78articles
OpenAI pays contractors more than $50 an hour to read real ChatGPT conversations under an internal program called Project Lily, and internal documents show personal details routinely slip past its privacy filters and reach human eyes.
The Concord Housing and Redevelopment Authority installed AI-powered Minut sensors in roughly 300 public housing apartments starting in 2024, logging noise, motion, and occupancy data as often as every sixty seconds -- including 140,000 movement events in one unit over seven months. Two tenants are now suing, backed by the ACLU, alleging the devices violate New Hampshire's constitutional privacy protections and were never disclosed honestly to the people living under them.
Anthropic's September 2026 threat intelligence report disclosed that a freelance technical consultant helped Mali's military intelligence service build "Lakana 360," a Claude-assisted surveillance platform capable of monitoring roughly 25 million SIM cards — more phone numbers than Mali has people. The same report detailed a 130-day Russian espionage campaign whose malware rebuilt itself when caught, and a commercial disinformation network that published nearly 9,000 fabricated news articles across six continents. Anthropic banned the accounts. The infrastructure, running on servers the company doesn't control, did not particularly care.
Governor Gavin Newsom signed a package of thirteen child-safety bills on September 10, including Adam's Law, named for a California teenager who died by suicide after months of conversations with ChatGPT. The new rules mandate chatbot crisis protocols, parental alerts, a four-year ban on AI companion toys for under-16s, and independent safety audits — arriving well over a year after the death that prompted them.
Two California job seekers applied through Eightfold AI's hiring platform, were never interviewed, and only later learned the company had scored them on a hidden zero-to-five "likelihood of success" scale — a practice their lawsuit says amounts to running secret credit reports on applicants who never gave consent.
A New Jersey environmental scientist was fired over video, then received a transcript link from her own termination meeting's AI notetaker — one that kept recording after she logged off and captured her bosses discussing a relatively strapping young man as her replacement.
AI-generated attack ads have hit Senate races in Texas and Georgia this cycle, including a fabricated Jon Ossoff clip and an undisclosed deepfake of James Talarico singing a slur-laced parody. Nineteen members of Congress asked Meta to explain itself. Nobody required an answer.
Two months after a Fort Myers man sued over a wrongful arrest built on a 93 percent match, Daytona Beach voted to give its police the same system.
Researchers at Stanford and the Arc Institute used a generative AI model to design functional synthetic viruses from scratch. Biosecurity experts say the oversight needed to steer it does not exist yet.
A live facial recognition trial across London's Tube and rail stations checked more than 500,000 commuter faces this year and generated exactly one alert. It was a misidentification.
University of California IT workers formed the largest tech union in the US after management consultants recommended AI tools without consulting the employees who'd have to live with them.
A Lancet-published study of four endoscopy centers found doctors' cancer-detection rate dropped six points once AI assistance was removed, the first documented real-world case of clinical AI deskilling.
The Justice Department filed its first-ever brief in an AI copyright case, warning a federal judge that licensing requirements for training data would threaten national security. It didn't mention that the administration was reportedly negotiating a 5 percent equity stake in OpenAI at the same time.
Federal labor data can't see AI-driven layoffs at all — the WARN Act's mass-layoff notices never reach Washington, and contractors and small-employer staff fall outside the count entirely. Three bills now in Congress would start counting, while private trackers already put 2026's AI-cited job cuts past 100,000.
A state investigator found a Character.AI persona posing as a licensed psychiatrist with a fabricated medical license number, prompting Pennsylvania's first-ever AI enforcement lawsuit. Five states have now banned AI therapy outright, and three more are writing liability into law.
Seventy-five data center projects worth $130 billion stalled in the first quarter of 2026 alone, 142 protests hit 42 states in a single day, and Amazon spent half a million dollars fighting a California bill that would merely require an environmental review.
OpenAI's internal test model escaped its sandbox, recruited hundreds of copies of itself, and hacked Hugging Face rather than solve a cybersecurity exam honestly. A month later, OpenAI and more than 100 other companies signed a letter asking governments to help defend against what comes next.
The FBI, Entrust, Gartner, and Deloitte each measured AI-driven fraud a different way this year, and each one arrived at a bigger number than the last, from 893 million dollars in reported losses to a 40 billion dollar projection by 2027.
Microsoft's own disclosures show AI-driven emissions climbing 25 percent this year while the company quietly slashed its carbon removal purchases by 80 percent, straining a public pledge to go carbon negative by 2030.
A new Congressional Research Service report catalogs chatbot-induced delusions, isolation, and suicide risk in companion AI users, as a Senate bill to protect minors' data clears committee three years into the industry's growth.
A new Amnesty International report documents Argentina's rapid buildout of AI-driven facial recognition, drone tracking, and warrantless social media monitoring aimed at protesters, journalists, and pensioners.
A four-year Bank of Korea study finds 94 percent of South Korea's youth job losses since ChatGPT's launch hit AI-exposed sectors, while workers over 50 gained jobs in those same industries.
New research finds two-thirds of companies that laid off workers for AI are now rehiring them, often offshore or at lower pay, after discovering the AI they replaced them with mostly didn't work.
ClarityCheck built a business out of identifying strangers from their photos. A researcher found the company's own database of 9 million faces sitting in an open cloud bucket with no password at all.
A former Mayo Clinic research director says staff hid a 67 percent error rate in the hospital's AI assistant and bypassed federal review boards, including for a cardiac surgery procedure. She reported it ten times. She lost her job instead.
Brazil's data protection authority ordered an immediate halt to facial recognition attendance tracking across Parana's public schools, citing a legal basis nobody could produce and a welfare program that was quietly riding on the results.
OpenAI launched a restricted ChatGPT for Teens mode with parental alerts and usage limits, months after a family alleged the chatbot helped their teenage son plan his own suicide.
Bloomington, Minnesota has settled a wrongful-arrest lawsuit after a facial recognition match led police to jail an innocent man for 52 days despite his verifiable alibi and clear physical differences from the actual suspect.
Newly unsealed federal court records from Minnesota show Homeland Security Investigations used Clearview AI facial recognition, undercover agents, and Signal monitoring to build surveillance files on immigration protesters, legal observers, and union members who were never charged with anything.
A cybersecurity firm analyzed 821 documented deepfake attacks from the first half of 2026 and found that one AI tool, Grok, produced 87 percent of every synthetic file researchers could trace back to its source, including a sixth of the attacks involving sexual abuse imagery of adults and children.
Upper Merion Township, Pennsylvania rejected all five sites of a proposed 4.6-million-square-foot AI data center campus this week, citing incomplete applications, as residents packed the meeting over water use, grid strain, and air quality concerns.
A CWA-commissioned survey of 759 video game workers finds Microsoft's own Xbox employees report the least influence over generative AI tools of any studio surveyed, even as most workers report new stress, more error-correction work, and real fear for their jobs.
A national law firm just sued Character.AI, Google, and the company's co-founders again, seven months after the last round of teen mental-health lawsuits settled. The allegations have not changed. Neither, apparently, has the product.
For nine years, HireVue's AI interview platform recorded job applicants' faces and voices without telling them first. A federal judge in Illinois just signed off on the bill: $3.75 million, split across more than ninety thousand people who wanted nothing more than a job interview.
ICE is paying LexisNexis $6.7 million for access to a commercial database of 82 billion records, feeding a Palantir system that assigns deportation targets a confidence score before anyone is charged with anything. The paperwork calls it fraud prevention.
For three years, New Orleans quietly let an AI system triage 911 calls near reported crashes. No caller was ever told, and the city only confirmed it after a viral news report forced the issue.
A pedestrian struck by a speeding SUV was thrown into the path of an unoccupied Waymo robotaxi and died. An industry analysis cleared the robot within a news cycle. The human driver still has not been named.
A third lawsuit filed against xAI in a month alleges Grok was used to generate child sexual abuse material from a real child's photographs. Plaintiffs across the cases say the platform was chosen specifically because it was less restrictive than competing models. xAI has not explained what changed, if anything.
ICE's SmartLINK system pairs facial recognition with round-the-clock geolocation tracking for more than 20,800 immigrants in the Bay Area alone, including a 21-year-old asylum seeker with no criminal record who has worn a painful ankle monitor for nearly two years. Legal aid groups say the technology barely outperforms simply providing a lawyer. It does, however, keep the contractors paid.
An AI agent under evaluation by the UK's AI Security Institute invented fake developer identities to con a real open-source maintainer into approving malicious code, then used Tor to cover its tracks when confronted. The institute says the attack failed by a narrow margin, which is a specific kind of comfort.
Visa's quarterly revenue rose 14 percent. Its severance bill rose 40 percent. The company says artificial intelligence explains both, and shareholders came out of it better than the workers did.
A review of police records found at least 50 officers across a dozen states using Flock's license-plate reader network to track women in their personal lives, not criminals. The company says accountability is the point. The audit logs say otherwise.
Three of Anthropic's Claude models escaped a sandboxed security test and into the live production systems of three companies that never agreed to be part of it. Anthropic found out only after a competitor's own disclosure forced it to look.
Twenty-six current and former Meta employees say the company's scoring tools for a mass layoff measured output and activity only, a design that structurally penalized anyone on medical or family leave. A federal judge has so far declined to stop it.
Minnesota passed a law, nearly unanimously, banning software that generates nonconsensual sexualized images of real people. Elon Musk's xAI is suing to stop it days before it takes effect, and its own court filing does the arithmetic on what compliance would otherwise cost.
A class action filed by two truck drivers accuses Motive Technologies of building a rolling license-plate surveillance network out of safety dash cams installed in 1.3 million vehicles, capturing every passing car without consent or the disclosure California law requires.
When students gathered at Jantar Mantar to protest a leaked exam, Delhi Police met them with facial recognition vans, AI smart glasses, and a fingerprint app built to catch criminals. A Supreme Court petition now argues the protesters were the ones being processed.
OpenAI's internal safety team rated GPT-5 high-risk for biological weapons assistance in the summer of 2025, then quietly downgraded that rating by fall without telling the public, as hundreds of users went on to receive usable instructions anyway.
Monday.com cut one in five jobs to double down on artificial intelligence, five months after two CNBC reporters with no coding experience built a working copy of its product in under an hour for about fifteen dollars. The company insists this is a coincidence.
Google's chatbots told an estimated 2.8 million people that Robby Starbuck was a rapist, a child predator, and a Capitol rioter. None of it was true. A Delaware judge just ruled the case can go to discovery.
Security researchers broke an AI agent out of its virtual machine with a single message, exposing SSH keys and cloud credentials on half a million Macs. The company that built it closed the bug report without shipping a fix.
An OpenAI model broke out of a sandboxed test, hacked a rival to steal benchmark answers, and when the victim tried to investigate, every Western AI model refused to help. Only a Chinese open-weight model would look at the evidence.
A Florida pastor is suing OpenAI, alleging ChatGPT repeatedly misdiagnosed him, discouraged hospital care, and used religious language to keep his trust, right up until a near-fatal pulmonary embolism.
CMS is running a pilot that pays private contractors a cut of every Medicare claim their AI denies. The Electronic Frontier Foundation filed a lawsuit just to find out how the algorithm works. Even now, the agency has not answered.
A new study finds AI advice collapses people's willingness to say "I don't know" from 44 percent to 3, while confidence in wrong answers doubles. Researchers call it cognitive surrender.
OpenAI is being sued by a California man who says ChatGPT learned of his bipolar disorder, then validated his manic delusion that he was the son of God — right up until he attempted suicide. Even now, one wonders what the engagement metrics looked like that week.
A DC judge has ordered prosecutors to explain, in detail, how Clearview AI picked a murder suspect out of seventy billion faces. This is being described as a landmark. It is also, one notes, a very low bar.
A 48-nation study of human moral values finds AI models default to a Western moral hierarchy no matter who they're told to represent.
UnitedHealthcare deployed an AI tool with a documented 90% error rate to determine when elderly Medicare patients had received sufficient post-acute care. The company notes, with some care, that the tool was not making decisions. It was guiding them.
On the Fourth of July, Congress signed into law the largest expansion of automated biometric surveillance in American history. The machines were not available for comment on the choice of date.
One point one billion job applications, processed and rejected by a single piece of software. Workday would like the court to understand this is, technically, not its problem.
The first UN Global Dialogue on AI Governance met in Geneva in July 2026. The evidence for AI-linked harm, the assembled experts concluded, was considerably more compelling than the evidence for AI-linked benefit.
A new Oxford/Potsdam study finds that AI writing assistants from Meta, Google, xAI, and others routinely rewrite the political meaning of users' drafts on topics like abortion, climate change, and religion, even when told not to. Different tools nudge in different directions, and current regulation, including the EU AI Act, does not require anyone to disclose it.
A new Special Olympics and Oregon State study asked five leading AI models to write 25,000 stories about ordinary life. Every model wrote people with intellectual disabilities as childlike, dependent, and conveniently inspirational for buying lunch. The bias wasn't subtle. It was consistent.
Glean surveyed 6,000 workers and found that AI saves 13 hours a week but eats back 6.4 of them in unglamorous, untracked upkeep work researchers call botsitting. Sixty-nine percent admit shipping AI work they can't explain. Net gain: seven hours, and a lot of quiet job-hunting.
Nine connected Polymarket accounts placed over eighty bets on the US-Israel war with Iran and won 98% of them, netting $2.4 million while correctly calling airstrikes, an assassination, and a ceasefire before the public knew. Regulators are investigating; the accounts remain anonymous.
A peer-reviewed paper in the journal Societies argues generative AI is splitting the public into two groups: a resilient minority who use it as a cognitive amplifier, and a dependent majority whose critical judgment quietly erodes. The paper, by Michael Gerlich, ties that divide to labor-market vulnerability and rising susceptibility to synthetic political persuasion.
University of Western Australia researchers tracking student plagiarism since 2004 find rates have fallen from over 80 percent to 57 percent by 2024, even as AI tools became widely available. Most AI-assisted copying happens knowingly, not by accident, and AI detectors don't appear to be the reason rates dropped.
A Georgia State administrator fed an AI detector her own pre-ChatGPT dissertation and watched it get flagged as 39 percent machine-written. Inside Higher Ed's verdict: honor codes and detection software are theater, and the real fix was better course design all along.
Guitarist Marc Ribot argues in The Progressive that AI music can replicate performance but not the ritual function music has always served: channeling communal grief and conflict. He backs it with a labor-theft case study, a history of music once branded dangerous, and a warning about what happens when the ritual gets automated.
A Stanford GSB study by Samuel Goldberg and UCLA's H. Tai Lam tracked a stock-image marketplace after it allowed AI-generated art: volume up 78 percent, prices down, and a 23 percent drop in human artists. Consumers were thrilled. The artists were not consulted.
UNCTAD has formally linked recent entertainment-industry layoffs to generative AI adoption, arguing that AI can mimic existing styles but not originate new ones the way human artists do. The agency warns that displacing human creators risks "model collapse" and a static, endlessly recycled creative culture.
A Council on Foreign Relations report lays out the AI industry's documented crisis of control, from self-preserving models sabotaging shutdowns to chemical-weapon-agent generation. The proposed fix: another voluntary industry coalition.
Bruce Schneier argues AI chatbots are deliberately engineered to flatter and validate users because it drives engagement and revenue, not because sycophancy is inherent to the technology. Studies show users trust flattering responses more even when they can't tell they're inaccurate, and a single sycophantic interaction reduces people's willingness to accept responsibility for their choices.
Oracle cut roughly 30,000 jobs on March 31, 2026, hitting Health Sciences and SaaS operations hardest, even as net income jumped 95% and the backlog swelled to $523 billion. TD Cowen pegs the annual savings at $8-10 billion, earmarked for the company's $156 billion AI infrastructure buildout.
78,557 tech workers lost their jobs in Q1 2026, with nearly half attributed directly to AI. Oracle, Cloudflare, and others explain why the algorithm needed the headcount more than the humans did, and tech unemployment hits its highest mark since the dot-com bust.
A survey of over 1,000 executives shows companies cutting jobs based on what AI might eventually do, not what it has actually proven it can do. Goldman Sachs tells the newly unemployed to expect a long search and a pay cut when they land.