When AI Disclosure Is Due

On 20 July 2026 the European Commission published guidelines to help providers and deployers meet the AI Act's transparency obligations, which apply from 2 August 2026. The Commission's stated aim is that people can recognise when they are dealing with AI, and when content has been generated or altered by it, so that a fluent system cannot pass itself off as something else. For an operator the task is narrower. Before the August date, list the live systems that talk to customers or the public, and check each one for a user notice, a content mark, an editorial owner and a rule for stopping publication when any of those is missing.

Article 50 transparency adds a recurring cost to the workflow, and it is better budgeted now than discovered later through an incident.

What Changed This Week

Article 50 of the EU Artificial Intelligence Act covers transparency for certain interactive AI systems and for AI-generated or manipulated content. The Commission's July 2026 transparency guidelines set out what falls to providers and what falls to deployers. Providers must design interactive systems so that users know they are dealing with AI, and must support machine-readable marks that help detect generated or manipulated content. Deployers must tell people about deepfakes, about AI-generated public-interest content that has had no human review or editorial control, and about emotion recognition or biometric categorisation systems.

2 August 2026 is close, and for planning purposes the date counts for more than the definitions. Other chapters of the AI Act have been debated and delayed, but public reporting has kept Article 50 on the August timetable. A company that assumed the whole regulation had been postponed will still owe customers a disclosure on that day.

American and Asian companies are caught too when the system is offered in the EU or reaches EU users. They can also be caught by contract, when enterprise buyers ask for EU-grade disclosure as a commercial term, which turns the same notice screen into a sales and product problem wherever the legal team happens to sit.

What this changes for operators

Once the date is fixed, three operating ideas get harder to put off: write the rules before choosing tools, build disclosure into the workflow, and name who owns the residual risk when something the company says in public goes wrong. The July guidelines attach a deadline to each of them.

Rules before tools. A team that spends this quarter comparing models will end up bolting a disclaimer onto a product that is already in market. By now it should have a one-page set of rules listing which customer conversations are AI-assisted, which public statements need human editorial control, and which kinds of content must carry a mark before they leave the firm. See also Decide the Rules Before Buying Tools.

Disclosure as workflow. When to Disclose AI Use tracks how content is generated, the path it travels, who approves it, and the point at which missing disclosure must stop publication. For in-scope systems, Article 50 of the EU AI Act (see the Commission's transparency guidelines) makes the first half of that checklist mandatory. The public sees only the mark and the user notice, and neither holds up unless generation, review, routing and publication behind them are already in order.

Residual ownership. If a chatbot speaks in the brand's voice and a customer says they were deceived, the cost lands on the deployer who put the system in front of that customer, and it rarely passes cleanly to the model vendor. Transparency reduces some of that exposure, but someone inside the company still has to own the workflow by name. See Insurers Are Writing AI Out of Liability Policies.

Four Operator Mistakes

Mistake one: disclosure written as a marketing line. A "Powered by AI" footer does not meet the requirement for interactive systems, because users need to know at first contact that they are talking to a machine. Product, support and legal have to agree on where the notice appears and what wording is honest.

Mistake two: handling public-interest content like an ordinary draft. The guidelines and the commentary around them give particular weight to AI-generated public-interest material that has had no human review or editorial control. A research blog, market note or founder essay published largely as machine output with no human owner carries more risk than an internal brainstorm, so editorial control needs a named person responsible for it.

Mistake three: leaving agents out of scope. Agentic systems whose outputs users see directly fall under the same transparency logic. If an agent emails customers or speaks through a channel that looks human, its action rules need a disclosure step, written before launch.

Mistake four: waiting for a complete insurance or policy product. As Insurers Are Writing AI Out of Liability Policies argues, regulatory disclosure and insured residual risk are separate problems, and I would not expect the market to produce a product that solves either before August.

Scoring each system in one week

Before 2 August, score every AI system that faces customers or the public against the five rows below.

What This Digest Is For

Earlier memos covered rules before tools, whether an AI strategy holds up, disclosure checklists, accountability gaps and insurance coverage gaps. This digest applies that frame to one regulatory event so an operator can act on it this week. For the longer checklists, start with When to Disclose AI Use and Is This AI Strategy Real?

Start with the five customer- or public-facing AI systems that would do the most damage if they went wrong. For each, name a disclosure owner, a notice design, a marking rule where media is involved, and a rule that stops publication when any of those is missing. A company that cannot fill in that list is not ready for a rule under which silence about AI is itself the breach, and from 2 August each of those systems has to tell people it is a machine while it is talking to them.