AI Regulation Tracker  /  Enforcement findings

Belgium Published Numbers on Undisclosed AI in Influencer Advertising

On June 30, 2026 the FPS Economy published the results of an international consumer-protection sweep its Economic Inspection coordinated. The publication is findings and best practices, not a new rule, but it tells you how a European regulator intends to read AI-generated creative under advertising law it already has.

By Anthony Guerriero  ·  Reviewed by Anthony Guerriero  ·  Published August 3, 2026  ·  Updated August 3, 2026  ·  Enforcement findings

Key facts

Regulatory briefing

Instrument

FPS Economy publication of ICPEN sweep findings, Influencers en AI: reclame nog te vaak niet duidelijk vermeld, June 30, 2026

Authority

FPS Economy, SMEs, Self-Employed and Energy, Economic Inspection; sweep run under ICPEN

Jurisdiction

Belgium, within the EU single market

Status

Published findings and best-practice recommendations. Not an instrument and not an AI Act action

Bindingness

Non-binding. The publication creates no new duty; the advertising-identifiability duty it enforces is pre-existing Belgian economic law

Effective date

Not applicable. Published June 30, 2026

Primary source

FOD Economie Newsroom, June 30, 2026

What Belgium published, and what it is not

Start with the classification, because it is the thing most trackers will get wrong. This is not a law, a decree, a royal decree or an AI Act measure. It is a press publication by the Belgian FPS Economy, SMEs, Self-Employed and Energy setting out what an international consumer-protection sweep found, together with best practices for content creators and a pointer to a reporting channel, ConsumerConnect. It imposes no obligation that did not exist on June 29, 2026. It names no company, no fine and no docket.

It is also not market surveillance under Regulation (EU) 2024/1689. The sweep was run under ICPEN, the International Consumer Protection and Enforcement Network, and the legal hook the Belgian authority cites is ordinary Belgian advertising law. If you see this described as an AI Act enforcement action, that description is wrong, and the distinction matters because the two regimes have different addressees, different procedures and different remedies.

So why cover it at all. Because the value of a document like this is not its legal force. It is the disclosure of how a national enforcement body intends to apply rules it already has to creative that is new. Regulators telegraph their reading before they run cases. This is a telegraph, and it is unusually specific.

Here is the passage that does the work, in the original Dutch:

Zo blijkt uit de enquête dat 69 % van de influencers die gebruikmaken van AI-gegenereerde of -bewerkte beelden dat niet stelselmatig aangaf. Daarnaast maakte meer dan de helft van de onderzochte virtuele influencers claims over persoonlijke ervaringen, zoals reiservaringen of gezondheidsresultaten. Dergelijke ervaringen kunnen echter niet als persoonlijk worden beschouwd als ze door een volledig door AI-gegenereerd profiel worden gepubliceerd. Zulke praktijken kunnen consumenten dus misleiden.FOD Economie, Influencers en AI: reclame nog te vaak niet duidelijk vermeld, June 30, 2026

In substance: 69% of the influencers using AI-generated or AI-edited images did not systematically say so; more than half of the virtual influencers examined made claims about personal experiences such as travel or health results; such experiences cannot be regarded as personal when published by a profile that is entirely AI-generated; and practices of that kind can therefore mislead consumers.

The numbers, and how to read them honestly

The sweep ran from March 23 to April 8, 2026. Twenty consumer-protection authorities took part. The Belgian release describes the work as carried out under the coordination of the Economic Inspection of the FPS Economy and refers to the study as being onder leiding van de FOD Economie, that is, under the FPS Economy's lead.

One counting caveat is worth stating up front, because it is the sort of thing that gets copied wrongly into a hundred summaries. The figure 228 appears in the Belgian primary source as 228 controles, meaning checks, of which the Belgian Economic Inspection performed 25. ICPEN describes 228 influencers examined. Those are not the same unit. If you are quoting the number in a memo, quote it as checks and note the discrepancy rather than asserting a headcount the Belgian source does not give.

On the general disclosure picture, the published findings are that 92% of the influencers examined posted commercial content while only about one in five disclosed it systematically. In Belgium specifically, 80% did not systematically disclose, which the release sets against 90% in the 2023 EU-wide study.

Read that movement carefully. It is an improvement in the Belgian non-disclosure rate against an earlier study, and it is still four in five. I would not present it internally as progress. The operative fact for a compliance function is that the baseline behaviour in this channel is non-disclosure, which means an authority looking at your campaign is starting from an expectation of failure rather than an expectation of compliance. That changes the posture of any conversation you have with it.

The AI-specific number, the 69%, is narrower and more useful. It is not a measure of how many influencers use AI. It is a measure of, among those who do use AI-generated or AI-edited images, how many failed to say so systematically. The qualifier systematically is doing real work in that sentence. It is not asking whether a creator ever disclosed. It is asking whether disclosure was the consistent practice. Occasional labelling counts as failure on that test, and in my experience occasional labelling is exactly what an unmanaged influencer programme produces: the careful creator labels, the fast one does not, and nobody at the brand is checking.

The virtual influencer finding is the sharp one

The disclosure statistics will get the headlines. The reasoning about virtual influencers is what I would actually brief a general counsel on.

The authority's position, as published, is that experiences cannot be regarded as personal when they are published by a profile that is entirely AI-generated, and that presenting them as personal can mislead consumers. More than half of the virtual influencers examined made exactly that kind of claim, and the examples given are travel experiences and health results.

Notice what that theory does not require. It does not require a new AI statute. It does not require proving that the audience was deceived about the existence of AI as such. It runs on the ordinary proposition that a factual claim in advertising has to be capable of being true, and a first-person testimonial from an entity that has never travelled anywhere and has no health is not. That is a familiar misleading-advertising argument wearing new clothes, and familiar arguments are the ones enforcement bodies are most comfortable running.

The health example should worry anyone in a regulated vertical. A synthetic persona saying a supplement, a device or a treatment worked for it is, on this reading, a claim about a personal result that cannot have occurred. Add a disclosure label and you have fixed the AI transparency problem while leaving the testimonial problem entirely intact. Those are two separate defects, and the sweep findings implicitly identify both.

I would treat that as the transferable lesson rather than as a Belgian curiosity. The record here is Belgian and I am not going to characterise any other country's law from it. What I will say is that the reasoning as published does not lean on anything AI-specific: it leans on the proposition that a personal-experience claim has to be capable of being personal. That is an argument built out of general advertising principles rather than out of Belgian drafting, and if you are running the same virtual-influencer script in several markets, it is worth having your own counsel test it against each of them rather than assuming the exposure stops at the Belgian border.

Who published it, and why the department matters

The publishing department is the FPS Economy, SMEs, Self-Employed and Energy. Its own AI Act page carries this sentence: "In België is de FOD Economie, K.M.O., Middenstand en Energie verantwoordelijk voor de algemene coördinatie van de implementatie van de AI Act." In English, the FPS Economy is responsible for the overall coordination of AI Act implementation in Belgium. That page carries a last-updated date of March 11, 2025.

So the body that coordinates AI Act implementation in Belgium is the same body whose inspection arm ran this sweep and wrote these findings. That does not convert the publication into an AI Act measure, and I am not going to pretend it does. But when the coordinating department publishes its reading of undisclosed synthetic content roughly five weeks before the AI Act's general application date of August 2, 2026, that is a reasonable indication of institutional temperament, and temperament is what you are trying to forecast when you plan a compliance budget.

On the wider Belgian governance question I want to be precise, because this is an area where secondary trackers and AI-generated summaries have been confidently wrong. The BIPT, the Belgian telecoms regulator, publishes AI Act guidance for operators on its own pages and states there that a digital helpdesk has not yet launched. Whether BIPT has been formally designated as a market surveillance authority under the AI Act is a separate question, and no adopted instrument making that designation was located in the sources reviewed as of August 3, 2026. Treat any claim that the designation has been made as unverified until you can point at the instrument.

Belgian Senate written question 8-98, put by Peter Van Rompuy on March 12, 2025 and answered on May 5, 2025 by the Deputy Prime Minister and Minister of Employment, Economy and Agriculture, confirms the structural requirement that national competent authorities comprise at least one market surveillance authority and one notifying authority, and at that point was still describing the Belgian governance model as under construction. Nor was any enacted Belgian AI Act implementing law identified in the sources reviewed. That is the honest state of the record: coordination is assigned, the full designation picture is not something I can evidence today.

What a US company should actually do

The exposure here is not exotic. If your creative reaches Belgian consumers, Belgian advertising-identifiability law is in play whether or not you have an entity there, and the sweep tells you what the inspectorate is looking at.

Four pieces of work, none of which needs a new budget line.

First, inventory. Find out which of your live influencer and paid-creator content uses AI-generated or AI-edited imagery. Most marketing organisations genuinely do not know, because the editing happens on the creator's machine and never touches an internal system. Until you can answer that question you cannot manage the risk, and the 69% figure suggests the honest answer is usually more than you think.

Second, make disclosure systematic rather than available. The finding is about consistency, not about whether a label exists somewhere in your programme. That means a contractual requirement in the creator agreement, a specified placement, and a sampling check by someone who is not the person who booked the campaign. A policy nobody audits produces exactly the pattern the sweep measured.

Third, separate the two defects. Labelling content as AI-generated addresses transparency. It does not cure a first-person claim that a synthetic persona cannot truthfully make. If you run virtual influencers, go through their scripts for experience claims, especially anything touching health outcomes, and rewrite them into claims the account can actually make. That is a copy-review task, and it is the cheapest item on this list.

Fourth, do not let the non-binding classification decide your priority. The AI Act's transparency and marking duties in Article 50 land on December 2, 2026 under the Digital Omnibus timetable, and the work you would do to satisfy a consumer-protection inspector on labelling is substantially the same work. Doing it once, now, on a schedule you control, is materially cheaper than doing it twice under two different regulators' clocks. That is the whole practical argument, and it does not depend on this publication being binding, which it is not.

Last verified: August 3, 2026 against the primary sources listed above.