A quieter Wednesday after a heavy week — but with a lead story that lands right on the nerve where AI, privacy and the workplace meet: Meta has paused an AI programme that was quietly watching how its own staff work.

Meta pauses the AI that was watching its employees

Meta has suspended an internal effort called the Model Capability Initiative — a programme that, since April, captured US employees' mouse movements, clicks, keystrokes and occasional screenshots to train AI on how people actually do their jobs. It hit pause this week after a security misconfiguration left that sensitive employee data readable far more widely inside the company than intended.

The programme was paused on Monday, after Reuters reviewed internal documents showing the data had been left accessible. A Meta spokesperson confirmed it on the record: the company has "no indication at this time that any data was improperly accessed by Meta employees, [but] we're pausing it while we investigate." The exposed material is said to have included private conversations, performance data and transcriptions — and internal pushback had already forced Meta to add a "30-minute tracking break" button earlier on.

The reason this matters beyond Meta: training AI on detailed recordings of employees is a privacy-and-labour minefield, especially under European rules. "Consent" is a shaky foundation when it's an employer asking — the power imbalance is exactly why GDPR treats workplace monitoring so carefully. It's a sharp early example of a tension that's coming for a lot of companies: the most valuable data for training workplace AI is often a detailed record of the workers themselves.

Mistral ships a new document-reading model

On the product side, French lab Mistral released OCR 4, a document-intelligence model aimed at the unglamorous-but-huge job of turning scanned and photographed documents into structured, machine-readable data. It returns not just text but bounding boxes, typed-block classification and per-word confidence scores, across what Mistral says is 170 languages.

Mistral claims best-in-class accuracy — a 72% average win rate against rival systems in blind human-preference tests, and top scores on the OlmOCR and OmniDoc benchmarks — at $4 per 1,000 pages, with a self-hostable single-container option. Those figures are the company's own, so treat them as a vendor's claims until independently tested. But the direction is clear: reliable document extraction is becoming cheap and commoditised, which is precisely the boring-but-foundational layer a lot of real-world AI automation is built on.

"It is time to come clean": the UN turns to AI's footprint

At London Climate Action Week on Tuesday, UN Secretary-General António Guterres put AI's environmental cost on the agenda, calling on AI companies to publicly disclose their carbon, water and land footprints and to power their facilities with 100% renewable energy by 2030. "No more hidden costs," he said. "It is time to come clean."

The numbers he cited frame the worry: data centres consumed roughly 1.5% of the world's electricity in 2025, a share that could double to around 3% by 2030, with coal still supplying about 30% of data-centre power against renewables' 27%. It's a transparency push, not a binding rule — but it's a notable shift in where the AI-policy conversation is heading: away from purely safety and capability, toward the physical bill for the build-out.

One to sit with: barely anyone uses AI to get their news

A useful piece of context from this week's Reuters Institute Digital News Report: despite the hype, using an AI chatbot to actually get the news remains rare — about 10% of people globally do so weekly, and just 4% in the UK, the lowest of any market surveyed. It's higher among the youngest cohort (17%) than the oldest (5%), and only a small fraction of people click through from a chatbot's answer to the original source.

For a news business, that last point is the one to watch — but the headline is a healthy reality check on a noisy year: the technology is moving far faster than the public's habits around it.

Quiet news day, honestly told. We're watching the WEF "Summer Davos" meeting in Dalian (running to Thursday) for any fresh AI announcements, and yesterday's two big threads — the Five Eyes cyber warning with OpenAI's defensive answer, and Oracle naming AI in its job-cut filing — for follow-ons.