenly, especially if cryptographically relevant quantum computing makes old encryption vulnerable. The legal risk does not depend on guessing the exact date. It depends on whether lawyers understand the terrain before clients, courts, and vendors begin asking questions they are not prepared to answer.

The best-known danger has a name: Q-Day. That is the day when a sufficiently powerful quantum computer is built that can break most of the public-key encryption now used to protect digital information. No one knows when Q-Day will arrive, but the danger begins before that day because encrypted data can be stolen now and stored for later. We know that is already happening. If the data is still valuable when quantum decryption becomes practical, today’s secure archive will become tomorrow’s open file cabinet. That is the “harvest now, decrypt later” problem, and it gives confidentiality an expiration-date problem lawyers cannot ignore. See NIST, What Is Post-Quantum Cryptography? and the companion paper, Post-Quantum Cryptography:

Q-Day is only part of the challenge. Quantum computing may also reshape how courts apply evidentiary standards and evaluate reliability, while creating new questions involving cryptography, privacy, liability, and insurance. See RAND, The Quantum Age and Its Impacts on the Civil Justice System (4/29/25).

One of the most important evidentiary shifts may be from Identity to Fidelity. Lawyers are accustomed to machines that produce the same answer every time. That is identity. Quantum systems may force courts to ask a different question: whether the process behaves faithfully within known error limits. That is fidelity. The issue will not be whether every run produces the same output, but whether the pattern of outputs can be explained, tested, and trusted.

That shift will force lawyers to ask practical questions. What was the model asked to do? What assumptions went in? What error rate is known? What was excluded? Can another qualified team test the process well enough to trust it?

This article only traces the outline of how law can prepare now, before Q-Day, and later, when quantum-generated evidence begins appearing in disputes. A fuller map is needed, and I have been working hard on that, but the first step is seeing the terrain.

A glass dome showcasing quantum science applications, including GPS, MRI technology, lasers, and transistors, with a scenic city backdrop and a classic telescope in the foreground.

Quantum Was Already Here, Just Quietly

Many lawyers hear “quantum” and think of science fiction and multiverses. That reaction is understandable. Quantum mechanics is strange, and lawyers are trained to distrust strange things unless they come with affidavits, exhibits, and a billing code.

In fact, quantum technology is not new. NIST’s Andrew Wilson explains that GPS, MRI machines, and laser pointers all depend on quantum science. From GPS to Laser Pointers, Quantum Science Is All Around Us. NIST’s Corey Stambaugh makes the same point in still broader terms. A Quantum Leap Forward: How Tiny Particles Can Bring Us Exciting New Tech.

Quantum computing is different because it uses quantum behavior to process information itself. NIST’s Quantum Computing Explained . The article is a useful starting point for lawyers because it explains the basic difference between classical computer bits and quantum bits, Qubits. The legal point is not the math or entangled superpositions. The point is that a different way of processing information will create new and different legal problems.

A futuristic setting depicting the concept of 'Q-Day', where quantum computers break encryption. It features a lock being illuminated by a blue beam, symbolizing decryption. In the foreground, an hourglass and metal filing cabinets labeled 'Encrypted Confidential Data' and 'Decrypted Access Granted Tomorrow' highlight themes of time and privacy.

Q-Day and the Old File Cabinet Problem

The most immediate legal problem is confidentiality. Imagine a law firm with an old litigation archive from a trade-secret case. The case settled years ago. The files are encrypted and stored in the cloud. The client has moved on. The lawyers have moved on. The archive sits quietly in a digital file cabinet, full of secrets everyone assumes are still safe.

Now ask how long those secrets must remain secret. Some secrets age out quickly. Others remain valuable long after the case is closed.

The legal danger begins before Q-Day because encrypted data can be stolen now and stored for later. The thief does not need to open the cabinet today. He only needs to steal it and wait. Q-Day is not just a future cybersecurity event. It is a present-day confidentiality problem for anyone holding secrets that must remain secret for years. See NIST, What Is Post-Quantum Cryptography?. The unpleasant feature of this risk is that it may mature suddenly. A file that was unreadable yesterday may become readable tomorrow if the lock protecting it was built on vulnerable cryptography.

Password protection is not encryption. A password controls access; encryption protects the contents. Q-Day is not a faster way to guess your great password. The risk pertains to the mathematics behind vulnerable public-key encryption. So, when a vendor responds to quantum-readiness questions by talking only about strong passwords, multi-factor authentication, or access controls, the vendor has not answered the real question. Ask what cryptography protects the data, who controls the keys, and whether there is a post-quantum migration plan.

NIST has already finalized its first three post-quantum cryptography standards to try to protect against this vulnerability: FIPS 203, FIPS 204, and FIPS 205. See NIST, Post-Quantum Cryptography FIPS Approved. See also Federal Register, Announcing Issuance of Federal Information Processing Standards FIPS 203, FIPS 204, and FIPS 205. For lawyers, the lesson is not to become cryptographers but to learn some of the basics, and to recognize that cryptographic migration has already moved from theory to standards.

A group of officials monitoring cybersecurity threats in a high-tech control room, with flags of the USA, China, and Russia in the background, and screens displaying alerts about data breaches and compromised encryption.

The National Security Shadow of Q-Day

The deepest Q-Day risk is not merely that old legal files with client secrets become readable. That is bad enough. The larger danger is strategic. If the first cryptographically relevant quantum breakthrough is achieved secretly by a hostile government, the result could be more than a cybersecurity incident. It could be a shift in military, intelligence, diplomatic, and economic power.

A state actor that can read previously secure communications may not announce the achievement. It may watch, wait, and exploit. It may use old, intercepted traffic to identify sources, compromise negotiations, expose military plans, manipulate markets, pressure companies, or weaken alliances. The first signs may not look like a quantum breakthrough at all. They may look like inexplicable intelligence failures, severe infrastructure disruptions, mass persuasion and social manipulation, followed by financial collapse and social unrest.

That is the nightmare scenario. Not a quantum computer on stage at a press conference, but a quiet advantage used in secret by an unscrupulous power. The problem is not limited to adversaries. If any military-intelligence system reached Q-Day first, the pressure to use that advantage would be immense. History teaches that strategic breakthroughs become instruments of state power long before civilian institutions understand them.

Do not think this will be like the mirage of Y2K. Although Q-Day has no known date, it is very real, may arrive in secret, and threatens power rather than malfunction.

Ben Buchanan and Andrew Imbrie’s important book, The New Fire: War, Peace, and Democracy in the Age of AI, is not a quantum book, but its warnings about advanced technology and state power apply here. Powerful computational tools can support science, medicine, and prosperity, but they can also intensify conflict, surveillance, and authoritarian control. The potential of quantum computing to vastly enhance mass surveillance and authoritarian control is especially worrisome. See my article, Escaping Orwell’s Memory Hole: Why Digital Truth Should Outlast Big Brother (March 2025), which may prove to be over-optimistic.

That is also why the CISA, NSA, and NIST have all urged organizations, especially those supporting critical infrastructure, to begin quantum-readiness planning now. See CISA, NSA, and NIST, Quantum-Readiness: Migration to Post-Quantum Cryptography. The advice is not theoretical. It is a practical checklist we should all follow, now, to begin preparations: (1) identify long-lived sensitive data; (2) build migration plans; and, (3) ask vendors what they are doing. Waiting for proof that Q-Day has arrived may mean waiting too long.

The best hope is that quantum breakthroughs occur in the open, with enough time for post-quantum defenses to be deployed across governments, courts, companies, and critical infrastructure. History offers little assurance that transformative strategic technologies will be introduced so politely.

A private company or university team might provide that warning if it reaches the threshold first and resists premature militarization. But even that hope is fragile. Once a technology can alter the balance of power, governments may smash through laboratory doors.

That is why quantum law is not just about future expert testimony or old encrypted archives. It is also about governance, secrecy, democracy, and stability in a world where computation is already a weapon of state power. Q-Day will make that weapon far more dangerous.

A man in a suit stands in a courthouse, gazing at a digital display with icons representing various aspects of the civil justice system, including privacy, liability, and digital evidence. In the background, a group of legal professionals is seated at a table, engaged in discussion.

The Civil Justice System Is Already on Notice

RAND has already examined the civil justice consequences of quantum computing, at least in a cursory way. Its 2025 report looks at what quantum computing may mean for courts, law firms, insurers, regulators, and related institutions. RAND, The Quantum Age and Its Impacts on the Civil Justice System. RAND is not selling magic crystals. It is telling the legal system to pay attention before the problems arrive fully formed.

The legal profession has been late before. It was late with email. It was late with e-discovery. It was late with cybersecurity. It is still catching up to generative AI. In Da Silva Moore v. Publicis Groupe, Judge Andrew Peck’s opinion became an early milestone in judicial acceptance of predictive coding in discovery. I served as lead technology counsel in that case, and the resistance to predictive coding was intense, to put it mildly.

Most lawyers in 2011-2012 treated machine learning in document review as dangerous speculation. Nearly all of the legal profession was in denial. If they heard the clarion calls of machine learning, they did not believe it. Cynics point out they had strong billable hour incentives not to. Today, fifteen years later, technology-assisted review is ordinary. Almost no one manually reviews a hundred thousand documents these days, much less a million.

Quantum law may follow that same pattern in some areas, but not all. Some quantum issues may arrive gradually through vendors, expert systems, and hybrid AI tools. Q-Day may not. If the first breakthrough occurs inside a military or intelligence program, the legal profession may receive no clear warning at all. We may instead see the consequences unfold in apocalyptic scenarios that are hard to imagine without grounding in AI and quantum computer capabilities.

That is one reason we must skip the denial phase that happened with AI predictive coding in 2011. The stakes are much higher now. Common sense and professional ethics require it. ABA Model Rule 1.1, Comment 8 states that lawyers should keep abreast of changes in law and practice, including the benefits and risks of relevant technology. See ABA, Rule 1.1 Competence – Comment. The ABA made a similar move for generative AI in Formal Opinion 512. See ABA, Formal Opinion 512 on Generative Artificial Intelligence Tools.

Quantum computing will require the same kind of professional adjustment: familiar duties applied to new facts. That is not a call for panic. It is a call for competent tracking of emerging technologies, especially in computing. Be prepared.

A cartoon character representing a quantum particle, Mr. Quantum, dressed in a top hat and suit, enters a courtroom, humorously addressing the audience with the phrase 'Pardon the wave-function.' Several surprised people are seated at a table, listening attentively, while a judge presides in the background.

When Quantum Evidence Walks Into Court

Encryption is the first practical problem. Evidence is the second. Strictly speaking, a quantum computer will not “testify” because witnesses are people, at least so far. But quantum systems may generate outputs that parties will want to use as proof. A manufacturer may use a simulation to test a material under stress. A pharmaceutical company may use a quantum method to model a molecule. A financial institution may use a hybrid quantum-classical process to test risk under market conditions.

Consider a product liability case involving a battery fire in an electric vehicle. The company’s emails show engineers debating heat risks. The testing logs are incomplete. One side says the risk was known and ignored. The other side says the accident resulted from misuse or unusual conditions. Then an expert offers a simulation of the battery chemistry under conditions close to the fire. The simulation does not produce one answer. It produces a pattern of outcomes.

That pattern may be powerful evidence, but it also creates practical courtroom questions. What inputs were used? Who selected them? What assumptions were built into the model? Were any runs excluded? Could another qualified team reproduce the distribution, even if not every individual result? How does the expert explain the error rate to a judge who has a docket full of ordinary human disputes waiting outside the door? This is just a rough outline of the new types of legal questions and analysis you will need for the future of quantum.

Federal Rule of Evidence 702 should work fairly well for this kind of work, even if quantum evidence will stretch it. The rule requires expert testimony to rest on sufficient facts or data, reliable principles and methods, and reliable application to the case. See Federal Rule of Evidence 702. Rule 901 also matters because Rule 901(b)(9) addresses evidence about a process or system that produces a result. See Federal Rule of Evidence 901. Also see Daubert directs courts to consider factors such as testing, peer review, error rate, standards, and general acceptance. These factors will remain critical.

An infographic illustrating the concepts of identity and fidelity in a legal context, featuring a courtroom scene with a judge and professional witnesses. It includes sections highlighting DNA evidence, epidemiology, and risk assessments, emphasizing how the law relies on probability.

Identity Versus Fidelity

Lawyers like identity. We like exact copies, matching signatures, stable timestamps, and hash values that confirm a file has not changed. That instinct served us well in e-discovery. A hash value is a beautiful thing. It uses straightforward mathematical analysis to show whether the file is the same. The same values appear each time the hash analysis of the document is run.

Quantum systems often require a different instinct. The key question may not be whether the machine gives the identical output every time. The question may be whether it behaves with fidelity. Identity asks whether we got the same answer again. Fidelity asks whether the system behaved as expected, within known error limits.

A courtroom analogy helps. Suppose a careful witness is asked three times whether the traffic light was red. On Monday she says she is almost certain it was red. On Tuesday she says she would put the probability very high. On Wednesday she says red is by far the most likely explanation. A cross-examiner hears only contradiction. A better lawyer hears the same judgment expressed in different language.

Quantum outputs can work in a similar way. Variation is not always unreliability. Sometimes variation is the form the reliable answer takes. Law already understands this better than it admits. DNA evidence, sampling, epidemiology, damages models, and risk assessments all rely on probability. We go into this in detail in the course using both published cases and hypotheticals. Quantum evidence will make probability too visible to ignore. It will add a new dimension to the core legal concept of causation.

Infographic illustrating the transition from traditional documents to model evidence in legal processes, featuring sections on document evidence, modeling examples, discovery requirements, and court decision-making.

From Documents to Models

For most of legal history, lawyers have been document hunters. Who wrote the email? What did the contract say? Where is the missing report? What did the board know? That world is not disappearing. Documents still show notice, intent, concealment, delay, agreement, and knowledge.

But AI and quantum systems push law toward model evidence. A model does not merely record what happened. It tests what likely would happen under stated conditions. Return to the battery-fire example. The emails may show that engineers discussed risk. The testing logs may show what the company actually checked. The simulation may show what the company failed to test.

The simulation does not replace the documents. It interrogates them. That is the practical shift from document-centric law to model-centric law. A discovery request may need more than the final report. It may need the inputs, assumptions, validation work, version history, and excluded runs. The producing party will raise burden, trade secret, and proportionality objections. The court will have to decide how much process disclosure is enough.

This is familiar territory in new clothing. We fought similar battles over metadata, native files, search terms, sampling, and predictive coding protocols. Quantum evidence will bring another version of the same fight: how much of the machine’s process must be disclosed before the result can be trusted?

Infographic illustrating the potential impact of AI on quantum computing timelines, featuring elements like a quantum computer, a panther representing acceleration, and a researcher analyzing data, along with text highlighting key points about error correction and research implications.

AI May Speed the Quantum Timeline

Lawyers are still adjusting to generative AI, but AI is already part of the quantum story. A 2025 Nature Communications review explains that AI is increasingly being used to help with quantum systems. In plain English, AI can help tune fragile machines, find errors, and keep quantum hardware closer to the narrow conditions required for useful work. See Artificial Intelligence for Quantum Computing, (Nature Communications, 12/02/25).

Google DeepMind’s AlphaQubit is one concrete example AI enhanced software. It identifies quantum-computing errors with greatly improved accuracy. Error correction is one of the central barriers to making quantum computers useful at scale. See AlphaQubit tackles one of quantum computing’s biggest challenges (11/20/24).

The practical point for lawyers is modest but important. Do not assume quantum development will proceed on a slow schedule convenient for law firm committees. AI may help researchers move faster. Quantum tools may later assist certain kinds of AI work, especially where optimization or simulation is the bottleneck. The feedback loop remains uncertain, but the first half of AI helping quantum is already underway. That matters because lawyers should not assume that quantum progress will move on a slow, linear timetable convenient for bar committees, vendor reviews, and CLE calendars. To me, a slow arrival would be shocking. I have seen an increase in the pace of change of technology my whole life. I see no reason this will not continue. The quantum floor is not a barrier; it is an opening.

An infographic featuring a man in a suit sitting at a desk with legal symbols around him, discussing advanced concepts like Willow, quantum echoes, and supercomputing. The background includes a futuristic corridor and text highlights on technological advancements and legal implications.

Willow, Quantum Echoes, and the Word Lawyers Should Notice

Google’s announcement in late 2024 of results achieved by its quantum computer, Willow, captured public attention. Willow performed a benchmark computation in under five minutes that would take our fastest AI supercomputers 10 septillion years. See Hartmut Neven, Google, Meet Willow, Our State-of-the-Art Quantum Chip. That claim naturally led to excitement, skepticism, and multiverse speculation. Quantum Leap: Google Claims Its New Quantum Computer Provides Evidence That We Live In A Multiverse (01/09/25, my all-time most read JDSupra article).

The multiverse is fascinating, but lawyers can leave it aside for practical purposes. The more important legal lesson comes from Google’s subsequent work on AI improved software, Quantum Echoes. In late 2025 Google described its new Quantum Echoes software as a step toward verifiable quantum advantage. The algorithm supposedly ran 13,000 times faster on Willow than the fastest supercomputers. See Google, The Quantum Echoes Algorithm Breakthrough (10/22/25).

The word that should matter most to lawyers is not “faster.” It is “verifiable.” A spectacular claim is not evidence merely because it sparkles. A courtroom claim must be tested, explained, challenged, and tied to the legal issue. Verification is the bridge between physics and proof.

What Lawyers Should Do Now

The first step is to stop treating quantum as trivia. You do not need to understand the math to recognize where it may matter. If a client has long-lived secrets, ask whether quantum risk belongs in the confidentiality analysis. If a vendor holds sensitive data, ask about cryptographic migration. If an expert relies on a simulation, ask for process evidence, not just conclusions.

For contracts, avoid vague comfort language. A clause promising “commercially reasonable security” may not tell you enough. Ask who controls the keys, what encryption is used, whether the vendor tracks NIST post-quantum standards, and how the vendor will notify customers when migration affects stored data. These are not physics questions. They are vendor-management questions. The time to learn this is now.

For litigation, start thinking about model evidence. When an expert relies on a simulation, ask for inputs, assumptions, validation work, excluded runs, and error analysis. Do not wait until the Daubert hearing to discover that the “black box” is really a locked box and nobody brought the key.

For courts, the task is not to become a laboratory. It is to insist on understandable explanations, fair disclosure, and honest limits. The judge’s job remains what it has always been: decide what is reliable enough to consider and what weight it deserves.

A promotional graphic for a law course titled 'Quantum Law Course', featuring a scenic landscape with a lawyer standing in the foreground. The image includes text that emphasizes the importance of preparation and knowledge in legal practice, with a laptop displaying the course website and various law books in the background.

Learn the Terrain Before the Emergency

Quantum computing will not eliminate legal judgment. It will make judgment more important. Some quantum issues may arrive quietly, hidden inside vendor tools, expert simulations, cybersecurity updates, and hybrid AI systems. Others may not arrive quietly at all. If Q-Day comes through a breakthrough in fault-tolerant quantum computing, the legal profession may not get a polite warning, a fixed deadline, or a long runway. It may come instead as a very rude awakening. Will you be prepared to answer the client calls?

Lawyers do not need to predict the exact date. They need to understand the questions that date will create, what the contours of the emergencies will be. Which old archives are worth protecting? Which vendor promises are too vague? Which expert models can be tested? Which court orders should require more than a final output? Which risks are speculative, and which are already present because data can be harvested now and decrypted later?

Those are not physics questions. They are legal judgment questions. The lawyer’s task is not to master the machinery, but to know enough to question the machinery, the vendor, the expert, and sometimes the client’s own assumptions. That is familiar work. The tools are new, but the professional responsibility is not.

That is why I created the online QuantumLawCourse.com.

The course is designed for legal professionals, not physicists. No math. Instead, it uses case law, legal reasoning, practical examples, and the kinds of concerns lawyers, legal tech professionals, and judges face every day. It focuses on confidentiality, evidence, expert testimony, cybersecurity, risk, and professional responsibility.

The goal is not to make you a quantum expert. It is to help you become an informed legal professional who understands enough to recognize the issues, ask better questions, and avoid learning the hard way when quantum law arrives in your own practice. This article only sketches the terrain. The course provides a full map.

A person in a suit standing on a path leading to a grand building, with columns on either side. The scene features elements representing privacy and judgment, such as locks and scales, with a futuristic cityscape in the background. The text "QUANTUM LAW" is prominently displayed above, along with the phrases "PRIVACY, PROOF, JUDGMENT" and "UNDERSTAND TODAY. LEAD TOMORROW."

Conclusion

The law has always had to judge under uncertainty. Quantum computing does not change that responsibility. It makes the uncertainty harder to ignore. Machines may calculate. Experts may explain. AI may help interpret. Vendors may package the result in polished dashboards. But courts, lawyers, regulators, and clients will still need reasons, evidence, standards, accountability, and courage.

Privacy, proof, and judgment are not abstract concerns. They are the daily work of the legal profession. Quantum computing may affect all three, sometimes gradually and perhaps someday, quite suddenly. Waiting until the emergency arrives is the surest way to fumble, struggle, and learn in public.

Consider taking the Quantum Law Course now, while the field is still emerging and there is time to prepare. The best time to learn a new legal technology is before it appears in your next emergency motion, vendor presentation, expert challenge, or board-level crisis.

Quantum law is coming. Lawyers who understand the terrain early will be better prepared to protect clients, question experts, contest false claims, and help courts make sound decisions.

1 Comment | AI Ethics, AI Instruction, Hacking, Lawyers Duties, New Rules, Quantum Law, Technology, VENDORS | Permalink
Posted by Ralph Losey

The Goblin in the Machine: What OpenAI’s “No-Pigeon Rule” Teaches Lawyers About AI Hallucinations

May 11, 2026

Ralph Losey, May 2026

This article is about a real event. It is not satire, parody, or metaphor. In late April 2026, OpenAI publicly explained why one of its frontier AI systems had developed an unusual tendency to mention goblins, gremlins, raccoons, trolls, ogres, pigeons, and similar creatures in places where they did not belong. OpenAI titled its official explanation “Where the Goblins Came From.” The title sounds fictional. The problem was not.

A humanoid robot with a friendly face sitting at a desk next to a coffee mug. The computer screen displays coding instructions and a highlighted warning about avoiding certain topics, including goblins and trolls, unless relevant to the user's prompt. — **Gremlins, Goblins and Pigeons. Oh my!**

If you take the time to study this strange episode, you will gain more than an amusing story about artificial intelligence. You will see, in unusually visible form, how Large Language Models can acquire unintended behavior from training incentives, how that behavior can spread beyond its original context, why prompt-level or developer-level instructions may be used to suppress it, and how the same root causes help explain the ongoing problem of AI hallucination. For lawyers, judges, e-discovery professionals, and legal technology vendors, this is not a curiosity. It is a warning label written in unusually memorable ink.

A collage of fantastical creatures including a green goblin, a mischievous gremlin, a large orange monster, a raccoon, a small brown creature, and a pigeon, all surrounding a glowing, swirling vortex in a cosmic background. — **Fact is sometimes stranger than fiction. This is one of those times.**

The Most Bizarre Codex Instruction of All Time

OpenAI’s example involved Codex, its AI coding agent. For non-programmers, Codex is not a fantasy product and not a casual chatbot. It is a professional software-development tool designed to help engineers plan, write, refactor, test, review, and release code. OpenAI describes Codex as “a coding agent that helps you build and ship with AI,” used for real engineering work across development tools.

That context matters. The now-famous instruction was not a joke inserted into a toy system. It was a developer-level instruction in a serious AI coding agent. According to reporting and OpenAI’s later explanation, Codex had been instructed not to talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless they were clearly relevant to the user’s request.

WIRED first reported the Codex CLI instruction that the model should “never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.” Maxwell Zeff, OpenAI Really Wants Codex to Shut Up About Goblins (WIRED, Apr. 2026). OpenAI, then responded with its own article, Where the Goblins Came From, OpenAI (Apr. 29, 2026), explaining that GPT-5.5 in Codex showed an affinity for goblin metaphors and tracing the behavior to training incentives connected with the “Nerdy” personality. It is well worth the read.

The facts are unusual enough that they do not need embellishment. Indeed, embellishment would weaken the point. The issue is not that an AI system said something funny. The issue is that a frontier model, shaped by modern training methods, developed a persistent behavior that its maker had to investigate, explain, and mitigate. That is precisely why lawyers should pay attention.

A whimsical scene featuring a wizard in a green robe controlling a machine labeled 'GPT-5.5/CODEX BEHAVIOR CONTROL.' In front of the wizard, there are two small goblin-like creatures and a pigeon, all looking towards the control panel. A sign reads 'Gremlins, Goblins, and Pigeons, OH MY!' in the background. — Pay no attention to the Codex instruction behind the curtain.

The “Goblin” Problem Was an Alignment Problem in Plain Sight

The legal technology world often discusses AI alignment in abstract language. We talk about bias, safety, truthfulness, reliability, explainability, auditability, and human values. Those are important words, but they can become bloodless. The goblin incident gives us something more concrete.

OpenAI explained that the behavior emerged from “many small incentives,” including training by AI of itself connected to its personality customization feature, especially an introversive “Nerdy” personality. That personality was designed to make the model more playful, intellectually enthusiastic, and metaphor-friendly. In the process, certain creature metaphors were rewarded often enough that the model learned to repeat and generalize them.

I have frequently written about the ability of AI to form fictitious sub-personalities for brainstorming purposes, and note the Devils Advocate character is especially effective. Fortunately he was not involved in this OpenAi fiasco. I never instructed AI to form a shy, super-nerd personality type for training purposes. If I ever do in the future (doubtful), I will obviously be very careful to provide strong human supervisions, something which was obviously missing here. This whole incident seems like over-delegation, where the humans in the loop were not paying attentions and so triggered this Gremlin crisis,

This brings up a key point. The AI model was not “thinking about goblins.” It was responding to patterns shaped by training data, reinforcement learning, preference signals, and later adjustments. If a certain style of answer receives favorable feedback, the model can learn that style as a useful pattern. If that pattern includes odd creature metaphors, those metaphors can become part of the model’s behavior.

OpenAI’s post-mortem is valuable because it shows something that usually remains hidden. Model behavior does not simply appear at deployment. It is cultivated. It is selected. It is rewarded. It is penalized. It is patched. It is monitored. Sometimes, it is suppressed by instructions that users never see. I never knew that before.

In this case, the visible symptom was bizarre. The underlying process was ordinary. That is what makes the episode important.

Infographic explaining the 'Goblin' problem in model training, featuring sections on inputs, emergent behavior, unintended outcomes, and mitigation strategies. Includes illustrations of goblins and reference to model training inputs like human feedback and evaluations.

What Are These “Instructions,” and Why Should Lawyers Care?

Modern AI systems are not governed only by the words users type into the chat window. They also operate under layers of instructions. Some instructions come from the system level. Some come from developers. Some come from product settings, safety policies, tool configurations, or specialized agent workflows. Some come from users themselves. The user may never see, nor even know about the developers instructions that shape the response to the user’s prompts.

A developer instruction is essentially a command placed above the ordinary user prompt. It tells the model how to behave in a particular product environment. In Codex, such instructions may shape how the model writes code, uses tools, comments on programming tasks, avoids certain behaviors, or responds within a software-development workflow.

That is not improper. In fact, layered instructions are necessary. A legal AI tool should be told to protect confidentiality, avoid unauthorized practice of law, cite sources, flag uncertainty, preserve privilege, and follow the user’s workflow. The problem is not the existence of instructions. The problem is invisibility, auditability, and as just mentioned, the lack of proper human supervision of the whole process. The humans in the loop were asleep at the wheel and as a consequence the dogs got out.

In legal work, hidden constraints can matter. If a model suppresses certain language (such as profanity), and favors certain categories (such as propriety), emphasizes certain risks (such as letting the dogs out), avoids certain conclusions (such as user is wrong), or changes behavior after an update (such as no hacking allowed, eh Claude), the lawyer may not know why. That matters in e-discovery, privilege review, contract analysis, legal research, expert preparation, and litigation strategy. Another layer of e-discovery open up.

The Codex no-goblin instruction is therefore not important because lawyers care about goblins. (I for one do not, although I do. care about ‘not letting the dogs out’). It is important because it reveals how behavioral control can operate behind the scenes.

Infographic titled 'Hidden Instructions. Real Impact.' illustrating the differences between user input and underlying model instructions. It shows an iceberg with 'User Prompt,' 'System Instructions,' 'Developer Instructions,' 'Tools & Data Sources,' and 'Model Behavior Shapers' listed under the waterline. A person is seen contemplating the information with a notebook and pen on the table, emphasizing the importance of understanding hidden instructions in AI output. — If a goblin ever appears in your AI response you will know why now. The super-nerd trainer slipped through the latest hidden instructions.

The Hallucination Connection

The goblin problem is not identical to hallucination, but the two issues share root causes.

The goblin problem involved an unintended stylistic habit. Hallucination involves plausible but false content. One produces irrelevant creature metaphors. The other produces fake cases, invented quotations, nonexistent statutes, false summaries, fabricated citations, or confident statements unsupported by the record.

The difference is obvious. The connection is deeper.

Both problems arise from the same basic fact: Large Language Models are not born as truth engines. They are trained to predict and generate language. Later training stages, including supervised fine-tuning, reinforcement learning, preference optimization, safety training, and evaluation systems, try to make that language helpful, accurate, safe, and aligned with user expectations.

But training incentives can misfire. Evaluation methods can reward the wrong behavior. A system can learn to produce answers that sound good rather than answers that are verified. It can learn fluency before truth, confidence before calibration, and completion before uncertainty. It could be trained to say, “I don’t know,” but it wasn’t. There is not much of that on the Internet. So, instead it just makes up an answer, one that it infers the user wants, because it is also trained to be a nice sycophant. Nobody wants a devils advocate around that disagrees with you. We should of course, and that is why lawyers have the potential to be great users of generative AI.

OpenAI made this point directly in its 2025 discussion of why language models hallucinate. Why Language Models Hallucinate, (OpenAI, Sept. 5, 2025). OpenAI explained that hallucinations persist in part because many evaluation systems reward accuracy alone, which can push models to guess rather than admit uncertainty. If a model guesses, it may get lucky and receive credit. If it says “I don’t know,” it may receive no credit at all. Over many evaluations, that scoring structure can make a guessing model appear more successful than a more careful model that abstains when it lacks reliable information.

That is the real connection between goblins and hallucinations. They are different failures, but they reflect the same training logic. In the goblin case, the rewarded behavior was playful metaphor, so the model learned to repeat and generalize playful creature references. In hallucination, the rewarded behavior is often answer-giving itself, so the model may learn to produce a confident response even when it lacks adequate grounding. In both cases, the model is not following truth as an independent legal or evidentiary standard. It is following patterns that its training, feedback, and evaluation systems have taught it to treat as successful.

The danger for lawyers is that hallucinations usually do not look strange. Goblins and pigeons are obvious intrusions. They announce that something has gone wrong. A fake citation does not. A fabricated quotation does not. A false summary of a contract clause, deposition answer, medical record, email thread, or judicial opinion may read with the same polish and confidence as a correct one. The surface quality of the prose may conceal the absence of reliable support.

That is why hallucinations are more dangerous than the goblin problem. The goblins expose the machinery because they look absurd. Hallucinations hide the machinery because they look professional. For legal work, that difference is critical. The risk is not merely that an AI system may be odd. The risk is that it may be wrong in a way that looks authoritative, usable, and ready to file.

An illustration featuring goblins and a bird discussing the concept of incentives and risks, contrasted with labels like 'Obvious,' 'Strange but Obvious,' and 'Plausible but Dangerous.' The central theme highlights differing risks associated with learned behaviors, with references to legal aspects and the importance of verification. — Don’t be a pigeon. Trust but verify.

This Is Not Just an OpenAI Problem

It would be a mistake to treat this as an OpenAI-only issue. The OpenAI goblin post-mortem is useful because it is unusually visible, candid, and memorable. But hallucination and unintended model behavior afflict all modern LLM systems under development, including Claude, Gemini, and other leading models.

Anthropic’s own Claude documentation expressly addresses hallucination reduction, warning that even advanced models can generate text that is factually incorrect or inconsistent with context, and recommending mitigation techniques such as allowing Claude to say it does not know, grounding answers in provided source material, using direct quotations, verifying with citations, and validating critical information. Anthropic, Reduce Hallucinations (Claude API Docs).

Google’s Gemini documentation similarly warns that Gemini for Google Cloud may produce hallucinations, including outputs that are plausible-sounding but factually incorrect, irrelevant, inappropriate, or nonsensical, and may even fabricate links to web pages that do not exist and have never existed. Google Cloud, Gemini for Google Cloud and Responsible AI (Google Cloud Documentation),

The vendors differ. The architectures differ. The safety philosophies differ. The product interfaces differ. But the fundamental problem is shared. These systems are trained to generate plausible language under complex incentives. Plausibility is not truth. Fluency is not verification. Confidence is not reliability.

This point should be stated carefully. It does not mean that all systems are equally risky, equally useful, or equally well governed. They are not. Some models perform better than others on particular tasks. Some products provide stronger grounding, citation, retrieval, logging, or enterprise controls. Some workflows are safer than others.

But no responsible legal professional should assume that hallucination and goblins are confined to one vendor. It is a structural limitation of current LLM technology.

An illustration emphasizing the responsibilities associated with AI models, featuring logos of OpenAI, Anthropic, and Google. The background includes law-related imagery and a checklist titled 'Lawyer's Checklist' with items for verifying information. — Advanced AI construction and use require human supervision and skills.

The Legal Technology Lesson

Legal professionals should not respond to this by rejecting AI. That would be the wrong lesson. It would also ignore the enormous value these tools already provide when used with care.

The correct lesson is disciplined adoption.

In e-discovery, we already understand this principle. Technology-assisted review is not accepted because someone declares the software intelligent. It is accepted when the process is reasonable, validated, documented, and proportionate. Sampling matters. Quality control matters. Human judgment matters. Reproducibility matters. Transparency matters.

The same discipline must now be applied to generative AI. Legal AI workflows should be designed to answer practical questions:

Can the output be traced to reliable source material?
Did the model actually use the cited source?
Can each legal citation be verified?
Can each quotation be checked against the original?
Can each factual assertion be tied to the record?
Can the workflow be reproduced if challenged?
Was the model permitted to say “I don’t know”?
Was uncertainty preserved, or did the workflow pressure the model into confident completion?
Were model version, prompt structure, source set, and review procedures documented?
Was a qualified human responsible for final legal judgment?

These questions are not anti-AI. They are pro-reliability. They are the questions that separate professional use from casual use.

Why This Matters for Courts and Clients

Courts do not need lawyers to become machine-learning engineers. Clients do not need their lawyers to understand every detail of transformer architecture. But both courts and clients are entitled to competent professional judgment.

That includes knowing when an AI output is grounded and when it is merely plausible. It includes knowing when a citation has been verified and when it has merely been generated. It includes knowing when an AI tool is being used for brainstorming, drafting, summarization, classification, legal research, or evidence analysis, because each use carries different risks.

The goblin incident offers a rare window into model behavior because the symptom was so visible. Most legally significant failures will not be so obvious. They will not involve fantasy creatures. They will involve a misstated holding, an omitted exception, a distorted fact pattern, a privilege call made too broadly, a missed document, or a confident statement about law that is no longer current. By the way, humans can all make the same mistakes, which is one reason we tend to do better working in small teams.

That is why the legal profession, indeed all of humanity, must treat generative AI as powerful but not self-validating.

An illustration depicting the balance between artificial intelligence (AI) and human judgment, emphasizing the importance of verification and accurate legal practices. The image shows a scale weighing truthful information against misleading data, with a group of professionals discussing documents at the bottom. — Seriously, why pigeons? None of my associates ever made that mistake.

Practical Guidance for Lawyers and Legal Tech Users

The practical response is straightforward:

Use AI, but verify.
Use AI for first drafts, issue spotting, summarization, brainstorming, and classification support, but do not outsource professional judgment.
Use retrieval, citations, and source-grounded workflows whenever factual accuracy matters.
Require the model to distinguish between sourced statements, inferences, and speculation.
Require explicit uncertainty when the record is incomplete.
For legal research, verify every case, statute, rule, quotation, and parenthetical against authoritative sources.
For e-discovery and document review, use sampling, validation, audit trails, and human quality control.
For AI vendor selection, ask what model is being used, how outputs are grounded, how hallucination risk is measured, what logs are preserved, what changes when the model is updated, and whether the workflow can be explained if challenged.
For judicial or regulatory settings, avoid vague claims that an AI tool is “aligned,” “safe,” or “accurate” without evidence. Ask what was tested, how it was tested, and under what conditions.

The lesson is not distrust. The lesson is earned trust.

A woman weighing scales in an office setting, emphasizing the importance of using AI tools while verifying information. Text highlights various uses for AI and verification methods.

Conclusion: The Promise and the Work Ahead

At the beginning of this article, I promised that this strange episode would offer more than an amusing story. It does.

OpenAI’s real no-goblin, no-pigeon instruction gives lawyers a concrete example of how modern AI behavior can be shaped by training incentives, generalized beyond its original setting, and later mitigated through hidden or semi-hidden instructions. The hallucination problem shows the same root issue in more serious form. When models are rewarded for fluent completion, confidence, and benchmark performance, they may learn to answer when they should abstain, to sound certain when they should qualify, and to generate plausible legal authority when only verified authority will do.

Users must learn these idiosyncrasies and adapt.

This is not just about OpenAI. It is not just about Codex. It is not just about goblins. It is about every legal professional’s duty to understand the tools now entering legal practice. It is about understanding how to use them properly.

Generative AI can help lawyers become faster, broader, more creative, and more effective. It can improve access to justice, reduce drudgery, accelerate document review, strengthen legal education, and help professionals see patterns they might otherwise miss. But these benefits will not be realized by pretending the risks are gone. They will be realized by confronting the risks directly and building better habits, better workflows, better audits, better training, and better professional norms.

The goblins are real in the only sense that matters here: real enough to show us how fragile model behavior can be. The hallucinations are more dangerous because they usually do not look strange at all.

That is the call to action. Legal professionals should not stand outside the AI revolution, arms folded, waiting for perfect machines. Nor should they rush in, eyes closed, dazzled by fluent output. We should do what good lawyers have always done with powerful evidence and powerful tools: question them, test them, document them, verify them, and use them responsibly.

The future of legal AI will not be built by blind trust or reflexive fear. It will be built by informed confidence.

And informed confidence begins with verification.

A woman in a suit standing with her back to the viewer, looking toward a bright horizon. Elements include a mythical creature on the left, a pigeon, an open laptop, a magnifying glass, and a scale of justice, all suggesting a theme of adaptation and learning.

For educational use only. Not legal advice.

Leave a Comment » | AI Ethics, AI Instruction, ChatGPT, Technology, VENDORS | Tagged: best practices, machine learning, science, technology | Permalink
Posted by Ralph Losey

Something Big Is Happening — But Not What You Think

February 23, 2026

Ralph Losey. February 23, 2026

A Response to Matt Shumer’s Viral Essay on AI Acceleration

A high-speed train in motion on railway tracks during sunset, creating a dynamic sense of speed with blurred background. — Acceleration without control is dangerous. Acceleration with judgment is transformative.

I. Something big is happening. On that much Matt Shumer and I agree.

The essay Something Big Is Happening was published on Matt Shumer’s personal blog on February 9, 2026. After he shared it widely on X, it drew more than 80 million views within days, rapidly becoming a focal point in public debates about AI and the future of work. Few essays about artificial intelligence have traveled that far, that fast.

Shumer’s central claim is straightforward: AI capability is accelerating so rapidly that large-scale displacement of white-collar work is imminent, perhaps within one to five years. He argues that recursive improvement loops are already underway, that benchmark curves are steepening, and that most people are underestimating what is about to happen.

It is a powerful narrative. It is also incomplete, and that matters more than its popularity suggests. So take a breath.

Before I explain why, a brief word of context. I have practiced law for over 45 years and have worked hands-on with AI in litigation for more than 14. I was involved in the first case approving predictive coding for e-discovery in federal court. Since 2023, I have written extensively about generative AI, hybrid human-machine workflows, and the emerging governance challenges of AI and quantum convergence. I am not skeptical of AI — I use it daily, teach it, and advocate its responsible adoption.

Acceleration is real. But acceleration demands adults – a calm, measured approach. That is why I take Shumer seriously, even as I disagree with his conclusions.

II. What Shumer Gets Right (and What He Exaggerates)

Let us begin where we agree. AI models have improved rapidly. Coding autonomy has advanced in ways that would have seemed implausible just a few years ago. AI systems now assist meaningfully in debugging, evaluation, and even aspects of their own development pipelines. Benchmarks measuring the duration of tasks that models can complete without human intervention have indeed increased.

There is rapid acceleration, but it is not a smooth, universal climb. It is jagged.

A. The Bar Exam Myth: Top 10% or Bottom 15%?

humer states: “By 2023, [AI] could pass the bar exam.” This has become a foundational myth in the AI-acceleration narrative. However, a rigorous study by Eric Martinez showed the truth of the vendor study. Re-evaluating GPT-4’s bar exam performance. Artif Intell Law (2024) (presenting four sets of findings that indicate that OpenAI’s estimates of GPT-4.0’s Uniform Bar Exam percentile are overinflated). Martinez found that when you limit the sample to those who actually passed the bar (qualified attorneys), the model’s percentile drops off a cliff. On the essay and performance test portions (MEE + MPT), GPT-4 scored in the ~15th percentile. In other words, bottom 15% among those who passed.

B. AI Hallucinations Are Not Ancient History

Shumer claims that the “this makes stuff up” phase of AI is “ancient history” and that current models are unrecognizable from six months ago. My daily use and objective tests tell a different story. Yes, it is getting better but we are not there yet, especially for most legal users.

Hallucination remains the number one concern for the Bench and Bar. Cross-Examine Your AI: The Lawyer’s Cure for Hallucinations (December 2025). Generative AI still has a persistent tendency to fabricate facts and law, leading to serious court sanctions. See e.g. Park v. Kim, 91 F.4th 610, 612 (2d Cir. 2024). Also see French legal scholar Damien Charlotin‘s catalogue of almost one thousand similar decisions worldwide in his AI Hallucination Cases.

Shumer’s claims that modern AIs no longer hallucinate and outperform most attorneys reflect optimism more than sustained exposure to legal work. After researching tens of thousands of legal issues over the course of my career, I can tell you that verification is not optional — it is the job.

C. The “Jagged Frontier” of AI Progress

Shumer envisions a wall of fast, inevitable advance. Research and personal experience of many experts suggests otherwise. The progress is jagged and uneven. See e.g., The New Stanford–Carnegie Study: Hybrid AI Teams Beat Fully Autonomous Agents by 68.7%. Also see research and reports of top expert teams in Navigating the Jagged Technological Frontier (Working Paper 24-013, Harvard Business School, Sept. 2023) and my humble papers, From Centaurs To Cyborgs, and Navigating the AI Frontier.

The New Stanford–Carnegie Study (November 2025) confirmed what Harvard researchers call the “Jagged Technological Frontier”. This research found that AI excels at specific programmable tasks but fails at messy, human-centric reality. In fact, the Stanford-Carnegie study showed that fully autonomous AI agents were significantly less reliable than hybrid human-AI teams, which outperformed solo agents by 68.7%.

D. “Team of Human Associates” or “Untested Sycophantic AI Experts”?

Shumer recounts a managing partner in a law firm who feels AI is like “having a team of associates available instantly.” I agree that every professional should be integrating AI into their daily workflow. But they must do so skeptically. Plus, it is nowhere near the same as having trained human associates. AIs are cheaper, sure, until they screw up and you are the one left cleaning it up.

In my 45-years of legal practice I have had the privilege of working with many excellent associates. They significantly exceed today’s AIs in many respects, so I must respectfully disagree with Shumer’s quote of an anonymous partner. There are many things that AI will never be able to do that all good professionals now do without thinking. The Human Edge: How AI Can Assist But Never Replace. I prefer humans with AIs – the hybrid approach – over AIs alone, even though, unlike humans, AI associates are always pleasant and they tend to agree with everything you say. Lessons for Legal Profession from the Latest Viral Meme: ‘Ask an AI What It Would Do If It Became Human For a Day? (Jan. 2026).

My testing of AI since 2023 has focused on the legal reasoning ability of AI, as opposed to general reasoning. For a full explanation of the difference, see Breaking New Ground: Evaluating the Top AI Reasoning Models of 2025. I have also spent hundreds of hours in hands-on independent testing of the AI legal reasoning abilities. See e.g., Bar Battle of the Bots, parts one, two, three and four. These articles reported multiple tests of Open AI and Google models in 2025, including tests using actual Bar exam questions, which they again failed. I have not seen substantial improvements in AI since then.

It is, in my experience a poor trade to use an AI instead of an associate-Ai team, and without extensive supervision, an invitation to sanctions and malpractice.

III. Benchmark Curves Are Not Civilization

Shumer relies heavily on task-duration benchmarks and exponential trend lines. The implication is clear: if models can complete longer and longer tasks autonomously, then large-scale displacement is imminent.

The problem with that is benchmark extrapolation is not societal destiny. In law, evidence does not decide the case. People do.

Most current autonomy benchmarks are domain-constrained. They focus heavily on software engineering and other structured digital tasks. Coding is not law. It is not medicine. It is not fiduciary duty. It is not governance.

Even when capability expands inside a benchmark, that does not mean institutions will move at the same speed. Courts, regulators, insurers, boards of directors, and compliance departments slow, shape, and channel technology. That is not inertia; it is risk management.

And assistance in model development is not the same as autonomous recursive self-governance. Humans remain deeply embedded in training, validation, and deployment. “AI helping build AI” makes for a compelling headline. It does not mean an intelligence explosion has detached from human control. AI extends cognition, but it does not replace stewardship.

That is the part Shumer’s curve does not capture: acceleration of capability is real but this increases the need for adult supervision. It does not eliminate the human role. It intensifies it. Just as it always has.

IV. Why Fear Travels Faster Than Wisdom

The viral success of Shumer’s essay is not accidental. It was designed to activate powerful psychological mechanisms.

It invokes the COVID analogy, reminding readers how quickly life changed in early 2020. It frames the reader personally: “you’re next.” It emphasizes exponential growth, which humans are notoriously poor at intuitively processing. It adopts insider authority: “I live in this world; I see what you don’t.”

Fear spreads faster than nuance because we evolved that way. A possible threat demands immediate attention. Social media algorithms amplify high-emotion content. Urgency increases engagement velocity. None of this necessarily makes Shumer insincere but it does explain why his article went viral. Acceleration narratives travel at super-fast computer speeds. Wisdom still travels at human speed.

V. Incentives Shape Narratives

It is also important to understand context. Shumer is a very young builder who lives in the code. His perspective is shaped by the possibility of the technology. My perspective, and the perspective of governance, is shaped by the consequences of the technology. Startup culture rewards speed; legal culture rewards survivability. These are different risk environments.

Recognizing that difference is not an attack. It is transparency. His incentives don’t invalidate his argument, but they do shape his narrative.

VI. A Structural Irony

Here is another irony worth reflecting on. We are now in an era where AI systems assist in drafting almost all persuasive content. Many viral essays, legal briefs, and opinion pieces share a similar highly optimized narrative arc—a cadence and structure that Large Language Models excel at producing.

If an AI is optimizing for “popularity” – to become the next great flash meme – it will naturally drift toward alarmism, because alarmism travels faster than nuance. It is entirely plausible that AI systems are increasingly shaping the very rhetoric used to warn us about AI. That is not necessarily a deception, but it is a reminder: persuasion optimization is not the same as civilizational wisdom.

VII. The Category Mistake: Doing the Task Is Not Being the Lawyer

Here is the deeper mistake in many inevitability arguments. They confuse task performance with personhood.

Yes, AI completes tasks. Sometimes very well. It predicts the next word, the next clause, the next block of code. At scale and at speed. But practicing law is not just completing text.

Human reasoning is not happening in a vacuum. It happens inside a body that can lose a license. Inside a reputation built over decades. Inside an ethical framework enforced by courts and bar regulators. Inside institutions that impose consequences.

AI does not stand in a courtroom or sign pleadings. AI does not carry malpractice insurance.

Law makes this distinction painfully clear. AI can draft a brief in seconds. I use it for that to start a review and verify process. But drafting is not signing. When a lawyer signs a motion, that signature attaches a human name, a bar number, a reputation, and a career to every word on the page.

If the brief is reckless, the AI does not get sanctioned. If the citation is fabricated, the AI does not face discipline. If the argument crosses an ethical line, the AI does not stand before a grievance committee. A probabilistic system cannot be disbarred.

Automation can transform tasks. It cannot assume moral agency. That distinction matters. And it will continue to matter, no matter how fast the models improve.

A close-up of a person's finger hovering over a laptop keyboard while signing a document electronically. The screen displays a signature field with the name 'John Smith' and a 'Confirm Signature' button. — Drafting is not signing. Accountability remains human.

VIII. Quantum Convergence Raises the Stakes

The need for adult supervision of accelerating technology becomes even more critical as we look at what is coming next. We are entering a new period where AI intersects with quantum computing. If AI is a race car, Quantum is the nitrous oxide. You do not put a novice driver behind that wheel.

Quantum-scale compute raises national security questions, cryptographic vulnerabilities, and governance complexity. More powerful systems require more sophisticated oversight frameworks. Power without governance is destabilizing; power with governance is transformative. The question is not whether capability grows—it is whether wisdom keeps pace.

The greatest short-term danger is not AI superintelligence overthrowing society, whether enhanced by quantum or not. It is over-delegation. It is professionals putting systems on autopilot. It is institutions adopting tools without supervision, audit trails, and verification. The solution is not panic. It is disciplined integration. Trust but verify.

IX. What Responsible Adoption Looks Like

Use AI seriously. Experiment daily. Adopt paid tools where appropriate. Automate repetitive tasks. I agree with Shumer on this.

But at the same time: Maintain human review. Preserve accountability. Document workflows. Understand limits. Teach younger professionals hybrid reasoning working with AI, not dependency.

The future belongs to those who combine human judgment with machine capability. Not to those who surrender to inevitability narratives

We have made this error before. We mistake acceleration for autonomy. We mistake tools for replacements. And each time, we rediscover that human responsibility does not disappear when machines improve. It intensifies.

X. Something Big Is Happening

Shumer is right that “something big is happening.” AI capability is advancing. Workflows are changing. New economic pressures are emerging. But history teaches us that technological acceleration does not eliminate the need for human beings. It heightens it.

This is where law and governance have to re-enter the conversation. Society should not allow its economic and moral direction to be set by the most amplified voices in tech, especially when those voices operate within incentive structures that reward urgency. We need engineers, not promoters. We need experience, not exuberance. We need wisdom, not just information.

Above all, we need adults in the room. Acceleration does not remove the human role. It demands judgment, accountability, and institutional memory.

A group of four professionals engaged in a discussion around a conference table, with laptops open and documents spread out, in a modern office setting. — Capability accelerates. Responsibility must keep pace.

Something big is happening. What happens next depends on whether we meet it with fear, or calm skepticism.

1 Comment | AI Ethics, AI Instruction, AI Prompt Engineering Instruction, knowledge, Lawyers Duties, Technology, VENDORS, wisdom | Permalink
Posted by Ralph Losey

2025 Year in Review: Beyond Adoption—Entering the Era of AI Entanglement and Quantum Law

December 31, 2025

Ralph Losey, December 31, 2025

As I sit here reflecting on 2025—a year that began with the mind-bending mathematics of the multiverse and ended with the gritty reality of cross-examining algorithms—I am struck by a singular realization. We have moved past the era of mere AI adoption. We have entered the era of entanglement, where we must navigate the new physics of quantum law using the ancient legal tools of skepticism and verification.

A split image illustrating two concepts: on the left, 'AI Adoption' showing an individual with traditional tools and paperwork; on the right, 'AI Entanglement' featuring the same individual surrounded by advanced technology and integrated AI systems. — In 2025 we moved from AI Adoption to AI Entanglement. All images by Losey using many AIs.

We are learning how to merge with AI and remain in control of our minds, our actions. This requires human training, not just AI training. As it turns out, many lawyers are well prepared by past legal training and skeptical attitude for this new type of human training. We can quickly learn to train our minds to maintain control while becoming entangled with advanced AIs and the accelerated reasoning and memory capacities they can bring.

In 2024, we looked at AI as a tool, a curiosity, perhaps a threat. By the end of 2025, the tool woke up—not with consciousness, but with “agency.” We stopped typing prompts into a void and started negotiating with “agents” that act and reason. We learned to treat these agents not as oracles, but as ‘consulting experts’—brilliant but untested entities whose work must remain privileged until rigorously cross-examined and verified by a human attorney. That put the human legal minds in control and stops the hallucinations in what I called “H-Y-B-R-I-D” workflows of the modern law office.

We are still way smarter than they are and can keep our own agency and control. But for how long? The AI abilities are improving quickly but so are our own abilities to use them. We can be ready. We must. To stay ahead, we should begin the training in earnest in 2026.

Here is my review of the patterns, the epiphanies, and the necessary illusions of 2025.

I. The Quantum Prelude: Listening for Echoes in the Multiverse

We began the year not in the courtroom, but in the laboratory. In January, and again in October, we grappled with a shift in physics that demands a shift in law. When Google’s Willow chip in January performed a calculation in five minutes that would take a classical supercomputer ten septillion years, it did more than break a speed record; it cracked the door to the multiverse. Quantum Leap: Google Claims Its New Quantum Computer Provides Evidence That We Live In A Multiverse (Jan. 2025).

The scientific consensus solidified in October when the Nobel Prize in Physics was awarded to three pioneers—including Google’s own Chief Scientist of Quantum Hardware, Michel Devoret—for proving that quantum behavior operates at a macroscopic level. Quantum Echo: Nobel Prize in Physics Goes to Quantum Computer Trio (Two from Google) Who Broke Through Walls Forty Years Ago; and Google’s New ‘Quantum Echoes Algorithm’ and My Last Article, ‘Quantum Echo’ (Oct. 2025).

For lawyers, the implication of “Quantum Echoes” is profound: we are moving from a binary world of “true/false” to a quantum world of “probabilistic truth”. Verification is no longer about identical replication, but about “faithful resonance”—hearing the echo of validity within an accepted margin of error.

But this new physics brings a twin peril: Q-Day. As I warned in January, the same resonance that verifies truth also dissolves secrecy. We are racing toward the moment when quantum processors will shatter RSA encryption, forcing lawyers to secure client confidences against a ‘harvest now, decrypt later’ threat that is no longer theoretical.

We are witnessing the birth of Quantum Law, where evidence is authenticated not by a hash value, but by ‘replication hearings’ designed to test for ‘faithful resonance.’ We are moving toward a legal standard where truth is defined not by an identical binary match, but by whether a result falls within a statistically accepted bandwidth of similarity—confirming that the digital echo rings true.

A digital display showing a quantum interference graph with annotations for expected and actual results, including a fidelity score of 99.2% and data on error rates and system status. — Quantum Replication Hearings Are Probable in the Future.

II. China Awakens and Kick-Starts Transparency

While the quantum future dangers gestated, AI suffered a massive geopolitical shock on January 30, 2025. Why the Release of China’s DeepSeek AI Software Triggered a Stock Market Panic and Trillion Dollar Loss. The release of China’s DeepSeek not only scared the market for a short time; it forced the industry’s hand on transparency. It accelerated the shift from ‘black box’ oracles to what Dario Amodei calls ‘AI MRI’—models that display their ‘chain of thought.’ See my DeepSeek sequel, Breaking the AI Black Box: How DeepSeek’s Deep-Think Forced OpenAI’s Hand. This display feature became the cornerstone of my later 2025 AI testing.

My Why the Release article also revealed the hype and propaganda behind China’s DeepSeek. Other independent analysts eventually agreed and the market quickly rebounded and the political, military motives became obvious.

A digital artwork depicting two armed soldiers facing each other, one representing the United States with the American flag in the background and the other representing China with the Chinese flag behind. Human soldiers are flanked by robotic machines symbolizing advanced military technology, set against a futuristic backdrop. — The Arms Race today is AI, tomorrow Quantum. So far, propaganda is the weapon of choice of AI agents.

III. Saving Truth from the Memory Hole

Reeling from China’s propaganda, I revisited George Orwell’s Nineteen Eighty-Four to ask a pressing question for the digital age: Can truth survive the delete key? Orwell feared the physical incineration of inconvenient facts. Today, authoritarian revisionism requires only code. In the article I also examine the “Great Firewall” of China and its attempt to erase the history of Tiananmen Square as a grim case study of enforced collective amnesia. Escaping Orwell’s Memory Hole: Why Digital Truth Should Outlast Big Brother

My conclusion in the article was ultimately optimistic. Unlike paper, digital truth thrives on redundancy. I highlighted resources like the Internet Archive’s Wayback Machine—which holds over 916 billion web pages—as proof that while local censorship is possible, global erasure is nearly unachievable. The true danger we face is not the disappearance of records, but the exhaustion of the citizenry. The modern “memory hole” is psychological; it relies on flooding the zone with misinformation until the public becomes too apathetic to distinguish truth from lies. Our defense must be both technological preservation and psychological resilience.

A graphic depiction of a uniformed figure with a Nazi armband operating a machine that processes documents, with an eye in the background and the slogan 'IGNORANCE IS STRENGTH' prominently displayed at the top. — Changing history to support political tyranny. Orwell’s warning.

Despite my optimism, I remained troubled in 2025 about our geo-political situation and the military threats of AI controlled by dictators, including, but not limited to, the Peoples Republic of China. One of my articles on this topic featured the last book of Henry Kissinger, which he completed with Eric Schmidt just days before his death in late 2024 at age 100. Henry Kissinger and His Last Book – GENESIS: Artificial Intelligence, Hope, and the Human Spirit. Kissinger died very worried about the great potential dangers of a Chinese military with an AI advantage. The same concern applies to a quantum advantage too, although that is thought to be farther off in time.

IV. Bench Testing the AI models of the First Half of 2025

I spent a great deal of time in 2025 testing the legal reasoning abilities of the major AI players, primarily because no one else was doing it, not even AI companies themselves. So I wrote seven articles in 2025 concerning benchmark type testing of legal reasoning. In most tests I used actual Bar exam questions that were too new to be part of the AI training. I called this my Bar Battle of the Bots series, listed here in sequential order:

Breaking the AI Black Box: A Comparative Analysis of Gemini, ChatGPT, and DeepSeek. February 6, 2025
Breaking New Ground: Evaluating the Top AI Reasoning Models of 2025. February 12, 2025
Bar Battle of the Bots – Part One. February 26, 2025
Bar Battle of the Bots – Part Two. March 5, 2025
New Battle of the Bots: ChatGPT 4.5 Challenges Reigning Champ ChatGPT 4o. March 13, 2025
Bar Battle of the Bots – Part Four: Birth of Scorpio. May 2025
Bots Battle for Supremacy in Legal Reasoning – Part Five: Reigning Champion, Orion, ChatGPT-4.5 Versus Scorpio, ChatGPT-o3. May 2025.

Two humanoid robots fighting against each other in a boxing ring, surrounded by a captivated audience. — Battle of the legal bots, 7-part series.

The test concluded in May when the prior dominance of ChatGPT-4o (Omni) and ChatGPT-4.5 (Orion) was challenged by the “little scorpion,” ChatGPT-o3. Nicknamed Scorpio in honor of the mythic slayer of Orion, this model displayed a tenacity and depth of legal reasoning that earned it a knockout victory. Specifically, while the mighty Orion missed the subtle ‘concurrent client conflict’ and ‘fraudulent inducement’ issues in the diamond dealer hypothetical, the smaller Scorpio caught them—proving that in law, attention to ethical nuance beats raw processing power. Of course, there have been many models released since then May 2025 and so I may do this again in 2026. For legal reasoning the two major contenders still seem to be Gemini and ChatGPT.

Aside for legal reasoning capabilities, these tests revealed, once again, that all of the models remained fundamentally jagged. See e.g., The New Stanford–Carnegie Study: Hybrid AI Teams Beat Fully Autonomous Agents by 68.7% (Sec. 5 – Study Consistent with Jagged Frontier research of Harvard and others). Even the best models missed obvious issues like fraudulent inducement or concurrent conflicts of interest until pushed. The lesson? AI reasoning has reached the “average lawyer” level—a “C” grade—but even when it excels, it still lacks the “superintelligent” spark of the top 3% of human practitioners. It also still suffers from unexpected lapses of ability, living as all AI now does, on the Jagged Frontier. This may change some day, but we have not seen it yet.

A stylized illustration of a jagged mountain range with a winding path leading to the peak, set against a muted blue and beige background, labeled 'JAGGED FRONTIER.' — See Harvard Business School’s *Navigating the Jagged Technological Frontier* and my humble papers, *From Centaurs To Cyborgs*, and *Navigating the AI Frontier.*

V. The Shift to Agency: From Prompters to Partners

If 2024 was the year of the Chatbot, 2025 was the year of the Agent. We saw the transition from passive text generators to “agentic AI”—systems capable of planning, executing, and iterating on complex workflows. I wrote two articles on AI agents in 2025. In June, From Prompters to Partners: The Rise of Agentic AI in Law and Professional Practice and in November, The New Stanford–Carnegie Study: Hybrid AI Teams Beat Fully Autonomous Agents by 68.7%.

Agency was mentioned in many of my other articles in 2025. For instance, in my June and July as part of my release the ‘Panel of Experts’—a free custom GPT tool that demonstrated AI’s surprising ability to split into multiple virtual personas to debate a problem. Panel of Experts for Everyone About Anything, Part One and Part Two and Part Three .Crucially, we learned that ‘agentic’ teams work best when they include a mandatory ‘Contrarian’ or Devil’s Advocate. This proved that the most effective cure for AI sycophancy—its tendency to blindly agree with humans—is structural internal dissent.

By the end of 2025 we were already moving from AI adoption to close entanglement of AI into our everyday lives

An artistic representation of a human hand reaching out to a robotic hand, signifying the concept of 'entanglement' in AI technology, with the year 2025 prominently displayed. — Close hybrid multimodal methods of AI use were proven effective in 2025 and are leading inexorably to full AI entanglement.

This shift forced us to confront the role of the “Sin Eater”—a concept I explored via Professor Ethan Mollick. As agents take on more autonomous tasks, who bears the moral and legal weight of their errors? In the legal profession, the answer remains clear: we do. This reality birthed the ‘AI Risk-Mitigation Officer‘—a new career path I profiled in July. These professionals are the modern Sin Eaters, standing as the liability firewall between autonomous code and the client’s life, navigating the twin perils of unchecked risk and paralysis by over-regulation.

But agency operates at a macro level, too. In June, I analyzed the then hot Trump–Musk dispute to highlight a new legal fault line: the rise of what I called the ‘Sovereign Technologist.’ When private actors control critical infrastructure—from satellite networks to foundation models—they challenge the state’s monopoly on power. We are still witnessing a constitutional stress-test where the ‘agency’ of Tech Titans is becoming as legally disruptive as the agents they build.

As these agents became more autonomous, the legal profession was forced to confront an ancient question in a new guise: If an AI acts like a person, should the law treat it like one? In October, I explored this in From Ships to Silicon: Personhood and Evidence in the Age of AI. I traced the history of legal fictions—from the steamship Siren to modern corporations—to ask if silicon might be next.

While the philosophical debate over AI consciousness rages, I argued the immediate crisis is evidentiary. We are approaching a moment where AI outputs resemble testimony. This demands new tools, such as the ALAP (AI Log Authentication Protocol) and Replication Hearings, to ensure that when an AI ‘takes the stand,’ we can test its veracity with the same rigor we apply to human witnesses.

VI. The New Geometry of Justice: Topology and Archetypes

To understand these risks, we had to look backward to move forward. I turned to the ancient visual language of the Tarot to map the “Top 22 Dangers of AI,” realizing that archetypes like The Fool (reckless innovation) and The Tower (bias-driven collapse) explain our predicament better than any white paper. See, Archetypes Over Algorithms; Zero to One: A Visual Guide to Understanding the Top 22 Dangers of AI. Also see, Afraid of AI? Learn the Seven Cardinal Dangers and How to Stay Safe.

But visual metaphors were only half the equation; I also needed to test the machine’s own ability to see unseen connections. In August, I launched a deep experiment titled Epiphanies or Illusions? (Part One and Part Two), designed to determine if AI could distinguish between genuine cross-disciplinary insights and apophenia—the delusion of seeing meaningful patterns in random data, like a face on Mars or a figure in toast.

I challenged the models to find valid, novel connections between unrelated fields. To my surprise, they succeeded, identifying five distinct patterns ranging from judicial linguistic styles to quantum ethics. The strongest of these epiphanies was the link between mathematical topology and distributed liability—a discovery that proved AI could do more than mimic; it could synthesize new knowledge

This epiphany lead to investigation of the use of advanced mathematics with AI’s help to map liability. In The Shape of Justice, I introduced “Topological Jurisprudence”—using topological network mapping to visualize causation in complex disasters. By mapping the dynamic links in a hypothetical we utilized topology to do what linear logic could not: mathematically exonerate the innocent parties. The topological map revealed that the causal lanes merged before the control signal reached the manufacturer’s product, proving the manufacturer had zero causal connection to the crash despite being enmeshed in the system. We utilized topology to do what linear logic could not: mathematically exonerate the innocent parties in a chaotic system.

A person in a judicial robe stands in front of a glowing, intricate, knot-like structure representing complex data or ideas, symbolizing the intersection of law and advanced technology. — Topological Jurisprudence: the possible use of AI to find order in chaos with higher math. Click here to see YouTube video introduction.

VII. The Human Edge: The Hybrid Mandate

Perhaps the most critical insight of 2025 came from the Stanford-Carnegie Mellon study I analyzed in December: Hybrid AI teams beat fully autonomous agents by 68.7%.

This data point vindicated my long-standing advocacy for the “Centaur” or “Cyborg” approach. This vindication led to the formalization of the H-Y-B-R-I-D protocol: Human in charge, Yield programmable steps, Boundaries on usage, Review with provenance, Instrument/log everything, and Disclose usage. This isn’t just theory; it is the new standard of care.

My “Human Edge” article buttressed the need for keeping a human in control. I wrote this in January 2025 and it remains a persona favorite. The Human Edge: How AI Can Assist But Never Replace. Generative AI is a one-dimensional thinking tool My ‘Human Edge’ article buttressed the need for keeping a human in control… AI is a one-dimensional thinking tool, limited to what I called ‘cold cognition’—pure data processing devoid of the emotional and biological context that drives human judgment. Humans remain multidimensional beings of empathy, intuition, and awareness of mortality.

AI can simulate an apology, but it cannot feel regret. That existential difference is the ‘Human Edge’ no algorithm can replicate. This self-evident claim of human edge is not based on sentimental platitudes; it is a measurable performance metric.

I explored the deeper why behind this metric in June, responding to the question of whether AI would eventually capture all legal know-how. In AI Can Improve Great Lawyers—But It Can’t Replace Them, I argued that the most valuable legal work is contextual and emergent. It arises from specific moments in space and time—a witness’s hesitation, a judge’s raised eyebrow—that AI, lacking embodied awareness, cannot perceive.

We must practice ‘ontological humility.’ We must recognize that while AI is a ‘brilliant parrot’ with a photographic memory, it has no inner life. It can simulate reasoning, but it cannot originate the improvisational strategy required in high-stakes practice. That capability remains the exclusive province of the human attorney.

Consistent with this insight, I wrote at the end of 2025 that the cure for AI hallucinations isn’t better code—it’s better lawyering. Cross-Examine Your AI: The Lawyer’s Cure for Hallucinations. We must skeptically supervise our AI, treating it not as an oracle, but as a secret consulting expert. As I warned, the moment you rely on AI output without verification, you promote it to a ‘testifying expert,’ making its hallucinations and errors discoverable. It must be probed, challenged, and verified before it ever sees a judge. Otherwise, you are inviting sanctions for misuse of AI.

VII. Conclusion: Guardians of the Entangled Era

As we close the book on 2025, we stand at the crossroads described by Sam Altman and warned of by Henry Kissinger. We have opened Pandora’s box, or perhaps the Magician’s chest. The demons of bias, drift, and hallucination are out, alongside the new geopolitical risks of the “Sovereign Technologist.” But so is Hope. As I noted in my review of Dario Amodei’s work, we must balance the necessary caution of the “AI MRI”—peering into the black box to understand its dangers—with the “breath of fresh air” provided by his vision of “Machines of Loving Grace.” promising breakthroughs in biology and governance.

The defining insight of this year’s work is that we are not being replaced; we are being promoted. We have graduated from drafters to editors, from searchers to verifiers, and from prompters to partners. But this promotion comes with a heavy mandate. The future belongs to those who can wield these agents with a skeptic’s eye and a humanist’s heart.

We must remember that even the most advanced AI is a one-dimensional thinking tool. We remain multidimensional beings—anchored in the physical world, possessed of empathy, intuition, and an acute awareness of our own mortality. That is the “Human Edge,” and it is the one thing no quantum chip can replicate.

Let us move into 2026 not as passive users entangled in a web we do not understand, but as active guardians of that edge—using the ancient tools of the law to govern the new physics of intelligence