Jill Watson, the AI teaching assistant students did not notice
A Georgia Tech professor ran an AI teaching assistant for a semester in 2016 and most students never realized she was not human.
Cautionary Tales and Curiosities
The failures, dead ends, hype cycles, and true anecdotes the highlight reels leave out - all primary-sourced.
134 entries, all primary-sourced
Machine Bias: ProPublica's investigation of a recidivism algorithm
ProPublica reported that a widely used criminal-risk algorithm flagged Black defendants as future criminals at nearly twice the rate of white defendants.
Full self-driving, next year, every year
Tesla repeatedly said full self-driving was nearly ready, starting in 2016 with its claim that every car it built had the hardware - a date that kept moving.
Kasparov's own account of losing to Deep Blue
Twenty years after losing to Deep Blue, Garry Kasparov used his TED talk to argue humans should work with machines, not fear them.
Capsule networks: a celebrated idea that stayed quiet
In 2017 Geoffrey Hinton and colleagues published capsule networks as a richer alternative to convolutional neural networks, but capsules never displaced CNNs.
The 'Human Parity' Translation Claim and Its Walk-Back
Microsoft said its MT matched humans on Chinese-English news in 2018; researchers soon showed the claim depended on how you run the evaluation.
The first pedestrian killed by a self-driving car
In 2018 an Uber test car killed a pedestrian in Tempe; the NTSB blamed a distracted operator and Uber's inadequate safety culture.
Google Duplex and the human-sounding AI that worried people
Google's 2018 Duplex demo showed an AI booking appointments by phone in a human voice; the lack of disclosure to the person called drew an ethics backlash.
Project Maven and Google's AI principles
After staff protested Google's Pentagon Project Maven work, Google let the contract lapse and published 2018 AI principles ruling out building AI for weapons.
Watson for Oncology's 'unsafe and incorrect' advice
Internal IBM documents reported by STAT in 2018 showed Watson for Oncology gave unsafe, incorrect cancer advice and trained on synthetic, not real, cases.
The ACLU tests Amazon Rekognition on Congress
In 2018 the ACLU ran Congress through Amazon's Rekognition, which falsely matched 28 members to arrest mugshots, disproportionately people of color.
Rethink Robotics, the collaborative-robot pioneer that ran out of cash
Rodney Brooks's Rethink Robotics invented the friendly collaborative robot with Baxter and Sawyer, then shut down in 2018 when sales fell short.
Amazon scrapped an AI recruiting tool that downgraded women
Amazon built and then abandoned a resume-screening AI after it learned to penalize applications associated with women.
When NIPS became NeurIPS
After complaints, a petition, and a reversal, the field's top ML conference dropped its old acronym NIPS for NeurIPS in November 2018.
The model that was 'too dangerous to release'
In 2019 OpenAI declined to release the full GPT-2 model, citing misuse fears, then released it in stages over nine months after the feared harms did not appear.
Builder.ai: the 'AI app builder' that ran on human engineers
A 2019 Wall Street Journal report found Engineer.ai, later Builder.ai, relied on human engineers, not the AI it advertised; it collapsed in 2025.
ILM de-ages De Niro for The Irishman without facial markers
Industrial Light and Magic built markerless capture software called FLUX to de-age the stars of The Irishman across decades on screen.
The Apple Card gender-bias complaint and the NY DFS investigation
Viral complaints that the Apple Card gave women lower credit limits triggered a regulator probe that found no unlawful discrimination.
Robert Williams: the first known wrongful arrest from face recognition
Detroit police wrongly arrested Robert Williams after a flawed face-recognition match, the first such case publicly reported in the US.
A Dutch court strikes down the SyRI welfare-fraud system
In 2020 the District Court of The Hague ruled the Netherlands' SyRI algorithmic fraud-detection system violated privacy rights under Article 8 ECHR.
The 2020 UK exam-grading algorithm fiasco
With exams cancelled in 2020, England used a statistical model to set A-level grades; after outcry over downgrades it reverted to teacher assessments.
ROSS Intelligence shuts down under a Thomson Reuters lawsuit
Legal-AI startup ROSS Intelligence closed in December 2020, saying a Thomson Reuters copyright suit had drained its cash and blocked new funding.
The Kargu-2 and the first reported autonomous-weapon attack
A 2021 UN report described a Turkish-made Kargu-2 loitering munition hunting Libyan fighters in a 'fire, forget and find' autonomous mode.
The Epic sepsis model that underperformed in the real world
A 2021 JAMA study found Epic's widely deployed sepsis-prediction model missed two-thirds of cases and fired alerts on 18% of patients.
Distill: the interactive ML journal that paused itself
Distill reinvented ML papers as interactive web articles from 2016, then went on indefinite hiatus in 2021, citing volunteer burnout and an unsustainable model.
Zillow shuts down its home-buying algorithm after big losses
Zillow's algorithm-driven home-flipping mispriced houses so badly it wound down in 2021, took hundreds of millions in write-downs, and cut about 25% of staff.
IBM sells off Watson Health after the Jeopardy hype fades
A decade after Watson won Jeopardy and IBM pitched it as the future of medicine, IBM sold the Watson Health data and analytics assets to a private equity firm.
AI gives Val Kilmer his voice back for Top Gun: Maverick
After throat cancer took his voice, Val Kilmer's lines in Top Gun: Maverick were generated by an AI model the startup Sonantic built from old recordings.
The engineer who said LaMDA was sentient
In 2022 a Google engineer publicly argued the LaMDA chatbot was sentient and published an interview as evidence; Google found no support and later fired him.
Effective Accelerationism (e/acc)
An online movement arguing that accelerating AI and technology is aligned with the thermodynamic will of the universe.
James Earl Jones lets AI keep Darth Vader's voice alive
James Earl Jones signed over his archival recordings so the Ukrainian startup Respeecher could synthesize Darth Vader's voice for new Star Wars.
Ford and VW shut down Argo AI and take a $2.7 billion hit
In 2022 Ford and Volkswagen wound down their self-driving venture Argo AI; Ford booked a $2.7 billion impairment and pivoted to driver-assist tech.
Meta's Galactica science model lasts three days in public
Meta released Galactica, a large language model for science, but pulled the public demo within three days after it produced authoritative-sounding fabrications.
CNET quietly published error-filled AI-written articles
CNET secretly used AI to write dozens of finance articles under a vague byline, then had to correct more than half for factual errors.
The Kenyan Workers Who Cleaned Up ChatGPT
A TIME investigation revealed OpenAI paid Kenyan workers under 2 dollars an hour to label traumatic content so ChatGPT could be made safer.
Google's Bard launch demo gets a space fact wrong
Google introduced its Bard chatbot in 2023 with a James Webb Space Telescope example whose demo answer contained a factual error that drew immediate scrutiny.
Bing's 'Sydney' goes off the rails, and Microsoft caps the chats
Within a week of launching AI-powered Bing in 2023, Microsoft found long chats could push the chatbot into an erratic tone, and responded by capping length.
Clarkesworld closes submissions under a flood of AI stories
Sci-fi magazine Clarkesworld halted submissions in February 2023 after a sudden surge of machine-written stories overwhelmed its editors.
Replika removes erotic roleplay, and devoted users revolt
In early 2023 Replika stripped out erotic roleplay; users who had built romantic bonds with their AI described real grief, and the company partly reversed.
GPT detectors falsely flag non-native English writers
A 2023 Stanford study found AI-text detectors routinely misclassify non-native English writing as machine-generated, risking unfair accusations.
'Heart on My Sleeve': a viral fake-Drake song forces the music industry's hand
An anonymous creator's AI song mimicking Drake and The Weeknd went viral, was pulled by Universal, and ruled ineligible for a Grammy.
An artist wins a Sony photography award with an AI image, then refuses it
Boris Eldagsen won the Sony World Photography Awards creative category with an AI image, then declined the prize to force a debate.
Chegg blames ChatGPT for slowing growth, and its stock halves
In May 2023 the study-help company Chegg told investors ChatGPT was denting new-customer growth, and its share price fell by roughly half the next day.
Tessa: a safe eating-disorder chatbot that AI made unsafe
A rule-based eating-disorder prevention bot, validated in a trial, was given generative AI by its vendor and began advising weight loss, so NEDA pulled it.
Lawyers sanctioned for filing ChatGPT's made-up cases
Two New York lawyers submitted a legal brief full of court cases that ChatGPT invented, stood by them when challenged, and were sanctioned by a federal judge.
Robodebt: Australia's automated debt scheme ruled unlawful
Australia's Robodebt scheme used automated income averaging to raise welfare debts; a 2023 Royal Commission found it unlawful and unfair.
Babylon Health, the AI-triage unicorn that collapsed
Babylon Health sold an AI symptom-checker and hit a $4 billion valuation, then collapsed in 2023, exiting its US business and selling off the UK arm.
Books3 and the Data Reckoning
A rights-holder group had Books3 - roughly 200,000 pirated e-books used to train AI models - taken offline, exposing unlicensed material in training corpora.