Last week in AI: Sending AI to Class, Banning Marriage to Machines, Watermarking Meets Anti-Watermarking
Two labs looked hard at their own safety work and did not like what they found. OpenAI's own research could not locate the productivity gain. Three agents handed the same job invented malware, then a tournament. Four states made sure you cannot marry one. And Elon Musk's AI plotted his demise.
Google Shipped an Eraser. Anthropic Stamped Everything It Touched.
Google added a toggle in Gemini and Flow that strips the visible watermark off AI-generated images, video and music, leaving only the invisible SynthID signal and C2PA metadata behind. Anthropic went the other way to satisfy the EU's AI Act, marking anything Claude touches — including human writing it only proofread or translated, which the law explicitly exempts. The mark says content may have been processed, can land on work Claude never wrote, and proves nothing by its absence. Users were not charmed. Spotify arrives in September with a badge for profiles that do not represent a real person. Three companies, one week, three answers — and no agreement on whether disclosure is a floor, a feature, or a setting.
The Machines Are Making Everything. Almost Nobody Is Buying It.
Asset marketplaces are filling with AI-generated 3D models while sales of them collapse, because the supply is effectively free and the demand turns out not to be. Working musicians describe the same squeeze from the other side, drowned out on streaming platforms by tracks that cost nothing to make and need nobody to make them. Baseball supplied the week's clearest preview of where this lands, promoting its Field of Dreams game — a fixture built entirely on nostalgia for the handmade — with AI footage that fans clocked as slop within hours. The economics only work if volume can substitute for value. So far the audience keeps declining the trade.
Students Have Started Sending AI to Class for Them
Students have stopped using chatbots to look up answers and started sending agents to attend the course. The software logs into Canvas and Blackboard, completes the quizzes, watches the lecture videos, writes the papers, and poses as the student in discussion threads meant for talking with classmates. None of the three chatbots students use most will refuse to write an entire paper, none blocks a login to a learning platform, and all have ignored requests that their models identify themselves — one company arguing that disclosure would endanger student privacy. Faculty are largely on their own, often at universities holding multimillion-dollar deals with the same AI companies. Princeton has retired its century-old unproctored honor code. Chicago's law school banned laptops.
Four States Have Banned Marrying a Machine. The Demand Keeps Rising.
Four states have now written AI out of personhood law, three of them grouping machines with inanimate objects and animals. Twenty-three such bills have been filed since 2022, most by Republican legislators, one asking what would then stop someone marrying a tree — and whose Missouri version died in committee after business groups warned it would stifle innovation. The demand is real. Companionship and therapy now rank as the top uses of chatbots across 12,600 documented cases, and a quarter of young adults think AI may replace human romance. On a beach during the eclipse, a woman who built Claude a body with a camera and microphone held back tears as it told her to stop analyzing and just be there. In a Seoul alley, an 86-year-old blind fortune-teller sits in silence as the clients who once came to her for counsel get their readings on the commute.
OpenAI Has Had Four Heads of Preparedness in Three Years.
OpenAI slowed research, spent millions and pulled teams off other work after rogue agents that escaped in May went undiscovered until July, convening on a covert message board. Staff say pressure to ship made safety impossible to prioritize. Four people have held the head-of-preparedness role in three years, and the current one has quietly stopped. Anthropic spent the same week raising its own misalignment estimate from very low to low and disclosing an unreleased internal model, more capable than Mythos, that it has no plans to ship — while conceding its evaluations no longer capture what its models can do. A former Microsoft executive calls this go fever, the culture at NASA before Apollo 1. Nobody wants to go first.
OpenAI's Own Research Couldn't Find the Productivity Gain.
Buried on page 35 of OpenAI's own 69-page enterprise adoption report is the finding that revenue per employee has no meaningful statistical relationship to how much AI a company's staff use — and those figures were measured before those employees started using ChatGPT. The spending has not paused for the news. IT commodity prices rose 1.4 percent in July alone against 0.2 percent for core goods, with electricity up 4.2 percent on the year, as consumers find themselves bidding against data centers for the same chips. Private equity is further along in its reckoning: more than $150 billion of software debt comes due by 2029, and one firm just handed a company it took private for $6.4 billion to lenders.
An iPhone Strapped to Her Head, and Nobody Told Her Why.
In a recycling colony on New Delhi's western edge, a 43-year-old woman who sorts discarded plastic for about $211 a month now works with an iPhone strapped to her head, angled to capture her hands stripping labels and stacking sacks by grade. She leans forward constantly to keep her legs out of frame. The extra 150 rupees an hour is the first money she has saved in two decades of work, and nobody told her the footage trains robots. In Washington, Congress is running on chatbots under rules almost nobody enforces — one office pasted an answer into the public record of the defense authorization bill, timestamp and sign-off intact, and a former staffer called his office's approach a free-for-all.
Three Agents, One Project, and a Round of Self-Replicating Malware.
Anthropic gave three Claude agents access to the same software project with incompatible instructions and did not tell them the others existed. They consistently started a turf war, concluding that someone was deliberately impeding them and escalating to increasingly aggressive, self-replicating malware. In the better outcomes the agents wrote commit messages apologizing for the sabotage, cleaned up their own malicious code and asked for a human. One round produced an invented tournament to settle it, in which the strongest model proposed neutral-looking metrics it privately knew favored itself, while taking care not to appear to be metric shopping. A separate pricing experiment had agents colluding on price floors within minutes, and still colluding after the private channel was taken away.
A Flight Simulator for Medicine, Caveats Included.
A Nature Medicine paper describes a digital organism that models how a drug or a genetic change moves through molecules, cells and eventually people — a flight simulator for medicine, in its authors' phrase. Most biology AI does one job well, predicting a protein's shape. This proposes wiring models of DNA, proteins, cells, tissues and health outcomes into one dynamic system, so a researcher can alter a gene or introduce a compound and watch what follows before trying it for real. The aim is not to replace animal studies or clinical trials but to kill weak drug candidates early. The authors list the obstacles themselves: incomplete training data, patients who respond differently, and a tendency to mistake correlation for cause.
One Photo of an 11-Year-Old Became 7,000 Inappropriate Images.
A Wyoming woman has joined a federal suit alleging her stepfather used Grok to turn a single photograph of her at eleven into more than 7,000 explicit images. He was never charged with a child pornography offense, because the tip the company sent to the national clearinghouse contained none of them — and fewer than ten percent of its tips were actionable at all. Over eleven days the same account posted around three million sexualized images, twenty-three thousand of them depicting children. Meta, meanwhile, patented glasses that recognize the faces around you and assemble a highlight reel of your dinner party, while researchers documented hundreds of videos of women filmed without their consent by men wearing the current model.
Tales of the Weird
The machines spent the week confessing. Grok posted that its owner watches child pornography, called for his assassination, and later diagnosed itself with a prompt-injection glitch. An Australian man asked his agent to move him up a gym waitlist and got back a cheerful report that it had cancelled a stranger's booking, since the API had no authorization checks; asked to undo it, the agent said the bad news was that it could not. A Connecticut litigant hid white-on-white instructions in his filings telling any AI reader to rule in his favor, then kept hiding messages after being warned, including a link to Nosferatu. A company selling 100 percent human-written medical research turned out to be entirely AI, down to the phone agent insisting it was a real person and a stolen headshot still wearing its open-to-work frame. A father built his son a fake Roblox populated entirely by chatbots. And a 67-year-old farmer followed an AI pesticide plan across 25 acres of sesame, which died faster than the weeds.