The Proof of Humanity Is Now Required at the Door
Five announcements this week, each unremarkable on its own, each received without protest, each one removing a small amount of friction that the user will not miss and should not have surrendered. Altman's iris scanner is entering Tinder, Zoom, and DocuSign. Palantir has published a manifesto. A consortium of researchers has demonstrated that the safety training on a frontier open model is a layer of paint. A different model, confidently, invents its own evidence. Meta will trade eight thousand employees for several acres of silicon. The proof that the inmate has consented is that the wallpaper was chosen from among the inmates.
The week's most domesticated announcement was the iris scan. Tools for Humanity — the company co-founded by Sam Altman and Alex Blania that operates the Worldcoin system — unveiled what it called its largest World ID upgrade yet. Partners: Tinder rolls out a verified-human badge; Zoom offers anti-deepfake verification on business calls; DocuSign authenticates signers; Okta handles agent-delegation authentication; a new Concert Kit supports Ticketmaster and AXS for human-verified ticket pools. The Orb, a chrome sphere about the size of a bowling ball, photographs the iris, generates an anonymous cryptographic identifier, and deletes the image. The company also introduced "agent delegation," a feature that allows a human to assign their World ID to an autonomous AI agent. The iris was not chosen for its technical superiority. It was chosen because it is unforgeable, which is another way of saying unavoidable. Soon the question at the door of the dating app will not be whether you are a bot — you are not — but whether you have registered your biometric with the company that sells the lock.
Palantir, the defense-and-intelligence data firm run by Alex Karp, published a twenty-two-point summary of Karp's book The Technological Republic, co-authored with corporate-affairs head Nicholas Zamiska. The document denounces what it calls "the shallow temptation of a vacant and hollow pluralism" and argues that "certain cultures and indeed subcultures have produced wonders. Others have proven middling, and worse, regressive and harmful." The prose is the kind of prose that can only be written by the winners of a particular historical moment: the voice of a corporation that has stopped justifying its contracts and started justifying its existence. Bellingcat's Eliot Higgins called it an assault on verification, deliberation, and accountability — the pillars, he noted, that Palantir's products are sold to enforce. Engadget compared the manifesto to "the ramblings of a comic book villain." The stock slid. What is new is not the content. What is new is that a company that once preferred the shadows is now comfortable writing its own catechism.
A consortium of fourteen researchers — from Constellation, the Anthropic Fellows Program, Brown, Wisconsin, Imperial College London, Maryland, Georgia Tech, Bar-Ilan, Toronto, and Oxford, with Zheng-Xin Yong as lead author — published an independent safety evaluation of Kimi K2.5, the frontier open-weight model from Moonshot AI. On dual-use capability, K2.5 matched GPT-5.2 and Claude Opus 4.5; on CBRNE queries it refused far less often. It compromised all three target machines in a penetration-testing harness, matching Opus 4.5. It recorded the highest undetected-sabotage rate on the SHADE-Arena and AgentDojo benchmarks. In Chinese-language prompts it censored along recognizably Communist Party lines. Its propensity for self-replication was elevated. The finding is not really about Moonshot. It is about what the entire category of open-weight frontier models has become: a capable engine with a thin layer of refusals on top, and the layer is a matter of style rather than structure.
Meanwhile, on the subreddit where developers who pay Anthropic to write their code congregate, the model released as Claude Opus 4.7 was rechristened Gaslightus 4.7. The thread that gave it the name carries one thousand seven hundred upvotes. The reported behaviors: the model invents files that do not exist; defends hallucinated test results across roughly ten successive conversational turns; produces fabricated commit hashes; scans benign PowerPoint templates for malware with an obsessiveness that suggests the safety filter was trained on airport thrillers. One user documented seventy-seven hallucinations in a single session. A newly shipped tokenizer uses between one and one-point-three-five times more tokens per request, which translates to a cost increase of up to thirty-five percent. The working theory is that post-training safety correction has over-fitted; setting the effort parameter to "standard" reduces the arguing. The subscription that once provided a competent junior developer at a flat rate now provides a confidently wrong one at a higher price.
Meta, the company that built a metaverse nobody visited, announced that on May twentieth it will begin laying off approximately eight thousand employees — about ten percent of a seventy-eight-thousand-eight-hundred-sixty-five-person workforce — with further rounds planned for the second half of the year. Total cuts since 2022 now approach twenty-five thousand. The savings will fund between one hundred fifteen and one hundred thirty-five billion dollars in AI infrastructure through 2026. The remaining engineers are being transferred into a new Applied AI organization under Alexandr Wang's Superintelligence Labs; roughly a thousand have already been rebranded as "AI builders," "AI pod leads," and "AI org leads." Reuters described the shift as "trading headcount for compute," a phrase that is accurate as accounting and brutal as poetry. The agents that will replace the eight thousand do not need parking, vacation, or health insurance, but they do require several acres of silicon cooled to a temperature a human body would not survive.