Everyone's work will carry the label, so the only thing that tells yours apart is whether it's accurate and of good quality.
Your Essential Report from the Frontlines of Applied AI
The EU's AI rules took effect on 2 August. Do you know what they mandate?
The AI companies now have to mark whatever their tools produce, in a form machines can detect. Text produced by AI will carry a marker you cannot see and a machine can.
You only have to declare AI yourself for deep fakes, and for text you publish to inform the public about something important. (Client decks, internal reports and proposals are not covered.) Even published text is exempt if someone genuinely read it and takes responsibility for it.
Our Institute's own view is that this changes very little. Everyone uses AI, so everyone's work will carry the mark, and nobody will stop because of it. A rule that applies to something everyone is already doing has low practical significance.
Which means the significance remains where it was: whether your AI-assisted work is accurate and of good quality.
Sources: Article 50 of the EU AI Act, and the European Commission's guidance on it.
So this week is about what checking actually involves.
Everyone's work will carry the label, so the only thing that tells yours apart is whether it's accurate and of good quality.
That makes checking more valuable than it was, and the rules agree — the exemption hangs on a person having genuinely reviewed the work. AI+ Librarians sets out five questions for it, called TRUST. Librarians hold them to a higher standard than most professions, because they are the standard-setters in their communities.
T — Traceable. Can every claim be traced to a source you can open? AI is confident about things it gets wrong, so don't publish a figure you can't source.
R — Relevant. Does it fit your actual situation? If it could have been written for anyone in any industry, it isn't close enough.
U — Unbiased. Is the topic represented fairly? Watch for overstatement, missing caveats, and anything a sceptical reader would question.
S — Specific. Does it actually say something? Sentences that sound meaningful without committing to anything usually mean the model is unsure.
T — Timely. Does it still hold? Rules, figures and platforms move, and training data has a cutoff.
How far you take it depends on where the work is going. A rough draft for yourself needs a glance; anything client-facing, regulated or public is worth all five, checked against primary sources.
What matters now is whether someone checked the work properly, and whether they would stand behind it.
- Advertising
New York now requires adverts to say when a performer is AI-generated
The Synthetic Performer Law took effect on 9 June: a conspicuous disclosure when an advert knowingly uses an AI-generated performer, with penalties of $1,000–$5,000. The first US labelling law aimed at AI creative work.
New York Synthetic Performer Law (statute); Manatt client alert
- Communications
The people writing the disclosures are largely unsupervised
Roughly 60% of PR and communications professionals name misinformation as their top AI concern. About the same proportion admit using generative AI without approval.
LexisNexis, AI in PR & Communications: 2026 Industry Report
- Public sector
UK government communicators have their own in-house AI
The Government Communication Service runs Assist, the Cabinet Office's own generative-AI toolset for government communicators — built rather than bought, and governed centrally.
UK Government Communication Service
- Leadership
76% of organisations now have a Chief AI Officer, up from 26% a year ago
The fastest-growing seat in the C-suite, and the clearest sign yet that AI governance has moved from a project to a permanent job.
IBM Institute for Business Value, 2026 CEO Study
A consultant presented an AI-drafted competitive analysis. Halfway through, the client's chief executive stopped her: where does this 34% market share figure come from? She checked afterwards — the AI had invented it, because a number in that range was plausible for a company of that type. The rest of the deck was sound. That one figure cost her an hour of damage control and weeks of rebuilt credibility.
I have the following factual claims from an AI-generated output. For each one, tell me: (1) whether you can verify it — naming the source and its date if you can, (2) your confidence level (high / medium / low / unable to verify), and (3) any corrections or caveats. If you cannot point to a source, say "unable to verify" rather than guessing. Here are the claims: [paste the claims].
Prompt 7.1 from AI+ Everyone, word for word.
Let AI take the first pass at 1.5 million paragraphs, and have a qualified professional sign off every result.
When MiFID II came into force, compliance teams faced roughly 1.5 million paragraphs to read, interpret and map to the parts of the business each obligation touched. No team covers that quickly by hand, and getting it wrong carries real regulatory risk.
In a 2018 pilot observed by the UK's Financial Conduct Authority, ING and Commonwealth Bank of Australia used AI to turn MiFID II into a structured set of obligations mapped to their business lines, with a law firm validating the legal interpretation. The AI did the first pass; qualified compliance professionals checked and owned the result, saving hundreds of hours. (The vendor puts the saving far higher — minutes against 1,800 hours — but that figure is the vendor's own and is not independently corroborated.)
The design is the point, and it matches this week's idea: the AI did the first pass and a qualified person took responsibility for the result, with a regulator watching.
