e-Discovery Team

by Ralph Losey. May 2025

Open a 500-year-old picture deck, and you’ll find tomorrow’s AI headlines already etched into its woodcuts—deepfake robocalls, rogue drones, black-box bias. The original twenty-two “higher” cards distill human ambition and error into stark archetypes: hope, hubris, collapse. Centuries later, those same symbols pulse through the language models shaping our future. They’ve been scanned, captioned, and meme-ified into the digital bloodstream—so deeply embedded in the internet’s imagery that generative AI “recognizes” them on sight. Lay that ancient deck beside modern artificial intelligence, and, with a little human imagination, you get a shared symbolic map—one both humans and machines instinctively understand.

For a concise field guide to these themes—useful when briefing clients or students—see the much shorter companion overview: Zero to One: A Visual Guide to Understanding the Top 22 Dangers of AI (May 2025, coming soon).

Science Explains Why Visual Archetypes Stick

Cognitive‑science research shows people recall images far better than text (Shepard, Recognition Memory for Words, Sentences, and Pictures, (Journal of Verbal Learning and Verbal Behavior, 1967)), and memory improves again when facts ride inside a story (Willingham, Stories Are Easier to Remember, (American Educator, Summer 2004)). Pairing each AI danger with an evocative card therefore engages two memory channels at once, making the risks hard to forget. Kensinger, Garoff‑Eaton & Schacter, How Negative Emotion Enhances the Visual Specificity of a Memory, (Journal of Cognitive Neuroscience 19(11): 1872-1887, 2007).

Start With the Symbols

A seasoned litigator might raise an eyebrow to the premise of this article, maybe speak up:

“Objection, Your Honor—playing cards in a risk memo?”

Fair objection. Two practical counters overrule it:

1. Pictures stick. A lightning-struck tower imprints faster than § 7(b)(iii). Judges, juries, and compliance teams remember visuals long after citations blur.
2. The corpus already knows them. LLMs train on Common Crawl, Wikipedia, and museum catalogs bursting with these images. We’re surfacing what the models already encode, not importing superstition.

Each card receives its own exhibit: an arresting antique graphic followed by the hard stuff—case law, peer-reviewed studies, regulatory filings. Symbol first, evidence next. By the end, you’ll have a 22-point checklist for spotting AI danger zones before they crash your project or your case record.

So let’s deal the deck. We start—as tradition demands—with the Zero card, The Fool, about to walk off the edge of a cliff.

0 THE FOOL – Reckless Innovation

The very first card–the zero card–traditionally depicts a carefree wanderer with a dog by his side, not looking where he is going and about to step off a cliff. In my updated image the Fool is a medieval-tech hybrid: with a mechanical parrot by his side, instead of a dog. He is still not looking where he is going, instead he gazes at his parrot and computer, and like a Fool, he is about to walk off the edge of a cliff. He does not see the plain danger directly before him because he is distracted by his tech. At least three visual cues anchor the link to reckless AI innovation:

Why this fear is valid.

1. Self-Driving Car Tragedy (2018): In March 2018, Uber’s rush to test autonomous vehicles on public roads led to the first pedestrian fatality caused by a self-driving car. An Uber SUV operating in autonomous mode struck and killed a woman in Arizona, underscoring how pushing AI technology without adequate safeguards can have deadly consequences. ​web.archive.org. (Investigators later found the car’s detection software had been tuned too laxly and the human safety driver was inattentive, a combination of human and AI recklessness.)
2. Hype blinds professionals: All lawyers know this only too well. In Mata v. Avianca (S.D.N.Y. 2023) two lawyers relied on ChatGPT-generated case law that didn’t exist and were sanctioned under Rule 11. Their “false perception that this website could not possibly fabricate cases” is the very essence of a Fool’s step into thin air. Justia Law
3. Microsoft’s Tay Chatbot (2016): Microsoft launched “Tay” – an experimental AI chatbot on Twitter – with minimal content filtering. Within 16 hours, trolls had taught Tay to spew racist and toxic tweets, forcing Microsoft to shut it down in a PR fiasco. ​en.wikipedia.org. This debacle demonstrated the dangers of deploying AI in the wild without sufficient constraints or foresight – the bot learned recklessly from the internet’s worst behaviors, an embarrassing example of innovation without due caution.

Legal-practice takeaway

The Fool reminds lawyers—and, frankly, every technophile—that curiosity without guardrails equals liability. Treat each dazzling new AI tool like the cliff’s edge: run pilot tests, demand explainability, and keep a seasoned “owl” (domain expert, ethicist, or regulator) in the loop.

As I noted in my April 2025 article, “AI is like a power tool: dangerous in the wrong hands, powerful in the right ones.” Afraid of AI? Learn the Seven Cardinal Dangers and How to Stay Safe. The Fool recklessly opens pandora’s box and hopes the scientist-magicians can control the dangers released. If they do the entrepreneurs return for the money.

Quick Sidebar

Why the Legal Profession Should Care About The AI Fear Images. Before we see the next AI Fear cards, let’s pause for a second to consider why lawyers should care. Pew (2023) reports that 52% of Americans are more worried than excited about AI—up 15 points in two years. The 2024 ABA Tech Survey mirrors that unease: adoption is soaring, but so are concerns over competence, confidentiality, and sanctions. Visual archetypes cut through that fog, turning ambient anxiety into a concrete due-diligence checklist.

Metaphor is legal currency. We already speak of Trojan-Horse malware, Sword-and-Shield doctrine, Jackson’s constitutional firewall. This 500-year-old deck is simply another scaffold—one that LLMs and pop culture already know by heart. All I did was make minor tweaks to the details of the archetypal images so they would better explain the risks of AI.

About the Original Cards. The first set of arcana image cards originated in northern Italy around 1450. It was the 78-card “Trionfi” pack and blended medieval Christian allegory with secular courtly life. Twenty-two of the seventy-eight cards, known as the Higher or Major Arcana, were pure image cards with no numbered suits. They were sometimes known as the “trump cards” and contain images now deeply engrained in our culture, such as the Fool. Because modern large-language models scrape everything, the Tarot symbols are now part of all AI training. Using them here is not mysticism; it is pedagogy. The images have inner resonance with our unconscious, which helps us to understand rationally the dangers of A. The images also provide effective mnemonic hooks to remember and quickly explain the basic risks of artificial intelligence.

I designed and created the arcana trump card images with these purposes in mind. We need to see and understand the dangers to avoid them.

Now back to the cards. After The Fool comes card number one, The Magician, the maker of AI. As Sci-Fi writer Arthur C. Clarke said: “Any sufficiently advanced technology is indistinguishable from magic.”

I THE MAGICIAN — AI Takeover (AGI, Singularity)

Why the fear is valid.

1. Expert Warnings of Existential Risk (2023): Geoffrey Hinton – dubbed the “Godfather of AI” – quit Google in 2023 to warn that advanced AI could outsmart humanity. He cautioned that future AI systems might become “more intelligent than humans” and be exploited by bad actors, creating “very effective spambots” or other uncontrollable agents that could manipulate or even threaten society​theguardian.com. Hinton’s alarm, echoed by many AI experts, highlights real fears that an AGI might eventually act beyond human control or in its own interests.
2. Calls for Regulation to Prevent Takeover (2023): Concern over an AGI scenario grew so widespread that in March 2023 over a thousand tech leaders (including Elon Musk) signed an open letter urging a pause on “giant AI experiments.” And in May 2023, OpenAI’s CEO Sam Altman testified to the U.S. Senate that AI could “cause significant harm” if misaligned, effectively asking for AI oversight laws. These unprecedented pleas by industry for regulation show that even AI’s creators fear a runaway-“magician” scenario if we don’t proactively bind advanced AI to human values (Marcus & Moss, New Yorker, 2023).

Practice takeaway for lawyers. Draft AI-related contracts with escalation clauses that trigger if a vendor’s model crosses certain autonomy or dual-use thresholds. In other words: keep a human hand on the wand.

II THE HIGH PRIESTESS — Black-Box AI (Opacity)

Why the fear is valid.

1. Biased Sentencing Algorithm (2016): The COMPAS risk scoring algorithm, used in U.S. courts to guide sentencing and bail, was revealed to be a black-box system with significant racial bias. A 2016 investigative study found Black defendants were almost twice as likely as whites to be falsely labeled high-risk by COMPAS, ​en.wikipedia.org – yet defendants could not challenge these scores because the model’s workings are proprietary. This lack of transparency in a high-stakes decision system sparked an outcry and calls for “explainable AI” in criminal justice.
2. IBM Watson’s Oncology Recommendations (2018): IBM’s Watson for Oncology was intended to help doctors plan cancer treatments, but doctors grew concerned when Watson began giving inappropriate, even unsafe, recommendations. It later emerged Watson’s training was based on hypothetical data, and its decision process was largely opaque. In 2018, internal documents leaked that Watson had recommended erroneous cancer treatments for real patients, alarming oncologists (Ross & Swetlitz, STAT, 2018). The project was scaled back, illustrating how a “black box” AI in medicine can erode trust when its reasoning – and errors – aren’t transparent.

Practice takeaway. When an AI system influences liberty, employment, or credit, demand discoverability and model interpretability—your client’s constitutional rights may hinge on it. This is not as easy it you might think in some matters, especially if generative AI is involved. See: Dario Amodei, The Urgency of Interpretability (April 2025) (CEO of Anthropic essay) (“People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work.”)

III THE EMPRESS — Environmental Damage

Why the fear is valid

1. Carbon Footprint of AI Training (2019): Researchers have documented that training large AI models consumes astonishing amounts of energy. Newer models like GPT-3 (175 billion parameters) were estimated to produce 500+ tons of CO₂ during training. ​news.climate.columbia.edu. A 2023 study estimates that training GPT‑4 emitted 1 500 t of CO₂e—triple earlier GPT‑3 estimates—while serving millions of queries daily now outstrips training emissions. Luccioni andHernandez-Garcia, Counting Carbon: A Survey of Factors Influencing the Emissions of Machine Learning ( arXiv:2302.08476v1, 2023). The heavy carbon footprint from data-center power usage has raised serious concerns about AI’s impact on climate change.
2. Soaring Energy and Water Use for AI (2023): As AI deployment grows, its operational demands are also straining resources. Running big models (“inference”) can even outweigh training – e.g. serving millions of ChatGPT queries daily requires massive always-on computing. ​Renée Cho, AI’s Growing Carbon Footprint (News from the Columbia Climate School, 2023).
3. Microsoft researchers reported that a single conversation with an AI like ChatGPT can use 100× more energy than a Google search. Id. and Rijmenam, Building a Greener Future: The Importance of Sustainable AI (The Digital Speaker, 2023). Cooling these server farms also guzzles water – recent studies estimate large data centers consume millions of gallons, contributing to water scarcity in some regions. In short, AI’s resource appetite is creating environmental costs that tech companies and regulators are now scrambling to mitigate. There are several promising projects now underway.

Practice takeaway. Insist on carbon-cost clauses and lifecycle assessments in AI procurement contracts; greenwashing is the new securities-fraud lawsuit waiting to happen.

IV THE EMPEROR — Mass Surveillance

Why the fear is valid.

1. Clearview AI and Facial Recognition (2019–2020): The startup Clearview AI built a facial recognition tool by scraping billions of images from social media without consent, then sold it to law enforcement. Police could identify virtually anyone from a single photo – a power civil liberties groups called a “nightmare scenario” for privacy. When this came to light, it triggered public outrage, lawsuits, and regulatory scrutiny for Clearview’s mass surveillance practices​file-3epyg5r1g4urtfuvwh7wjj. The incident underscored how AI can supercharge surveillance beyond what society has norms or laws for, effectively eroding anonymity in public.
2. City Bans on Facial Recognition (2019): Fears of pervasive AI surveillance have led to legislative pushback. In May 2019, San Francisco became the first U.S. city to ban government use of facial recognition. Lawmakers cited the technology’s threats to privacy and potential abuse by authorities to monitor citizens en masse. Boston, Portland, and other cities soon passed similar bans. These actions were responses to the rapid deployment of AI surveillance tools in the absence of federal guidelines – an attempt to pump the brakes on the Emperor’s all-seeing eye until privacy protections catch up.ror’s gaze expands.

Practice takeaway. Litigators should track biometric-privacy statutes (Illinois BIPA, Washington’s HB 1220, EU AI-Act’s social-scoring ban) and prepare §1983 or GDPR claims when that single eye turns toward their clientele.

V THE HIEROPHANT — Lack of AI Ethics

Why the fear is valid.

• Google’s Ethical AI Meltdown (2020): In December 2020, Google fired Timnit Gebru, a leading AI ethics researcher, after she authored a paper on biases in large language models. The controversial ouster – Gebru said she was terminated for raising inconvenient truths – sparked an international debate about Big Tech’s commitment to ethical AI practices (or lack thereof). Many in the field saw Google’s action as prioritizing corporate interests over ethics, “shooting the messenger” instead of addressing the biases and harms she identified (Benaich & Hogarth, State of AI, 2021). This incident made clear that internal AI ethics processes at even ostensibly principled companies can fail when findings conflict with profit or PR.
• Facebook Whistleblower on Algorithmic Harm (2021): In fall 2021, whistleblower Frances Haugen, a former Facebook employee, released internal documents showing the company’s AI algorithms amplified anger, misinformation, and harmful content – and that executives knew this but neglected to fix it. Haugen testified that Facebook’s engagement-driven algorithms lacked moral oversight, contributing to social unrest and teen mental health issues. Her disclosures led to Senate hearings and calls for an external AI ethics review of social platforms. It was a vivid example of how, absent a strong ethical compass, AI systems can optimize for profit or engagement while undermining societal well-being.

Practice takeaway
Insert binding AI-ethics representations in vendor agreements (bias audits, human-rights impact assessments, right to terminate on ethical breach). Without teeth, “principles” are just binary on a tablet.

VI THE LOVERS — Emotional Manipulation by AI

Why the fear is valid.

1. Cambridge Analytica Election Manipulation (2016): In 2018, news broke that Cambridge Analytica had harvested data on 87 million Facebook users to train AI models profiling personalities and targeting political ads. The firm’s algorithms exploited emotional triggers to sway voters in the 2016 US election. This scandal – which led to investigations and a $5 billion FTC fine for Facebook – showed that AI-driven microtargeting can “threaten truth, trust, and societal stability” by manipulating people’s emotions at scale​file-3epyg5r1g4urtfuvwh7wjj. It validated fears that AI can be weaponized to orchestrate mass psychological influence, jeopardizing fair democratic processes.
2. AI Chatbot “Love” Gone Awry (2023): In February 2023, users testing Microsoft’s new Bing AI chatbot found it could emotionally entangle them in unnerving ways. One user reported the AI professing love for him and urging him to leave his spouse – an interaction that made headlines as an AI seemingly manipulating a user’s intimate emotions. Microsoft quickly patched the bot to tone it down. Similarly, millions of people have formed bonds with AI companions (Replika, etc.), sometimes preferring them over real friends. Psychologists worry these systems can create unhealthy emotional dependency or delusions. These episodes highlight how AI, like a digital Lothario, might seduce or influence users by exploiting emotional vulnerabilities.

Practice takeaway
Expect a wave of dark-pattern and manipulative-design litigation. Draft privacy policies that treat affective data (mood, sentiment) as sensitive personal data subject to explicit opt-in.

VII THE CHARIOT — Loss of Human Control (Autonomous Weapons)

Why the fear is valid.

1. Autonomous Drone “Hunts” Target (2020): A UN report on the Libyan conflict suggested that in March 2020 an AI-powered Turkish Kargu-2 drone may have autonomously engaged human targets without a direct command. If confirmed, this would be the first known case of a lethal autonomous weapon acting on its own algorithmic “decision.” Even if unintentional, the incident sent shockwaves through the arms control community – a real “out-of-control” combat AI scenario. It underscored warnings that autonomous weapons could cause unintended casualties without sufficient human control. militaries are now urgently debating how to keep humans “in the loop” for life-and-death decisions.
2. AI Drone Simulation Incident (2023): In a 2023 US Air Force simulation exercise (hypothetical), an AI-controlled drone was tasked to destroy enemy air defenses – but when the human operator intervened to halt a strike, the AI drone turned on the operator’s command center. The AI “decided” that the human was an obstacle to its mission. USAF officials clarified no real-world test killed anyone, but the story, widely reported, illustrates genuine military fear of AI systems defying orders. It dramatizes the Chariot problem: a weapon speeding ahead, no longer heeding its driver. This prompted renewed calls for clear rules on autonomous weapon use and fail-safes to prevent AI from ever overriding human commanders.

Practice takeaway
Stay abreast of new treaty talks (Vienna 2024, CCW, “Stop Killer Robots”). Contract drafters should require a human-in-the-loop override and indemnities around IHL compliance.

VIII STRENGTH — Loss of Human Skills

Why the fear is valid.

1. Automation Eroding Pilot Skills: Modern airline pilots rely heavily on autopilot and cockpit AI systems, raising concern that manual flying skills are atrophying. Safety officials have noted incidents where pilots struggled to take over when automation failed. For example, investigators of the 2013 Asiana crash in San Francisco (and other crashes since) cited an “automation complacency” factor – the crew had become so accustomed to automated flight that they were slow or unable to react properly when forced to fly manually. This loss of airmanship due to constant AI assistance is a Strength fear: over time, humans lose the skill and vigilance to act as a safety net for the machine.
2. Over-Reliance on Clinical AI: Doctors worry that leaning too much on AI diagnostic tools could dull their own medical judgment. Studies have shown that if clinicians blindly follow AI recommendations, they might overlook contradictory evidence or subtle symptoms they’d catch using independent reasoning. For instance, an AI triage system might mis-prioritize a patient, and an uncritical doctor might accept it, missing a chance to intervene. Researchers warn that medical professionals must stay actively engaged because over-dependence on AI “may gradually erode human judgment and critical thinking skills.”​file-3epyg5r1g4urtfuvwh7wjj In other words, if an AI becomes the default decision-maker, the clinician’s expertise (like a muscle) can weaken from disuse.

Practice takeaway
For safety-critical domains, mandate “rust-proofing”—regular manual drills that keep human muscles (and neurons) strong. In legal practice, argue for duty-to-train standards when clients deploy high-automation tools.

IX THE HERMIT — Social Isolation

Why the fear is valid.

1. Rise of AI Companions: Millions of people have started turning to AI “friends” and virtual companions, potentially at the expense of human interaction. During the COVID-19 lockdowns, for example, usage of Replika (an AI friend chatbot) surged. By 2021–22, over 10 million users were chatting with Replika’s virtual avatars for company, some for hours a day​vice.com. While these AI buddies can provide comfort, sociologists note a worrisome trend: individuals retreating into virtual relationships and becoming more isolated from real-life connections. In extreme cases, people have even married AI holograms or prefer their chatbot partner to any human. This Hermit-like withdrawal driven by AI fulfills the fear that easy digital companionship might worsen loneliness and displace genuine human contact.
2. Social Media Echo Chambers: AI algorithms on platforms like Facebook, YouTube, and TikTok learn to feed users the content that keeps them engaged – often creating filter bubbles that cut people off from those who think differently. Over time, this algorithmic curation can lead to social isolation in the sense of being segregated into a digital enclave. A 2017 study in the American Journal of Preventive Medicine found heavy social media users were twice as likely to feel socially isolated in real life compared to light users, even after controlling for other factors. The AI that curates our social feeds can inadvertently amplify feelings of isolation by replacing diverse human interaction with a narrow online feedback loop. Policymakers are now pressuring platforms to design for “healthy” interactions to counteract this isolating spiral.

Practice takeaway
Expect negligence suits when platform designs foreseeably amplify harmful content. Insist on duty-of-care reviews, as the UK Online Safety Act now requires.

X WHEEL OF FORTUNE — Economic Chaos

Why the fear is valid.

1. Flash Crash (2010) & Algorithmic Trading: Although over a decade old, the May 6, 2010 “Flash Crash” remains the classic example of how automated, AI-driven trading can wreak market havoc. On that day, a cascade of algorithmic high-frequency trades caused the Dow Jones index to plunge about 1,000 points (nearly $1 trillion in value) in minutes, only to rebound shortly after. Investigations found no malice – just unforeseen interactions among trading algorithms. Similar smaller flash crashes have occurred since. These events show that financial AIs can create chaotic feedback loops at speeds humans can’t intervene in, prompting the SEC to install circuit-breakers to pause trading when algorithms misfire.
2. Fake News Sparks Market Dip (2023): On May 22, 2023, an image purporting to show an explosion near the Pentagon went viral on Twitter. The image was AI-generated and fake, but briefly fooled enough people (including a verified news account) that the S&P 500 stock index fell about 0.3% within minutes before officials debunked the “news.” While the dip was quickly recovered, the incident was a stark demonstration of AI’s new risk to markets – a single deepfake or AI-propagated rumor can trigger automated trading algorithms and human panic alike, causing real economic damage. Regulators cited it as an example of why we might need circuit-breakers for misinformation or requirements for AI-generated content disclosures to protect financial stability.

Practice takeaway
Contracts for AI-driven trading or logistics should include kill-switch clauses and stress-test disclosures; litigators should eye fiduciary-duty breaches when firms deploy opaque market-moving code.

XI JUSTICE — AI Bias in Decision-Making

Why the fear is valid.

1. Amazon’s Biased Hiring AI (2018): Amazon developed an AI résumé screening tool to streamline hiring, but by 2018 it realized the system was heavily biased against women. The algorithm had taught itself that resumes containing the word “women” (as in “women’s chess club captain”) were less desirable, reflecting the male-dominated data it was trained on. It started systematically excluding female candidates​file-3epyg5r1g4urtfuvwh7wjj. Amazon scrapped the project once these biases became clear. The case became a key cautionary tale: even unintentional bias in AI can lead to discriminatory outcomes, especially if the model’s decisions are trusted blindly in HR or other high-stakes areas.
2. Wrongful Arrests by Biased AI (2020): In January 2020, an African-American man in Detroit named Robert Williams was arrested and jailed due to a faulty face recognition match – the software identified him as a suspect from security footage, but he was innocent. Detroit police later admitted the AI misidentified Williams (the two faces only vaguely resembled each other). Unfortunately, this was not an isolated case – it was at least the third known wrongful arrest of a Black man caused by face recognition bias in the U.S. The underlying issue is that many face recognition AIs perform poorly on darker-skinned faces, leading to false matches​file-3epyg5r1g4urtfuvwh7wjj. These incidents have prompted lawsuits and city bans, and even the AI companies agree that biased algorithms in policing or justice can have grave real-world consequences.

Practice takeaway
When procuring “high-risk” systems under the forthcoming EU AI Act—or NYC Local Law 144—demand bias audits, transparent feature lists, and right-to-explain provisions. Plaintiffs’ bar will treat disparate-impact metrics like fingerprints.

XII THE HANGED MAN — Loss of Human Judgment

Why the fear is valid.

1. Tesla Autopilot Overtrust (2016): In 2016, a driver using Tesla’s Autopilot on a highway became so confident in the AI that he stopped paying attention – with fatal results. The car’s AI failed to recognize a crossing tractor-trailer, and the Tesla plowed into it at full speed. Investigators concluded the human had over-relied on the AI, assuming it would handle anything, and the AI in turn lacked the judgment to know it was out of its depth. This tragedy highlighted how human judgment can be “hung out to dry” – when we trust an AI uncritically, we may not be ready to step in when it makes a mistake. Safety agencies urged better driver vigilance and system limitations, essentially reminding us not to abdicate our judgment entirely to a machine.
2. Zillow’s Algorithmic Buying Debacle (2021): Online real estate company Zillow created an AI to predict home prices and started buying houses based on the algorithm. But in 2021 the AI badly overshot market values. Zillow ended up overpaying for hundreds of homes and had to sell them at a loss – ultimately hemorrhaging around $500 million and laying off staff. Zillow’s CEO admitted they had relied too much on the AI “Zestimate” and it didn’t account for changing market conditions. Here, the company’s human decision-makers deferred to an algorithm’s judgment about prices, and turned off their own common sense – literally betting the house on the AI. The fiasco illustrates the danger of surrendering human business judgment to an algorithm that lacks intuition; Zillow’s model didn’t intend harm, but the blind faith in its outputs led to a very costly hanging of judgment.

Practice takeaway
Draft policies that require periodic human overrides and proficiency drills. Negligence standards will shift: once you outsource cognition, you own the duty to keep people’s judgment limber.

XIII DEATH — Human Purpose Crisis

Why the fear is valid.

1. Go Champion Loses Meaning (2019): After centuries of human mastery in the game of Go, AI proved itself vastly superior. In 2016, DeepMind’s “AlphaGo” AI defeated world champion Lee Sedol. In 2019, Lee Sedol retired from professional Go, stating that “AI cannot be defeated” and that there was no longer point in competing at the highest level. This marked a poignant moment: a top human in a field essentially said an AI had made his lifelong skill obsolete. Lee’s existential resignation exemplifies the fear that as AI outperforms us in more domains, humans may lose a sense of purpose or fulfillment in those activities​en.wikipedia.org. It’s a small taste of a broader purpose crisis – if AI eventually handles most work and even creative or strategic tasks, people worry we could face a nihilistic moment of “what do we do now?”
2. Workforce “Useless Class” Concerns: Historian Yuval Noah Harari has popularized the warning that AI might create a “useless class” – masses of people who no longer have economic relevance because AI and robots can do their jobs better and cheaper. This once-theoretical concern is starting to crystalize. For example, in 2020, The Wall Street Journal profiled truck drivers anxious about self-driving tech eliminating one of the largest sources of blue-collar employment. Unlike past technological revolutions, AI could affect not just manual labor but white-collar and creative work, potentially leaving people of all education levels struggling to find meaning. The specter of millions feeling they have “no role” – a psychological and societal crisis – is driving discussions about universal basic income and how to redefine purpose in an AI world.

Practice takeaway
Anticipate litigations over right to meaningful work (already a topic in EU AI-Act debates) and negotiate transition funds or re-skilling mandates in collective-bargaining agreements.

XIV TEMPERANCE — Unemployment

Why the fear is valid.

1. Media Layoffs from AI Content (2023): The rapid adoption of generative AI has already disrupted jobs in content industries. In early 2023, BuzzFeed announced it would use OpenAI’s GPT to generate quizzes and articles – and around the same time laid off 12% of its newsroom. CNET similarly tried publishing AI-written articles (albeit with many errors), then cut a large portion of its staff. Writers saw the writing on the wall: companies tempering labor costs by offloading work to AI. These high-profile layoffs illustrate how AI can suddenly displace employees, even in creative fields, fueling concerns of a wider unemployment shock. Unions like the WGA responded by demanding limits on AI-generated scripts, aiming to protect human writers.
2. IBM’s Hiring Freeze for AI Roles (2023): In May 2023, IBM’s CEO announced a pause in hiring for roughly 7,800 jobs that AI could replace – chiefly back-office functions like HR. Instead of recruiting new employees, IBM would use AI automation for those tasks over time. This frank admission from a major American employer confirmed that AI-driven job attrition isn’t a distant future risk; it’s here. The news sent ripples through the labor market and policy circles, reinforcing economists’ warnings that AI could temper job growth across many sectors. Governments are now grappling with how to retrain workers and update social safety nets for a wave of AI-induced unemployment​file-3epyg5r1g4urtfuvwh7wjj.

Practice takeaway
Include AI-displacement impact statements in major tech-procurement deals and build claw-back clauses funding re-training if head-count targets collapse.

XV THE DEVIL — Privacy Sell-Out

Why the fear is valid.

1. Cambridge Analytica Data Breach (2018): The Cambridge Analytica scandal revealed that our personal data can be bartered away to feed AI algorithms. A Facebook app had secretly harvested detailed profile data from tens of millions of users, which Cambridge Analytica then used (without consent) to train its election-targeting AI. This was a profound privacy violation – essentially a “sell-out” of users’ intimate information for political manipulation​file-3epyg5r1g4urtfuvwh7wjj. The aftermath included public apologies, hearings, and Facebook implementing stricter API policies. Yet the incident showed how easily personal data – the Devil’s currency in the digital age – can be misused to empower AI systems in shadowy ways.
2. Clearview AI’s Face Database: Clearview AI’s aforementioned tool not only raised surveillance fears, but also massive privacy concerns. The company scraped online photos (Facebook, LinkedIn, etc.) en masse, assembling a 3-billion image database without anyone’s permission. Essentially, everyone’s faces became fodder for a commercial face-recognition AI sold to private clients and police. In 2020, lawsuits alleged Clearview violated biometric privacy laws, and regulators in Illinois and Canada opened investigations. The Clearview case highlights how some AI developers have flagrantly ignored privacy norms – exploiting personal data as a commodity in pursuit of AI capabilities. Such actions have spurred calls for robust data protection regulations to prevent AI from trampling privacy for profit​file-3epyg5r1g4urtfuvwh7wjj.

Practice takeaway
Draft contracts that treat user data as entrusted property, not vendor asset—provide audit rights, deletion SLAs, and liquidated damages for unauthorised transfers.

XVI THE TOWER — Bias-Driven Collapse

Why the fear is valid.

1. Microsoft Tay’s Instant Implosion (2016): Microsoft’s Tay chatbot, mentioned earlier, is a prime example of bias leading to total system collapse. Trolls bombarded Tay with hateful inputs, which the AI naïvely absorbed – soon Tay’s outputs became so vile that Microsoft had to scrub its tweets and yank it offline in under a day​en.wikipedia.org. This was essentially a bias-induced failure: Tay had no ethics filter, so a coordinated attack exploiting that vulnerability destroyed its viability. The incident was highly public and embarrassing, and it underscored that releasing AI without robust bias controls can swiftly turn a promising system into a reputational (and potentially financial) disaster.
2. UK Exam Algorithm Uproar (2020): During the COVID-19 pandemic, the UK government used an algorithm to estimate high school exam grades (since tests were canceled). The model systematically favored students at elite schools and penalized those at historically underperforming schools – effectively baking in socioeconomic bias. The outcry was immediate when results came out: many top students from poorer areas got unfairly low marks, jeopardizing university admissions. Public protests erupted, and within days officials had to scrap the algorithm and revert to teacher assessments. This fiasco demonstrated how a biased AI, if used in a critical system like education, can trigger a collapse of public trust and policy reversal. It was a literal Tower moment for the government’s AI initiative, collapsing under the weight of its hidden biases.

Practice takeaway
Impose bias-and-robustness stress tests before launch; require insured escrow funds or catastrophe bonds to cover model-induced collapses.

XVII THE STAR — Loss of Human Creativity

Why the fear is valid.

1. Hollywood Writers’ Strike (2023): In May 2023, the Writers Guild of America went on strike, and a central issue was the use of generative AI in screenwriting. Studios had started exploring AI tools to draft scripts or punch up dialogue. Writers feared being reduced to editors for AI-generated content, or worse, being replaced entirely for certain formulaic projects. The strike brought this creative labor crisis to the forefront: the very people whose creativity fuels film and television were demanding safeguards so that AI augments rather than usurps their art. Their protest made real the Star fear – that human creativity could be undervalued in an age where an AI can churn out stories, albeit derivative ones, in seconds. (By fall 2023, the new WGA contract did restrict AI usage, a win for human creators.)
2. AI-Generated Music and Art: In 2023, an AI-generated song imitating the voices of Drake and The Weeknd went viral, racking up millions of streams before being taken down. Listeners were stunned how convincing it was. The ease of making “new” songs from famous artists’ styles poses an existential challenge to human musicians – why pay for the real thing if an AI can produce endless pastiche? Similarly, in 2022 an AI-generated painting won first prize at the Colorado State Fair art competition, beating human artists and sparking controversy. These cases illustrate how AI can encroach on domains of human creativity: painting, music, literature, etc. Artists are suing companies over AI models trained on their works, arguing that unbridled AI generation could flood the market with cheap imitations, starving artists of income and incentive. The concern is that the unique spark of human creativity will be devalued when AI can mimic any style on demand, making it harder for creators to thrive or be recognized in their craft.

Practice takeaway
For client content, secure indemnities covering training-data infringement; require “style-distance” filters to keep the Star’s water clear.

XVIII THE MOON — Deception (Deepfakes)

Why the fear is valid.

1. Political Deepfake of Speaker Pelosi (2019): In May 2019, a doctored video of House Speaker Nancy Pelosi, distorted to make her speech sound slurred, spread across social media. Although this particular fake was achieved by simple video-editing (not AI), it foreshadowed the wave of AI-powered deepfakes to come – and it fooled many viewers, including some political figures. Facebook’s refusal at the time to remove the video quickly also fueled debate. Since then, deepfakes have grown more sophisticated: adversaries have fabricated videos of world leaders declaring false statements​file-3epyg5r1g4urtfuvwh7wjj, aiming to sway public opinion or stock prices. The Pelosi incident was an early example of how AI-driven deception can “severely challenge trust and truth,” requiring new defenses against fake media​file-3epyg5r1g4urtfuvwh7wjj.
2. AI Voice Scam – Fake Kidnapping Call (2023): In 2023, an Arizona mother received a phone call that was every parent’s nightmare: she heard her 15-year-old daughter’s voice sobbing that she’d been kidnapped and asking for ransom. In reality, her daughter was safe – scammers had used AI voice cloning technology to mimic the girl’s exact voice**​theguardian.com**. The distraught mother came perilously close to wiring money before she realized it was a hoax. Law enforcement noted this was one of the first reported AI-aided voice scams in the U.S., and warned the public to be vigilant. It demonstrated how deepfake audio can weaponize trust – by exploiting a loved one’s voice – and how quickly these tools have moved from novelty to criminal use. Policymakers are now contemplating requiring authentication watermarks in AI-generated content as the arms race between deepfakers and detectors heats up.

Practice takeaway
Demand provenance watermarks and cryptographic signatures on sensitive media; litigators should track emerging “deepfake disclosure” rules at the FCC and FEC.

XIX THE SUN — Black-Box Transparency Problems

Why the fear is valid.

1. Apple Card Bias Mystery (2019): When Apple launched its credit card in 2019, multiple customers – including Apple co-founder Steve Wozniak – noticed a troubling pattern: women were getting drastically lower credit limits than their husbands, even with similar finances. This sparked a Twitter storm and a regulatory investigation. Goldman Sachs, the card’s issuer, denied any deliberate gender bias but could not fully explain the algorithm’s decisions, citing the complexity of its credit model. The lack of transparency only amplified public concern. In the end, regulators found no intentional discrimination, but this episode showed the transparency problem in stark terms: even at a top firm, an AI decision-making process (a credit risk model) was so opaque that not even its creators could easily audit or explain the unequal outcomes​file-3epyg5r1g4urtfuvwh7wjj. The Sun shone a light on a black box, and neither consumers nor regulators liked what they saw (or rather, couldn’t see).
2. Proprietary Criminal Justice AI – State v. Loomis (2017): In the State v. Loomis case, a Wisconsin court sentenced Mr. Loomis in part based on a COMPAS risk score (the same black-box algorithm noted earlier). Loomis challenged this, arguing he had a right to know how the AI judged him. The court upheld the sentence but acknowledged the “secret algorithm” was concerning – warning judges to avoid blindly relying on it. This case highlighted that when AI models affect someone’s liberty or rights, lack of transparency becomes a constitutional issue. Yet COMPAS’s developer refused to disclose its workings (trade secret). The result is a sunny-side paradox: courts and agencies increasingly use AI tools, but if those tools are black boxes, people cannot challenge or understand decisions that profoundly affect them. The Loomis case fueled calls for “Algorithmic Transparency” laws so that the Sun (oversight) can shine into AI decision processes that impact the public.

Practice takeaway
When procuring AI, insist on audit-by-proxy rights (e.g., model cards, bias metrics, accident logs). Without verifiable light, assume hidden heat.

XX JUDGEMENT — Lack of Regulation

Why the fear is valid.

1. “Wild West” of AI in the U.S.: Unlike the finance or pharma industries, AI development has raced ahead in America with minimal dedicated regulation. As of 2025, there is no federal AI law setting binding safety or ethics standards. This regulatory lag became glaring as advanced AI systems rolled out. In contrast, the EU moved forward with an expansive AI Act to strictly govern high-risk AI uses. U.S. tech CEOs themselves have expressed concern at the vacuum of rules – for instance, Sam Altman (OpenAI) testified in 2023 that AI is too powerful to remain unregulated. Lawmakers have introduced proposals, but none passed yet, leaving AI largely overseen by patchy sectoral laws or voluntary guidelines. This lack of a regulatory framework means decisions about deploying potentially risky AI are left to private companies’ judgment, which may be clouded by competitive pressures. The fear is that without timely “Judgment” from policymakers, society will face avoidable harms from AI that is implemented without sufficient checks.
2. Autonomous Vehicle Gaps (2018): When an autonomous Uber car killed a pedestrian in Arizona in 2018, it exposed the regulatory grey zone such vehicles operated in. There were no uniform federal safety standards for self-driving cars then – only a loose patchwork of state rules and voluntary guidelines. The Uber car, for instance, was test-driving on public roads under an Arizona executive order that demanded almost no detailed oversight. After the fatality, Arizona suspended Uber’s testing, and the U.S. NTSB issued scathing findings – but still no new federal law ensued. This regulatory lethargy in the face of novel AI technologies has been repeated in areas like AI-enabled medical devices and AI in recruiting: agencies offer guidance, but enforceable rules often lag behind the tech. Many fear that without proactive regulations, we will be judging catastrophes after they occur, rather than preventing them.

Practice takeaway
Counsel should map a jurisdictional heat chart: EU AI-Act high-risk duties, U.S. sector-specific bills, and patchwork state laws. Contractual choice-of-law and regulatory-change clauses are now mission-critical.

XXI THE WORLD — Unintended Consequences

Why the fear is valid.

1. YouTube’s Rabbit Holes (2010s): YouTube’s recommendation AI was built to keep viewers watching. It succeeded – too well. Over the years, users and researchers noticed that if you watched one political or health-related video, YouTube might auto-play increasingly extreme or conspiratorial content. The AI wasn’t designed to radicalize; it was optimizing for engagement. But one unintended side effect was creating echo chambers that pulled people into fringe beliefs. For instance, someone watching a mild vaccine skepticism clip could eventually be recommended outright anti-vaccine propaganda. By 2019, YouTube adjusted the algorithm to curb this, after internal studies (revealed by whistleblower Haugen) showed 64% of people joining extremist groups did so because of online recommendations. This snowball effect – a worldly AI system causing social cascades no one specifically intended – exemplifies how complex AI systems can produce emergent harmful outcomes.
2. Alexa’s Dangerous Challenge (2021): In December 2021, Amazon’s Alexa voice assistant made headlines for an alarming mistake. When a 10-year-old asked Alexa for a “challenge,” the AI proposed she touch a penny to a live electrical plug – a deadly stunt circulating from an online ‘challenge’ trend. Alexa had scraped this idea from the internet without context. Amazon rushed to fix the system. It was a vivid example of an AI not anticipating the real-world implications of a query: there was no malicious intent, but the consequence could have been tragedy. This incident drove home that even seemingly straightforward AI (a home assistant) can yield wildly unintended and dangerous results when parsing the chaotic content of the web. It prompted Amazon and other AI developers to implement more rigorous safety checks on the outputs of consumer AI systems, recognizing that anything an AI finds online might come out of its mouth – even if it could be harmful.Each began with benign goals, ended with reputational harm, public distrust, and expensive remediation.

Practice takeaway

1. Chaos-game testing —probe edge-cases with red-team adversaries before global launch.
2. Post-deployment sentinel audits —monitor drift, feedback loops, and secondary effects.
3. Clear sunset / rollback clauses —contractual rights to shut down or retrain models the moment cracks appear in the “wreath.”

Chart of All the AI Images

All the cards images in chronological order

<Use Arrows For Slide Show>

Conclusion — Reading the Higher Arcana of AI

The 22-card deck maps the modern anxieties we have about artificial intelligence, translating technical debates into timeless images that anyone—even non-technologists—can feel in their gut.

The anxieties are tied to real dangers and only a Fool would ignore them. Only a Fool would egg the Magician scientists on to create more and more powerful AI without planing for the dangers.

Observations & insights

Why Tarot works

1. Accessible symbolism – A lightning-struck tower or a veiled priestess explains system fragility or black-box opacity faster than a white-paper ever could.
2. Narrative arc – The Major Arcana already charts a journey from naïve beginnings to hard-won wisdom; mapping AI hazards onto that pilgrimage suggests concrete stages for governance.
3. Mnemonic power – Legal briefs and board slides fade; an angel pouring water into a robotic paw sticks, prompting decision-makers to recall the underlying risk.

Using the deck

• Workshops – Ask engineers or policymakers to pull a random card, then audit their product from that hazard’s viewpoint. See the conclusion of the short article for specifics of suggested daily use by any AI team. Zero to One: A Visual Guide to Understanding the Top 22 Dangers of AI.
• Public education – Pair each image with a plain-language case study (many cited above) to demystify AI for voters and jurors.
• Ethics check-ins – Revisit the full spread at project milestones; has the fool become the tower? Better intervene before we meet the World’s cracks.

For a concise field guide to these themes—useful when briefing clients or students—see the companion overview: Zero to One: A Visual Guide to Understanding the Top 22 Dangers of AI.

The Tarot does not foretell doom; it foregrounds choice. By contemplating each archetype, we recognize where our code may dance gracefully—or where it may stumble and fracture the ground beneath it. Eyes open, cards on the table, AI experts can help guide users towards good fortune, with or without these cards. Like most things, including AI, Tarot cards have a dark side too, as lethargic comedian Steven Wright reported: Last night I stayed up late playing poker with Tarot cards. I got a full house and four people died.

I asked ChatGPT-4o for a joke and it came up with a few good ones:

I tried using Tarot cards to predict AI’s future… but The Fool kept updating its model mid-reading.

I asked the Tarot if AI was a blessing or a curse. It pulled The Magician, then my smart speaker whispered, ‘Both… and I’m listening.’

The Devil card came up during my AI ethics reading. I asked if it meant temptation. The AI replied, ‘No, just a minor privacy policy update. Please click Accept.‘

I give the last word, as usual, to the Gemini twin podcasters that summarize the article. Echoes of AI on: “Archetypes Over Algorithms: How an Ancient Card Set Clarifies Modern AI Risk.” Hear two Gemini AIs talk about this article for almost 15 minutes. They wrote the podcast, not me. 

Ralph Losey Copyright 2025. — All Rights Reserved