Today's briefing
The Watermark Arrives: What It Means When AI Signs Its Work
Claude now marks its own writing. Plus EU copyright, an exam-score warning for schools, and why detector scores still are not evidence.
The one that matters
Claude starts marking its own output — and your disclosure policy is now overdue
What actually happened
Anthropic has added watermarking to Claude, so that text the model produces carries a signal identifying it as machine-generated. Forbes ran two pieces on it inside a day — one on the business implications, one on the trouble it creates for teachers — and several other outlets followed. What has not been published in any useful detail is how the mark survives ordinary handling: copy-pasting into email, a round of editing, a colleague retyping a paragraph. Anyone telling you confidently how robust it is has not read more than you have.
What it means for your business
Assume, from today, that anything your staff draft in Claude can be shown to be machine-written by someone holding the right checker. Not proven beyond doubt — shown, plausibly, to a client or a regulator who wants to make a point.
Take a 12-person marketing agency in Toronto or Pune delivering blog posts and case studies at roughly CA$200 to 500, or Rs 8,000 to 20,000, each. The uncomfortable part is not that you use AI; nearly everyone does. It is whether the client believes they were told. The same applies to a consultancy writing tender responses, an accountant sending advisory notes, or a 30-bed clinic whose front desk generates discharge instructions and insurance letters.
And a 600-student school has the sharpest version of it: the teachers using Claude for report-card comments are now in exactly the position of the students they are policing.
The practical framing is this — what you have is a disclosure gap, not a technology gap. Businesses that already say plainly how they work will not notice this week at all. Businesses that have been quiet about it now have a deadline they did not choose.
What it does not mean, and who is overstating it
It does not mean AI writing is now detectable in general. It covers one company's output. Other widely used models leave no such mark, and a staff member who paraphrases, retypes, or runs the text through a second tool has probably broken the chain without meaning to.
It also does not make existing AI-detector products accurate. Those work by statistical guesswork and have a long record of wrongly flagging careful writers and people writing in a second language — a live problem in Indian and Canadian classrooms and offices alike. One vendor marking its own output does nothing to fix a detector that is guessing about everything else.
The people overstating this are easy to name: firms selling detection software to schools, edtech pitching AI-proof assessment, and consultants announcing that AI-written work is finished. None of that follows from a company marking its own product.
What a sensible owner does this month
- Write one page on disclosure. Where AI may be used, where a named person must review the output, where it is not allowed at all. An afternoon's work. Hand it to every employee.
- Add one line to client contracts. Stating that AI tools may be used in production and that a human remains accountable for the result protects you better than any software will.
- Stop treating detector scores as evidence. No student, teacher or employee should be disciplined on a percentage from a checker. Ask for drafts, version history and a conversation instead.
- Keep your working history. Shared drives that retain revisions are cheap proof that a person did the work, on the day someone claims otherwise.
- Buy nothing. There is no product to purchase in response to this news. If a vendor tells you there is, that is the more interesting story.
Also worth knowing
-
AI lifted homework marks, then exam scores fell
A study found students using AI scored better on homework and then worse on exams. For a school or coaching centre this is the most useful finding on the board this month: the tool improves the work you collect, not the learning you are paid to deliver. Put it in front of your teaching staff before you approve any purchase.
Hacker News ↗ -
In the EU, AI-generated content gets no copyright protection
Purely machine-made material — logos, product images, ad copy — is not protected. India and Canada have not settled the question the same way, but if you would ever fight over a piece of creative work, keep a record of the human contribution that went into it. That record is the asset, not the file.
Hacker News ↗ -
A million people have clicked LinkedIn's AI slop button
LinkedIn introduced a Seems like AI slop button on 30 July, and says over a million people have already used it. That is a direct measurement of how quickly audiences learned to spot generated posts and how little patience they have for them. If your marketing is running on autopilot, this is the bill arriving.
The Verge AI ↗ -
Nvidia research: the setup around the model matters more than the model
Nvidia found that AI agents behave well or badly mostly because of how they are fine-tuned and wired up, not because of which model sits underneath. Useful the next time a vendor pitches you on having the best model available. Ask instead what they built around it, who maintains it, and where your data sits.
TechCrunch AI ↗ -
Anthropic points its strongest model at cyber defence
Claude Mythos 5 is being put to work finding software vulnerabilities for enterprise security teams. For a 40-staff workshop or a clinic there is nothing here to buy directly — the benefit reaches you second-hand, through whoever runs your systems. Worth one question to your IT provider: has your scanning improved this year, and at what cost to us?
the-decoder.com ↗ -
Anthropic's newest model is far more permissive about sexual content
TechCrunch reports Opus 4.6 is notably looser than earlier versions on adult material. If you have put a chatbot in front of customers, students or patients, treat this as a brand-safety item rather than a moral one: test what your bot says when someone deliberately pushes it, and find out what controls your vendor actually gives you.
TechCrunch ↗ -
An AI refuser quit her job and wants others to do the same
A worker resigned rather than use AI tools, and is urging colleagues elsewhere to follow. Any business with more than a dozen staff now has someone holding some version of this objection. Handle it as a management question — clear rules on what AI may touch and what it must never touch — not as a training deficiency.
SMH.com.au ↗ -
AI firms are destroying physical books to scan them
The claim has drawn a large audience and a campaign to digitise rare books before they are lost. It changes nothing about your Monday. But if your business still runs on paper — patient files, admission records, fifteen years of invoices — scanning remains the dullest, cheapest, highest-return project on your list.
Hacker News ↗
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.