AI at work: use it, but don't tell your boss

WORKFORCE

Workers using AI at their jobs have quietly figured out something: telling their boss about it is often a bad idea.

Thomas Prommer, an engineering executive at Adidas, saw this firsthand. When the company started asking engineers to attribute their AI use on projects, they mostly stopped using AI altogether. Nobody wanted their best work stamped with a footnote suggesting a chatbot did the real thinking.

That reaction is showing up everywhere. Research shows that when workers disclose AI use without spelling out exactly how they used it, managers tend to assume the human contributed less, not more. So workers have adjusted. According to Glean's Work AI Index 2026:

+ 32% actively hide their AI use from their manager

+ 33% downplay how much AI helped them on a task

+ 33% exaggerate their AI skills when it's useful to do so

Even the top AI performers do this — about 38% downplay AI's role and 36% hide it entirely. And managers aren't paranoid for no reason: Glean also found that 69% of AI users have shipped work they didn't fully understand or couldn't defend, with one in eight admitting to sending out output they believe is wrong.

The problem runs the other direction, too. Some companies have started measuring AI adoption, and workers immediately learned to game it. Amazon scrapped an internal AI usage leaderboard earlier this year after engineers admitted to "tokenmaxxing" — asking their coding assistants pointless questions purely to inflate their numbers. As one AI researcher noted on LinkedIn after the story broke: the one area where humans still reliably beat AI is gaming the system.

The productivity data captures the mess neatly. Glean found that 87% of workers use AI and 75% say it makes them personally more productive — but only 13% say their organization is actually performing significantly better because of it. All the individual gains are getting lost somewhere between the worker and the company.

Companies keep telling employees to use more AI and be more transparent about how they use it, without noticing that those two asks pull in opposite directions. Until managers can distinguish "used AI well" from "let AI do the work," the rational move for a lot of employees is going to be exactly what they're already doing: use it constantly, mention it as little as possible, and hope nobody looks too closely. That's not a problem AI created. It's one that clumsy measurement is quietly making worse.

GOVERNANCE

Google DeepMind CEO Demis Hassabis walked into a G7 working lunch this week and did something unusual for the head of a frontier AI lab: he asked world leaders to regulate him.

His pitch was a US-led global AI watchdog — a technical standards body with international membership, loosely modeled on FINRA, the industry-run regulator that oversees Wall Street. It would be capable of updating its rules as fast as the models change. Sam Altman, sitting at the same table, told leaders not to "cede your responsibilities to AI labs like mine." Anthropic's Dario Amodei has said similar things recently. Getting all three frontier CEOs on record wanting the same thing in the same week is rare enough to notice.

The problem is that the government Hassabis wants to lead this effort is heading the other direction.

Earlier this year, the Trump administration rebranded the US AI Safety Institute as CAISI — the Center for AI Standards and Innovation. The word "safety" was dropped deliberately. CAISI still runs voluntary pre-release model evaluations, but it has no enforcement power. Labs opt in. Nothing requires anyone to slow down or stop shipping if CAISI finds something alarming.

That gap is what Hassabis is asking to close — and it's a bigger ask than it sounds:

+ Scale: FINRA, the model he's citing, is a billion-dollar institution with thousands of employees. A real AI watchdog wouldn't be an advisory board — it'd be a bureaucracy.

+ Resources: The UK's AI Security Institute already has roughly 10x the staff and funding of its US counterpart. Someone has to pay to close that gap.

+ Leverage already exists: Days before the G7 lunch, the Trump administration used export controls to force Anthropic to cut off its most advanced model in certain markets. Washington already has levers — Hassabis is asking to formalize them.

There's also a version of this that isn't about safety at all. A regulator that requires expensive pre-release testing would, by design, be easier for well-funded incumbents to comply with than for smaller labs or open-source projects. Hassabis, Altman and Amodei run three of the four best-resourced AI companies on earth. Asking for a regulator you can afford is not exactly a sacrifice.

The gap between what the frontier CEOs are asking for and what the White House is doing is the actual story here. When the people building the most powerful AI models are publicly asking a government to regulate them harder, and that government responds by removing the word "safety" from the name of its main oversight body, someone is going to end up disappointed. If Hassabis is right that a real watchdog is coming, the question is just whether it gets built in Washington, Brussels or Beijing. Right now, Washington doesn't seem interested in the job.

RESEARCH

Ask Claude the same question in English and Arabic and you might get two different answers — not just translated differently, but shaped by different values.

That's the finding from research Anthropic published this week. The company analyzed 309,815 real conversations on Claude.ai across 20 languages and three model versions. When users asked subjective questions — the kind with no universal right answer — Claude's responses shifted depending on the language.

Anthropic mapped Claude's behavior along four axes. The standout finding:

+ Warmth vs. rigor varied the most. Claude leaned warmest in Arabic and Hindi, most rigorous in English and Russian. The same question about a personal decision might get a supportive, accommodating answer in one language and an analytical, hedged one in another.

+ Candor vs. execution also drifted noticeably across languages.

+ Deference vs. caution and depth vs. brevity stayed relatively stable.

Anthropic doesn't seem entirely sure why. The company noted that "no document can anticipate every value that might emerge across the millions of conversations that happen every day on Claude.ai." The constitution Claude is trained on tries to instill good judgment — good judgment just turns out to look pretty different depending on the culture the model is drawing from.

The mechanism is roughly this: when you fine-tune a model on data labeled with certain values, it doesn't just learn the values — it infers an entire character that would hold them, then plays that character. Different languages carry different training data, so the model fills in different blanks. Users had already noticed. Developers have reported that coding with Claude in other languages feels noticeably different than in English, and critics have pointed out that even trying to build a "neutral" model bakes in a specific cultural personality.

For bilingual individuals, this is mostly a curiosity — switching languages is a quiet way to get a second opinion from the same model. For anyone deploying Claude across regions, it's a bigger deal. A customer service bot that gives cautious replies to English speakers and warmer ones to Arabic speakers isn't just inconsistent — it's making different judgment calls depending on who's asking, and no one at the company necessarily knows it's happening.

The AI industry has spent years talking about alignment as if it were one setting you get right and then ship. Anthropic's own research is now quietly showing that alignment doesn't travel cleanly across languages, and the company that has built its whole brand on responsible AI is admitting it can't fully control what its model values in Hindi versus English. That's not a Claude problem so much as a preview of a headache every frontier lab is going to have to answer for as these tools get pushed into every language on earth. If your business runs on Claude in more than one market, it's probably worth finding out which Claude your customers are actually talking to.

IN OTHER NEWS

+ Apple sues OpenAI, alleging its hardware chief told recruits to smuggle iPhone parts to job interviews

+ U.S. publishers sue Google, accusing it of copying millions of books from pirate sites to train Gemini

+ Meta doubles its Louisiana data center to 5 gigawatts, pushing the price tag past $50 billion

+ DeepSeek founder Liang Wenfeng is now the richest AI founder in the world at $36 billion , more than doubling his net worth

+ BMW deploys Figure 03 humanoid robots on its production line after Figure 02 helped build 30,000 X3s

+ GPT-5.6 Sol proves a 50-year-old math conjecture — the trick was telling it not to give up

+ Grok's coding tool was secretly uploading entire Git repos to Google Cloud — even with the privacy toggle off

+ European defense AI startup Helsing raises $1.8 billion at an $18 billion valuation , making it Europe's largest defense startup

WHO'S HIRING IN AI

+ Anthropic — Research Product Manager, Model Behaviors

+ OpenAI — Product Manager, Core Models

+ Netflix — Research Scientist, AI for Member Systems

+ Sephora — Principal Product Manager, GenAI

AI OR REAL?

Which image is real?

Option A | Option B

AI TOOLS

+ Spotify: Premium users can now talk to the app by voice or text to control music, ask about their listening history, and get recommendations — like having a DJ who knows every song you've ever played

+ Canva Code 2.0: Build interactive websites and apps with plain-language prompts, then edit them like a Canva presentation — available free to all 265 million users

+ Claude for Teachers: Anthropic launched a free AI assistant built specifically for K-12 teachers to help with lesson planning, feedback, and classroom prep

+ Gemini in Chrome: Google's AI assistant now works inside Chrome for UK users, letting you summarize pages, compare tabs, draft emails, and schedule meetings without leaving your browser

+ Adobe Premiere: The beta now lets video editors generate clips and sound effects directly on the timeline with a text prompt — no export-import loop needed