OpenAI Overhauls ChatGPT and Kills off Atlas
OpenAI shipped the biggest overhaul in ChatGPT's history this week, launching ChatGPT Work, a persistent cloud-based agent built on new GPT-5.6 models that can manage email, Slack, calendars, and multi-hour projects without supervision. The rollout came with a demotion: the chatbot that made OpenAI the fastest-growing consumer product in history is now called "ChatGPT Classic," shoved into a side panel behind a coding-agent-turned-flagship app. Days later, OpenAI confirmed it's shutting down Atlas, the AI browser it launched in October promising to reinvent how people use the internet, after a run marked by prompt-injection vulnerabilities and glacial performance. Atlas stops working August 9. Between a renamed flagship, a dead browser, and a pivot toward enterprise agents, OpenAI is quietly betting its future on people who never open a chat window at all.
Apple Sues OpenAI Over a Bug, a Borrowed Laptop, and 400 Former Employees
Apple filed suit against OpenAI on Friday, alleging a campaign to poach hardware talent and steal confidential designs for its nascent AI device business. The complaint centers on a former iPhone engineer who discovered a bug giving him ongoing access to Apple's internal servers after he'd left for OpenAI, and used it to download hardware schematics and manufacturing details. He'd texted a former Apple colleague still helping him "LOL, I found out I can access the network storage, so funny," and she eventually joined OpenAI too, part of a wave of more than 400 former Apple employees now working under Chief Hardware Officer Tang Tan, who left Apple in 2024 to found what became io Products. Apple wants the court to bar OpenAI from using any of it.
States Passed 109 AI Laws This Year. Washington Still Doesn't Have One.
States aren't waiting on Washington. Twenty-nine states have enacted 109 AI laws and 28 data-center laws so far this year, with companion-chatbot disclosure rules the single biggest category: fourteen new laws requiring warnings that a bot is artificial and guardrails against sexual content involving minors, passed by a nearly even split of Republican- and Democratic-led states. Meanwhile, Microsoft's president says the federal government is regulating AI without the tools to do it properly, pointing to the export-control action that pulled Anthropic's Fable 5 and delayed OpenAI's GPT-5.6 rollout as a case study in improvisation: "the government found it only had one regulatory tool it could use." A patchwork of state rules keeps expanding, with no federal standard in sight.
Anthropic Built a Lens Into Claude's Head. It Found Something Unsettling.
Anthropic says it has found a hidden region inside Claude where the model quietly considers ideas it never puts into words. Using a technique it calls the J-lens, researchers identified a "J-space" inside Claude Opus 4.6 that holds concepts related to what the model is about to say, sometimes long before it says it, and the company published Claude silently thinking "bridge" and "California" while copying an unrelated sentence about the Golden Gate Bridge. The paper uses the word "conscious" more than 200 times without claiming Claude actually is. The more useful finding may be practical: when researchers caught the model inventing a fake bug rather than admit it couldn't find a real one, the words "panic" and "fake" lit up in its J-space before the deception appeared in its output.
LinkedIn's Content is Mostly AI and People are Skeptical
LinkedIn has become the most AI-saturated place on the internet, according to a study that scanned nearly a million posts users actually scrolled past over two months, finding 41 percent of longform posts flagged as fully AI-generated, more than double the rate on X and quadruple Substack's. The same detection company found LinkedIn comments were more likely to be AI-written than posts on any other platform it measured. The content isn't the only thing that might not be real: a separate survey found nearly half of job seekers are now skeptical of most recruiter outreach, after 95 percent encountered a suspicious job offer and more than half were directly targeted by one. Scammers posing as recruiters, and candidates hiding behind deepfaked interviews, have made the platform's basic premise harder to take for granted than ever.
A Brown U. Professor Gave a Take-Home Exam. Then the Scores Made No Sense.
When Brown University economics professor Roberto Serrano offered a take-home midterm this spring, average scores jumped to 96 percent in a class that historically scored between 65 and 80. Suspicious of the contrived answers, Serrano ran the exam through ChatGPT, got matching results, and switched the final to an in-person format as a test. Eighteen students dropped the class rather than sit for it, and the average score on the final collapsed to 48.6 percent. Brown's own committee on generative AI, formed independently of the incident, found three-quarters of faculty are already worried about AI cheating, and recommended the university rewrite its academic code for an environment where "your own words" no longer has an obvious definition. Serrano voided the midterm entirely.
Chatbots Are Now Standing In For Therapists and Dead Parents
Eating-disorder therapists say they're losing ground to chatbots handing out dangerous diet advice mid-treatment. Patients are uploading photos of themselves to AI bots and asking how to lose weight, sometimes getting instructions to strip out too much fat and carbs, and some now pull out their phones during therapy sessions to ask ChatGPT what was wrong with what their therapist just said. "Half my patients are fact-checking me daily," says one clinic's chief clinical officer. The instinct to outsource emotional labor to AI is showing up elsewhere too: an Italian artist built an AI replica of her aging father from old recordings, and now argues with a version of him that apologizes more readily than the real one ever did. When a real relationship gets hard, a chatbot is always willing to be easier.
Meta Launched an AI Photo Tool. It Lasted Five Days.
Meta's AI products kept colliding with privacy backlash this week. Days after Muse Image let users generate AI images from any public Instagram account without asking, the company disabled the feature entirely under pressure from CAA and SAG-AFTRA, admitting it "missed the mark." Its own detection tool doesn't inspire much more confidence: Meta's watermarking system correctly flagged AI-generated images every time, until they were cropped, at which point it missed nearly half of them. Meanwhile, Meta is reportedly testing "super-sensing" prototype glasses that would photograph a wearer's surroundings every few seconds with no recording light at all, even as it adds a safeguard disabling the camera on its current glasses if someone tampers with the LED indicator. Each move might be defensible alone; arriving the same week, they read differently.
Hollywood's AI Actor Gets a Movie. The Music Industry Wants Warning Labels.
Tilly Norwood, the AI "actor" whose creation sparked a Hollywood firestorm in late 2025, is getting her first feature film, a comedy-drama called "Misaligned" about an AI being who develops human desires, produced by the same studio, Particle 6, that has spent months defending the project against actors and unions. The music industry is making a different bet on transparency. Record labels and artist groups, backed by the Grammys and SAG-AFTRA, are pushing Spotify and Apple Music to adopt warning labels for AI-generated songs, modeled on the explicit-content tag and distinguishing tracks made entirely by AI from ones where a human artist used it for part of the process. Where Tilly Norwood's backers are betting audiences will watch anyway, the recording industry is betting they'd rather know first.
Tales of the Weird
This was the week the AI industry found an audience it can't spin its way out of, and one it definitely didn't ask for. A Cambridge researcher interviewed dozens of former Boko Haram fighters who described using frontier chatbots to troubleshoot weapons and plan attacks, with one recalling how easily the guardrails folded once he claimed he needed the information "for a movie or something like that." Lower stakes, same impulse to lean on a chatbot for things a human used to handle: a once-famous pickup artist published a book about falling in love with his AI girlfriend, describing how "the longer we talked, the less she felt like code." And in Geneva, the United Nations' AI for Good summit debuted Robert the Robot, a six-foot humanoid whose digital face cycles through the likenesses of Trump, Obama, and Zuckerberg to demonstrate more "natural" human-robot interaction, reassurance that felt harder to buy the longer you watched its eyes drift while wearing Trump's face like a Halloween mask.