01A quarter of long social posts are machine-written, and Instagram won't sort them for you
Two months ago, Pangram shipped a Chrome extension that scans social feeds as users scroll and flags the posts its algorithm reads as AI-generated. The company sells AI detection and describes itself as research-first. It bundled in an opt-in toggle, letting users anonymously share their scan statistics, then watched what the feeds looked like from inside the tool.
The first months of that data say machine text is everywhere. Across every platform Pangram's users scanned, the average AI rate landed at 13.8%. Length changed the picture fast. On four of five platforms, longer posts were more likely to be machine-written than short ones. Across all of them, 25.72% of items over 250 words were fully AI-generated. One in four long posts, by the company's count, had no human behind it.
LinkedIn carried the heaviest load, the professional network where a fluent paragraph is the whole point. Substack ran the opposite way. There, the share of fully AI content stayed roughly flat no matter the length, the rare corner of the feed where writing longer did not mean writing with a bot.
Pangram frames this as the view from a tool built to fight what it calls slop, and it wants users making informed choices about where their attention goes. The platforms hosting the content have decided that choice is not theirs to pre-empt.
Instagram head Adam Mosseri said as much this month. "I don't think we should filter out AI content," he said on Lenny Rachitsky's podcast, adding that the company should instead "let you know" when a post is synthetic. His position for anyone who dislikes the stuff was blunt: if you don't want it, "then you shouldn't have it in your feed." The work of keeping it out, in that framing, belongs to the user.
That leaves a labeling gap someone has to fill. Mosseri points to disclosure over removal. Pangram sells the removal decision back to users as a browser plug-in, one that reports a quarter of long posts already failing the test. The practical result is a professional feed where telling a person from a generator now runs through a third-party detector rather than the platform itself.
02Meta Killed Its New Instagram AI-Image Feature Three Days After Launching It
Meta announced a feature this week that let Instagram users generate AI images from any public account's content by tagging it. By Friday, the company had switched it off.
The mechanic was simple and, to critics, the problem. Tagging a public account fed its posts into Meta's image tools, and the account owner had no say. Content from any public profile could be pulled into an AI creation without permission. That default routed strangers' faces and posts into generated images by design, not by accident.
The backlash was fast enough that Meta reversed course in the same news cycle it had used to promote the tool. "Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way," the company said in a blog post, per TechCrunch. "We've heard the feedback that this feature missed the mark, so it's no longer available."
The gap sits inside that sentence. Meta framed the feature as giving people control over whether their content gets referenced. As built, the control ran the other way: public posts were opt-out at best, and the objections centered on likeness and consent, not creative utility. A tool pitched as user empowerment shipped as a way to make images of people who never agreed.
Pulling the feature does not signal a retreat from the larger push. Meta has spent the year threading generative AI through its feeds, and this was one entry point among several. What changed is a single product decision, reversed under public pressure, three days after Meta presented it as a launch worth announcing. The company has not said whether a reworked version, with explicit consent gates, will return.
The reversal is the news here. Earlier coverage documented the feature going live; this is the official walk-back, with Meta conceding the design "missed the mark."
03OpenAI's newest flagship shipped only after the government said yes
OpenAI released GPT-5.6 to the public only after the Trump administration signed off, according to The Verge. For roughly two weeks before that, the model lived in a "limited preview" open exclusively to government-approved organizations. The company's chief executive, Sam Altman, called it "the best model we have ever produced." He could not put it in front of most users until Washington cleared it.
The distance between a finished model and a shippable one used to be measured in load testing and red-teaming. Now it runs through an outside approval. GPT-5.6 was built, benchmarked, and praised by its own maker before a regulator decided when the public could touch it.
A second signal points the same direction. OpenAI stood up a Bio Bug Bounty, a program that pays outside testers to probe whether its models can help with biological threats. The company is paying people to find the reasons a model should not ship, and doing it as a named, structured process rather than an internal check.
Read together, the two moves describe a release pipeline gaining pre-conditions. One gate is regulatory: a government body decides the timing of public access. The other is a safety threshold the company built for itself, aimed squarely at biological risk. Neither existed as a public, formalized step in earlier frontier launches, when a vendor set its own ship date and announced the model the same day it turned it on.
For the model itself, the delay was temporary. GPT-5.6 is now public. What changed is the sequence: approval and biosafety evaluation moved ahead of the launch instead of trailing it, and the vendor no longer holds sole authority over when a frontier system reaches users.
That shifts what founders and developers can assume about release cadence. A capable model finishing training no longer means it ships on the builder's schedule. The bottleneck sits with whoever grants the greenlight, and with whatever a bio bounty surfaces before launch.
GPT-5.6 stays the default in Microsoft Copilot amid breakup talk OpenAI said its new GPT-5.6 family remains the "preferred model" for Microsoft's Copilot 365 suite. The statement lands as reports circulate about strain between the two companies over their commercial and compute arrangement. techcrunch.com
Anthropic will charge Claude subscribers usage fees for Fable 5 Anthropic will require paid subscribers to pay usage-based fees to reach Claude Fable 5, its top consumer model. Flat-rate access no longer covers the best tier, shifting consumers toward metered billing. wired.com
OpenAI rebrands Codex as an agent that runs for hours OpenAI relaunched its Codex coding tool as an agent that executes multi-step workflows independently, running "for hours if needed." The pitch targets tasks a developer hands off rather than supervises line by line. arstechnica.com
OpenAI's head of safety Johannes Heidecke resigns Johannes Heidecke, who led safety at OpenAI, is leaving the company. His exit coincides with OpenAI merging its research and safety teams into a tighter structure. wired.com
OpenAI publishes a GPT-5.6 proof of the Cycle Double Cover Conjecture OpenAI posted a PDF claiming its GPT-5.6 Sol Ultra model produced a proof of the Cycle Double Cover Conjecture, a long-open problem in graph theory. Mathematicians have not yet verified the result. cdn.openai.com
The AI return-on-investment debate reopens at trillion-dollar scale Analysts renewed the argument over whether AI spending pays off, now framed around a $3 trillion question. Rising capital outlays raise the stakes for buyers who cannot yet show clear returns. techcrunch.com
Microsoft will ship more security fixes per Patch Tuesday using AI Microsoft said it now uses AI to find flaws earlier, so each Windows 11 security release will bundle more fixes. The company tied the change to attackers, including amateurs, increasingly using AI to find bugs. theverge.com
1X gives its Neo home robot dexterous new hands 1X updated its Neo home-chore robot with more tactile, fast-moving fingers aimed at fine manipulation. The soft-bodied unit targets household tasks that demand precise grip. wired.com
Deutsche Telekom moves customer service and network ops onto OpenAI models Deutsche Telekom is deploying OpenAI models across customer service, employee workflows, network operations, and voice. The telecom operator describes the effort as becoming an "AI-native telco." openai.com
Estonia builds an AI tool to catch legal errors before laws pass Estonia deployed an AI system to flag mistakes in draft legislation after a single wording error cost the government $28 million. The tool is part of a wider push to automate state functions. wired.com
OpenAI hires a product manager to build ChatGPT features for families OpenAI posted a job for a product manager to design ChatGPT experiences for families, caregivers, and older adults. The role signals a push into household and multi-generational use. techcrunch.com