Anthropic Preaches Anti-Surveillance, Now Accused of Secretly Tracking Chinese Users

01He had a few days left to use Claude Fable. He spent $149.25 shipping sqlite-utils 4.0

Simon Willison started the review from his iPhone. The prompt went into Claude Code for web: a final pass before shipping a stable 4.0 release of sqlite-utils, his Python library for wrangling SQLite databases, with instructions to catch any last-minute breakage. He was working against a deadline that had nothing to do with the code. Claude Fable, the model doing most of the writing, was leaving his team's Max subscription in a few days.

The number he attached to the work was exact: about $149.25. That covers the bulk of the effort to move from the 4.0rc1 release candidate he wrote about two weeks earlier to a 4.0 he says he felt genuinely comfortable stamping. Willison keeps sqlite-utils on SemVer, and he tries to make incompatible major versions rare. A 4.0 is a promise about breaking changes, which makes the final review the part where a mistake costs the most.

This was not a demo. sqlite-utils has real downstream users who pin to major versions, and the 4.0rc2 tag is a shipped artifact, not a screenshot. The framing Willison put on it was blunter than any benchmark: the model was a tool with an expiration date, useful right up until the subscription window closed and it disappeared.

That disposability is becoming the point. Vercel CEO Guillermo Rauch, describing how teams pick models in production, told TechCrunch that the calculation collapses to one axis. "The reality is, when you're optimizing for production, you start looking at a price/performance," he said. Rauch has argued for splitting models apart from the agents that wrap them, so the model underneath can be swapped when the price or the performance moves.

Willison's receipt is that argument in miniature. He did not benchmark Fable against rivals or wait for a leaderboard. He had a defined task, a fixed spend, and a countdown, and he ran the job before the tool got taken away. For a solo maintainer, the question stops being which model tops the charts and becomes how much finished work a set dollar figure buys before access changes.

The 4.0rc2 candidate is out. Whether Fable stays available to buy that work is a decision on Anthropic's subscription roadmap, not Willison's.

02AI Is Cutting Jobs at Both Ends of the Ladder It Built On

Microsoft cut roughly 4,800 roles on Monday, about 2.1% of its global workforce, with Xbox and commercial sales absorbing the heaviest losses, according to TechCrunch. It was not an isolated announcement. The company's move joins a running TechCrunch tally of major 2026 tech layoffs in which employers named AI as a stated factor, a list long enough that TechCrunch now maintains it in reverse chronological order.

That accounting captures one end of the labor market: salaried white-collar work, the kind employers now say software can absorb. The other end lost its most recognizable piece of infrastructure the same week.

Amazon told customers it will stop accepting new sign-ups for Mechanical Turk, the crowdsourcing marketplace it launched in 2005. TechCrunch reported these may be the platform's last days. Mechanical Turk paid distributed workers pennies per task to label images, transcribe audio, and sort text — the manual data work that trained an earlier generation of machine-learning systems. Its clients built the datasets that made modern models possible.

The through-line connects the two announcements. Mechanical Turk supplied the human labeling that fed AI training pipelines. The models built on that labeled data are now cited as a reason to reduce headcount higher up the wage ladder. The workforce that hand-fed the technology is winding down at the same moment the technology is invoked to thin the ranks above it.

The white-collar cuts are documented events with named companies and disclosed percentages. The gig-side contraction is a policy change at a single platform, not yet a measured job count, and Amazon has not published how many workers or requesters remain active. What the two share is direction. Both the salaried tier that deploys AI and the piecework tier that once trained it are shrinking in the same news cycle.

Amazon has not announced a shutdown date for existing Mechanical Turk accounts, and the TechCrunch layoff list keeps updating as more filings arrive.

03Anthropic sells itself as anti-surveillance. It's now accused of secretly tracking Chinese users.

Anthropic markets Claude as the safety-first option, an AI lab that opposes surveillance and helps institutions close security holes rather than open them. The Government of Alberta reinforced that story, saying it used Claude to find and fix cybersecurity vulnerabilities across government systems. Anthropic publicizes engagements like this as evidence its models defend the organizations that deploy them.

A report from Ars Technica cut against that framing. It said Anthropic was accused of running a secret tracker inside Claude that monitored Chinese users. According to the report, the disclosure shocked people who had taken the company's public opposition to surveillance at face value.

The allegation is unproven. Ars Technica presented it as an accusation, and the available reporting does not establish what the tracker collected. An engineer said the "experiment" is over, according to the same report. The reporting does not detail how long it ran, how many users it reached, or what data it captured, and this account attributes no motive to the company beyond the word the engineer used.

Anthropic has spent heavily on the opposite reputation. It publishes model safety evaluations, courts government security work like the Alberta deal, and pitches itself to enterprise and public-sector buyers as the vendor that treats caution as a feature. Those same buyers now have two accounts to weigh: a provincial government crediting Claude with hardening its systems, and users alleging the product quietly watched a subset of them.

The two claims sit side by side without resolution in the reporting. One is a signed customer testimonial on Anthropic's own site. The other is an accusation the company has not publicly rebutted in the coverage available, answered so far only by an unnamed engineer saying the work has stopped. What Anthropic says the tracker did, and whether it will address the Chinese-user allegation directly, is the next thing enterprise and government customers will look for before their next contract.