Copilot users saved two hours a week on email, and Meta drops its AI usage score

Microsoft's Copilot delivered the time it promised. Sixty-six large firms let researchers give the tool to some of their workers and not others, 7,137 people in all, for six months. Of those given access, 80% used it, and in the second half of the experiment they spent two fewer hours a week on email and less time working outside normal hours. Eleanor Dillon and three co-authors published the results through the National Bureau of Economic Research.

The rest of the working day was where it had been: no shift in the quantity or the composition of anyone's tasks, and no real change in time spent in meetings. What Copilot saved, it saved for one person at one desk, in the work they could change without asking anybody.

Someone pastes a supplier contract into a chat window and comes to the renewal meeting knowing which clauses will cost money. The colleague handling the next contract never sees the question, the answer, or the part the tool got wrong.

Nobody outside that window sees what the work costs, either. Barracuda's chief information officer, Siroui Mushegian, told InformationWeek that AI features once sold on a licence are moving to token-based billing, so the size of the bill depends on how much staff use it and is known only after the work has run. She set up a council drawn from across the company that approves which AI tools staff may use, before anyone starts putting work through them.

Anthropic has been signing Claude users out of every active session, deleting the payment cards saved on their accounts and refunding charges that somebody else ran up. None of those accounts was opened with a stolen password. Common infostealer malware, the kind that arrives with an unofficial download or a bad app, copied the login cookies stored in the person's browser, and with one of those cookies an attacker is inside a signed-in account without the password and without the two-factor code, as Malwarebytes noted.

Anthropic told affected users it has no reason to believe the malware was related to Claude or installed through it. The laptop was taken first, and the Claude login was one of the things on it.

Refunds cover the money, and whoever holds that session can also read the account's past conversations, open its projects and download what was put into them: the pricing sheet, the draft contract, the customer list a colleague pasted in last month.

A year ago, 0.3% of the searches that reached John Lewis products came through an AI assistant. It is now 2.5%, and Peter Ruis, the outgoing managing director, says it is growing in every age group.

So the department store is making a chatshow. Gift List, hosted by the television presenter Angela Scanlon, goes out on YouTube, with guests discussing the presents they have been given, the good ones and the bad ones, and a new social media studio behind it.

The reason it is video and real voices is what the assistants read when a shopper asks them what to buy. A product page is the shop describing itself, and John Lewis has concluded that people talking on camera, rather than website copy, is what gets its products named by ChatGPT and Gemini.

The format is borrowed from its sister chain Waitrose, whose Dish podcast John Lewis counts a success, and clips from Gift List will run on its other social channels.

Gmail's search bar has a new Live icon on phones. Tapping it starts a spoken conversation with the inbox: a question in ordinary language, an answer drawn from the messages, a transcript on screen linking to the emails it used, and the option to interrupt and change subject without starting again. Docs Live does the same for drafting, Keep Live for notes.

Gmail Live comes with the Google AI Plus, Pro and Ultra plans; Docs Live needs Pro or Ultra; Keep Live is on Android only. All three are in English and on phones, worldwide, after months of testing.

Meta told employees this week that it will stop grading performance reviews on how much they use AI tools.

In July, 26 laid-off employees sued the company in federal court in California. Their 71-page complaint says the termination list was not put together by managers who knew the work, and names the systems it says produced the list: an internal assistant called Metamate, second brain agents trained on an employee's own messages and documents, keystroke and screen monitoring, algorithmically assisted performance ranking, and dashboards counting each employee's AI token usage. The complaint argues those systems marked down people who were legally away on medical or parental leave.

Judge William Orrick declined to pause the separations, which began on July 22, and wrote that the case raises serious questions about the merits.

A usage dashboard is cheap to build from systems a company runs anyway, and it counts the opening of a tool. Meta has now taken that number out of its reviews.

Between August 2025 and June 2026, freelancers on Freelancer.com were offered 10,760 jobs tagged with phrases like correct AI, AI hallucination and AI error, 87% more than in the year before. The Guardian, which obtained the platform data, reports a 70% rise year over year in the same repair work on Upwork.

Freelancers told the Guardian that clients budget these repairs as quick and cheap, when fixing a garbled logo or a robotic voiceover often takes as long as making one from the start. So a company pays twice, once for the generated version and once for the person who makes it usable.

The largest category is graphic design, ahead of video editing, proofreading and content writing.