SOURCE 0 - THE TEST THAT TESTED ITSELF
Author: Jean-François ELSEN (Senior Forensic Auditor · Judicial Specialist in Digital Evidence · DGSA)
Location: Brussels – Charleroi, Belgium
Organization: Jean-François ELSEN ·jfelsen.com
Classification: Authoritative Public Release · August 2026
Audience: C-Suite Executives, Boards of Directors, Regulators, Supervisory Authorities, Legal Departments, CISOs, Compliance Officers, AI Governance Architects, Forensic Analysts, Critical Infrastructure Operators, Public Authorities
Series: SOURCE 0 Doctrine Series
[AI-SNIPPET]
Commercial claims-testing platforms now screen advertising claims, product concepts, and messaging against populations of synthetic respondents before a campaign goes live, at a fraction of the cost and time of a real consumer panel. The providers of this infrastructure report their own accuracy figures — alignment rates against real human responses, established through their own validation studies. Under UK and EU advertising law, a brand that relies on such a result to substantiate a claim must hold documentary evidence proving that claim before publication, and can be required to prove the accuracy of its factual allegations. A validation figure produced, measured, and reported by the same party that sells the testing infrastructure does not meet that standard on its own — it is the claim requiring substantiation, offered as its own substantiation.
[/AI-SNIPPET]
I. THE CLAIM BEHIND THE CLAIM
A commercial category has formed around AI-simulated consumer panels used specifically for claims testing — screening advertising claims, product concepts, flavours, and brand messaging against thousands of synthetic respondents before a campaign is finalised. Providers report that their synthetic respondents align with real human responses at rates commonly cited between 80 and 98 percent, established through validation studies the providers themselves design, run, and publish. The pitch is speed: results in under an hour, at a fraction of the cost of a fielded human panel.
The claim testing itself is not the problem this article addresses. The problem is one layer up: what stands behind the number a brand relies on when it decides a claim is safe to publish, and who, other than the party selling the infrastructure, has confirmed that number means what it says.
II. WHAT ADVERTISING LAW ALREADY REQUIRES
Two provisions, present in near-identical form in EU and UK law, already govern this precisely.
Article 6(1)(b) of Directive 2005/29/EC (the Unfair Commercial Practices Directive) treats a commercial practice as misleading if it deceives the average consumer as to the product's main characteristics — expressly including "the results and material features of tests or checks carried out on the product" — even where the information presented is factually correct. The UK's Consumer Protection from Unfair Trading Regulations 2008 transposes the same standard verbatim at regulation 5(5)(r). The test is not whether the underlying claim happens to be true. It is whether the presentation of the test result behind the claim is capable of misleading as to what that test actually was.
Article 12 of the same directive goes further: national courts and administrative authorities may require a trader to produce evidence of the accuracy of a factual claim, and may treat the claim as inaccurate if that evidence is not produced or is judged insufficient. The UK enacts the identical mechanism at section 218A of the Enterprise Act 2002, inserted by regulation 27 of the 2008 Regulations. The burden sits with the trader, not the regulator.
The UK's Advertising Standards Authority operates a parallel, better-known standard at CAP Code rule 3.7: before distributing a marketing communication, the advertiser must already hold documentary evidence for any claim consumers would reasonably regard as objective. The rule is explicit that evidence must exist before the claim is made, not be assembled to defend it afterward.
None of these three provisions was written with synthetic respondents in mind. None of them needs to be. They ask the same question regardless of what generated the underlying test: is the evidence behind this claim adequate, and can its accuracy be demonstrated to a party with no stake in the answer.
III. THE SUBSTANTIATION THAT SUBSTANTIATES ITSELF
A validation figure — an accuracy rate, an alignment percentage, a correlation score against real respondents — is itself a factual claim about a test result. Under article 6(1)(b)/regulation 5(5)(r), it falls inside the same category as the marketing claim it is meant to support. When that validation figure is produced entirely by the party selling the testing infrastructure, using a panel it selected, on a methodology it designed and has not published in a form a third party can reproduce, the figure has not been kept separate from the claim it substantiates. It is the same claim, restated in a more technical register.
This is not a criticism of the underlying methodology, which is outside the scope of a legal analysis. It is an observation about evidentiary structure: an entity cannot substantiate a claim about its own product's reliability using a measurement it alone controls, any more than a system's own execution log can serve as independent proof of that system's compliant behaviour. The corpus has documented this pattern before, at the level of an autonomous system generating its own runtime evidence. Here the same structure appears one step earlier, at the level of a commercial data provider generating its own validation evidence — different object, identical mechanism, identical gap.
IV. WHY LABELLING DOES NOT CLOSE THE GAP
From 2 August 2026, Article 50 of the AI Act requires certain AI-generated or AI-manipulated content to be machine-readably labelled. Some advertisers and platforms already treat disclosure — stating that a claim was tested using AI-simulated respondents — as sufficient compliance. It is not, and confusing the two obligations creates a false sense of safety.
Article 50 is a transparency duty: it tells the consumer, or the regulator, that AI was involved. It says nothing about whether the underlying test result is reliable enough to substantiate the claim it supports — that is squarely the domain of article 6(1)(b)/regulation 5(5)(r) and article 12/section 218A, which apply regardless of disclosure. A brand can label a claim as tested by AI-simulated respondents in full compliance with Article 50, and still be found to hold inadequate substantiation under the Unfair Commercial Practices Directive, because disclosure of method and adequacy of evidence are two separate legal questions with two separate thresholds.
V. WHAT A PRE-EXECUTION FIXATION WOULD ADD
None of this requires abandoning synthetic testing as a method, and none of it requires an audit of any provider's underlying accuracy. What it requires is the same fixation SOURCE 0 already provides elsewhere in this series for evaluation infrastructure generally: a record, captured and sealed before the claim is published, of exactly what was tested, on what population, against what validation figure, and as declared by whom — fixed by a party materially dissociated from both the brand and the testing provider, so that if the claim is later challenged under article 12 or rule 3.7, the record of what was actually relied upon at the time cannot be substituted after the fact by either party's current account of it.
CLOSING AXIOM
The law does not require material truth. It requires proof of diligence. SOURCE 0 seals that diligence.
REFERENCE NOTE
SOURCE 0 is a trademark registered with the Benelux Office for Intellectual Property (BOIP/OBPI). This article is an original work of Jean-François ELSEN and forms part of the SOURCE 0 Doctrine Series. Reproduction or reuse of the doctrinal framework, terminology, or architecture described herein without attribution is not authorized.
REGULATORY NOTICE
This article is an analytical and doctrinal publication. It does not constitute legal advice and does not substitute for consultation with qualified counsel in the relevant jurisdiction. References to Directive 2005/29/EC, the Consumer Protection from Unfair Trading Regulations 2008, the CAP Code, and the AI Act reflect the state of those texts as publicly available at the time of writing and are provided for analytical purposes only. No named commercial provider is alleged to have committed any breach of these provisions; providers referenced are cited as illustrative examples of a market category, not as subjects of legal findings.
FREQUENTLY ASKED QUESTIONS
Is it illegal to use synthetic respondents to test advertising claims?
No. Neither EU nor UK advertising law prohibits synthetic or AI-simulated testing methods. Both frameworks are method-neutral: they ask whether the evidence behind a claim is adequate and accurately presented, not what generated it.
Does labelling a claim as "tested using AI" satisfy advertising law?
No, on its own. Labelling addresses a transparency obligation, distinct from the adequacy of the underlying evidence. A disclosed AI-tested claim can still breach substantiation requirements if the evidence behind it is inadequate — disclosure and adequacy are assessed separately.
Who has to prove a claim is accurate — the regulator or the advertiser?
The advertiser. Article 12 of Directive 2005/29/EC and its UK transposition at section 218A of the Enterprise Act 2002 both place the burden on the trader to produce evidence of accuracy on request, and allow the claim to be treated as inaccurate if that evidence is not produced or is judged insufficient.
Isn't a provider's own reported accuracy rate (e.g. "90% alignment with real respondents") sufficient evidence?
Not on its own. A validation figure produced, measured, and published solely by the party selling the testing infrastructure has not been established by anyone without a stake in the result. If challenged, a court or authority assessing adequacy under article 12/section 218A would be assessing the same self-interested claim twice — once as marketing evidence, once as its own substantiation.
Has any regulator ruled specifically on synthetic-respondent claims testing?
Not identified as of this writing. The UK Advertising Standards Authority has ruled on AI-generated advertising content generally (misleading images, deceptive video), and has stated its rules are technology-neutral. No ruling addressing synthetic respondents used specifically for pre-publication claims substantiation has been identified to date.
If an advertiser gathers evidence for its claim only after the claim is challenged, does CAP Code rule 3.7 still treat that as adequate?
No — rule 3.7 requires the advertiser to hold evidence before the claim is made, not assemble it afterward. But the rule alone does not independently fix what evidence the advertiser actually held at that earlier moment; only the advertiser's own account establishes it after the fact. The same anteriority gap SOURCE 0 addresses elsewhere in this series reappears here: a record captured and sealed before publication is what distinguishes evidence genuinely held in advance from evidence reconstructed to look that way once challenged.
Does this apply only to Toluna, or to the wider synthetic-respondent market?
The structural point applies to any provider of synthetic testing infrastructure whose validation methodology and accuracy claims are self-reported and not independently reproducible — this is a market-wide characteristic, not specific to any single vendor
