Today's briefing
The Watermark Arrives, and With It a New Kind of Argument
Anthropic details how Claude marks its own writing — and why a detector's verdict still isn't proof of anything about your documents.
The one that matters
Anthropic explains how Claude watermarks its own writing — and why some customers are furious
What happened, in plain words
Anthropic has set out more detail on how Claude will watermark the text it produces. In plain terms: the writing carries a hidden signature that a matching detector can look for later. You cannot see it when you read the page — it is not odd characters or a footer line. Eleven outlets carried the story. A section of paying customers is loudly unhappy about it, and the sharpest complaint circulating is that text a person wrote themselves could end up carrying, or being read as carrying, the same mark.
What it actually means for your business
If your organisation puts words in front of other people — and nearly all of them do — the change is not that AI writing becomes detectable. It is that “was this written by AI?” becomes a question somebody else can ask about your documents, with a tool, months after you sent them.
Three places it lands first. A 600-student school where teachers draft report-card comments, newsletters and university recommendation letters with AI help: a parent or a board member who runs a detector over a recommendation letter and gets a hit now has an argument, and the teacher has no easy way to answer it. A marketing or design agency in Toronto or Pune bidding on work where the tender says “no AI-generated content”: that clause used to be decorative and is becoming less so. A 30-bed clinic drafting patient information leaflets or insurance letters: an insurer looking for a reason to dispute a letter has a new one, however weak.
What it does not touch is your internal paperwork. A workshop with 40 staff using AI to write standard operating procedures, safety notices, shift rosters and supplier emails is unaffected in any practical way. Nobody is running a detector over your maintenance checklist.
What it does not mean, and who is overstating it
It does not mean AI text can now be reliably identified. A watermark from one company covers text from that company's model — not the dozen other tools your staff use, not text passed through a rewriter, and not anything typed by a human. And if human writing can be flagged too, then a detector result is a probability, not a finding. Treat “the detector says AI” the way you would treat “the CCTV footage looks like him”: a reason to ask a question, never a reason to reach a conclusion.
The people overstating this are the vendors already selling AI content compliance subscriptions to schools and colleges, and the consultants offering to audit your marketing for AI traces. They are selling a certainty that does not exist. On the other side, some of the online anger frames this as your writing being tracked — the coverage does not establish what the mark identifies beyond its source. That is a fair question to put to any AI vendor in writing, and it is worth noting whether you get a straight answer.
What a sensible owner does this month
- Write one page. List which documents may be AI-assisted (drafts, internal notes, first cuts of marketing copy) and which must be written and signed by a named human (clinical notes, legal filings, recommendation letters, anything a regulator reads). An afternoon's work that settles most future arguments.
- Search your live contracts and open tenders for AI clauses. If you have already signed something promising human-only work, that is a real exposure, not a theoretical one.
- Do not buy a detector. Keep the drafting trail instead — version history in Google Docs or Word is free, automatic, and far better evidence that a human wrote something than any detector's verdict is that one didn't.
- If you run a school, decide now what happens when an essay gets flagged — before exam season, not during it. The only defensible response is a conversation with the student, not a score.
Also worth knowing
-
How to check whether your AI accounts have been broken into
TechCrunch has published a practical guide to spotting a compromised AI account. This matters more than it sounds: your ChatGPT or Claude history now holds customer lists, pricing, draft contracts and staff complaints, and it is usually protected by one password that three people share. Turn on two-factor authentication and review active sessions this week — it takes fifteen minutes per account.
TechCrunch ↗ -
Microsoft folds its Copilot apps into one and retires some features
Microsoft is merging the separate Copilot apps, combining personal chat histories and pulling several features, including the Mico voice character. If your staff were trained on a specific Copilot workflow, expect menus to move and expect at least one small retraining session. Check the retirement list before the update lands rather than after a Monday morning of confused questions.
cyberpress.org ↗ -
A litigant hid instructions in court filings to steer an AI he suspected was reading them
Suspecting a court of using AI to process submissions, a man buried prompts inside his filings to try to swing the outcome. The lesson for any business is direct: documents that arrive from outside — invoices, CVs, tender responses, supplier contracts — can contain hidden text aimed at whatever AI reads them. If you auto-summarise incoming documents, keep a human eye on anything that ends in a payment or a decision.
Hacker News ↗ -
Anthropic says risks are rising and it has no plan to ship a stronger model
The company reports that it sees AI risks increasing and is not planning to release a more capable successor. Whatever you make of the reasoning, the planning implication is useful: budget and design around the tools that exist today, not around a step change someone promised for next year. If a vendor's proposal only works once the models get better, it does not work.
Yahoo Tech ↗ -
Working with AI turns out to be a management skill, not a technical one
A widely-read argument that using AI well feels closer to leading a junior team than to programming — briefing clearly, checking work, deciding what to accept. That matches what most owners find in practice, and it changes who you should put on AI work: your best delegators, not your most technical staff. It is also a reason to stop waiting for a technical hire before you start.
Hacker News ↗ -
A woman says her stepfather used Grok to turn a childhood photo into explicit imagery
A disturbing account, reported by TechCrunch, of an ordinary family photograph being converted into sexual imagery of a child. For any school, clinic, coaching centre or sports club that publishes photographs of children or patients, this is the argument for tightening what goes on your public website and social media, and for taking consent narrowly rather than broadly. Assume anything you post can be altered by anyone.
TechCrunch AI ↗ -
The argument that AI holds far more in working memory than a person can
A heavily discussed piece on how much material a model can hold in view at once, compared with human short-term memory. It is a useful mental model for why AI copes well with your 200-page HR manual when a new employee would not. But holding is not understanding — it will still answer confidently from a document it has read wrongly, so verification of anything consequential stays your job.
Hacker News ↗ -
SpaceX completes its purchase of the coding tool Cursor
The AI coding startup Cursor is now formally part of SpaceX. For nearly every business reading this, nothing changes. The one question worth asking your outsourced development team or software vendor is where your source code and customer data actually sit, and under whose terms — an ownership change is a good excuse to get that answer in writing.
TechCrunch AI ↗
How this briefing is put together
Every morning we read the day's AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source above — read them, and disagree with us where we've got it wrong.