When the Law Is Right but the System Is Wrong
An “AI” label can tell us that artificial intelligence was used. It cannot tell us who had the idea, how much the human contributed, or whether the final result is true.
The thought behind this essay did not begin with a particularly complicated legal provision. It began with a very simple question: what exactly do we learn when a text, image or video carries a label stating that it was created with artificial intelligence?
We learn something about how it was produced. We do not learn whether it is true.
The distinction appears obvious. Yet it is often lost in public debate. Disclosure of AI use is gradually treated as a warning, a quality assessment and an indirect sign of reduced credibility at the same time—as if the origin of a piece of content were enough to determine how much we should trust it.
This is precisely where a legally coherent framework risks describing the real system it is trying to regulate incorrectly.
Europe’s reasonable response
The European Union does not approach artificial intelligence only through content labelling. The AI Act is much broader. It includes prohibited practices, rules for high-risk systems, obligations for providers and deployers, requirements for general-purpose AI models and a wider framework of transparency and accountability.
It would therefore be unfair and inaccurate to claim that European AI regulation has been reduced to an “AI” label.
Within the narrower field of generated content, however, labelling and technical detectability are indeed the principal regulatory responses. Since 2 August 2026, Article 50 of the AI Act has required, among other things, machine-readable marking for certain synthetic outputs, disclosure when a person interacts directly with an AI system, and visible labelling of deepfakes or public-interest text published without substantive human review or editorial responsibility. The European Commission’s guidance on Article 50 explains the scope and exceptions in detail.
The logic is understandable. When citizens communicate with a system, they should know that they are not speaking to a person. When they see a convincing video of a real person saying something that person never said, they should know that the material was artificially generated or manipulated. Transparency is essential when confronting deception, impersonation and manipulation.
The problem is not that labelling is wrong. The problem begins when we treat it as a sufficient answer.
The law needs categories; reality operates on a continuum
Every law needs definitions. It must determine who is a provider, who is a deployer, when content is considered AI-generated, who is responsible for marking it and which exemptions apply. Without clear boundaries, there can be no implementation, enforcement or assignment of responsibility.
Human–machine collaboration, however, does not fit easily into binary categories.
A text may begin with an entirely human idea. A person may formulate the central position, define the essential parameters and ask a model to propose an initial structure. They may then reject much of the output, rewrite entire sections, add personal examples, verify the sources and assume full responsibility for the finished work.
In another case, someone may type a single sentence, accept the first response without examining it and publish it unchanged.
Both works could carry exactly the same label: “Created with AI.”
The label might be technically correct in both cases. Yet it would describe two fundamentally different creative, cognitive and accountable processes as though they were equivalent.
Who had the idea?
This leads to a question that regulation has not answered convincingly—and may never be able to answer.
Was the idea human or machine-generated?
The easy response is that the person wrote the prompt, so the intention belonged to them. But that does not cover every situation. A system may suggest a connection the person had not considered. The person may reject it, modify it, or recognise within it a different idea and develop that idea in a direction the model never anticipated.
At what precise point was the final idea born?
Can we say that AI contributed 20, 50 or 80 per cent? What unit would measure that contribution? The number of generated words retained in the final text? The original concept? The structure? The difficulty of the research? The decisions that led the author to reject incorrect suggestions?
Creative work is not a spreadsheet. A single sentence can change the reasoning of an entire essay, while hundreds of machine-generated words may contribute nothing of value. Selection, judgement and rejection are also creative acts, even though they leave no visible trace in the final output.
The more iterative the collaboration between a person and a system becomes, the less meaningful an exact allocation of authorship becomes. The process is not a simple transfer of work from one participant to another. It is a cycle of proposal, evaluation, correction, reframing and final choice.
The law can require disclosure that a tool was used. It cannot observe with certainty where a thought was born.
The limits of the label itself
The European approach partly recognises this difficulty. Under the Article 50 guidance, public-interest text that has undergone substantive human review or editorial control does not require the same visible disclosure. Grammar correction alone is not enough. The substance must be examined by a person with relevant knowledge, and a human must hold ultimate editorial responsibility.
That distinction matters. It shows that the regulation itself understands that human participation is not always decorative.
But it remains a coarse threshold: meaningful human review either took place or it did not. It does not explain who conceived the position, who selected the data, who constructed the argument, who detected the mistakes or how the cognitive work was actually distributed.
Most importantly, it does not tell us whether the result is valid.
Disclosure of origin is information about the production process. It is not evidence of truth, quality or reliability.
A completely human-written text may be false, careless, unsupported or deliberately manipulative. A text created with substantial AI assistance may have been rigorously examined, grounded in primary sources and be more accurate than many texts written entirely by people.
The label does not tell us which of the two we are reading.
When origin replaces argument
There is a deeper danger. A label may stop functioning as information and begin functioning as a way to weaken a voice.
This is an old practice with a new target. When we do not want—or are unable—to answer what is being said, we shift the discussion towards the person saying it. We examine their identity, motives, status or affiliation and leave the argument itself untouched.
In its familiar form, this is an attack on the person. In the case of AI-assisted content, it is closer to what logic calls the genetic fallacy: accepting or rejecting a claim because of its origin rather than because we examined the evidence supporting it.
AI use may therefore become an easy route out of substantive debate. Instead of asking, “Is this correct?”, we ask, “Was this written by AI?” Instead of examining the data, we evaluate the genealogy of the text.
We risk replacing the evaluation of a claim with the evaluation of its ancestry.
That is not progress in critical thinking. It is a new way to avoid it.
Origin matters—but not in the same way everywhere
It would be equally mistaken to move to the opposite extreme and claim that origin never matters.
It matters enormously when we are evaluating a deepfake, a piece of evidence, a scientific publication, a medical or legal opinion, an instance of plagiarism or a work for which intellectual-property rights are being claimed. It matters when we need to know whether a photograph records a real event, whether a quotation is authentic and who is responsible for an automated decision.
In those contexts, provenance, chain of custody and auditability are not secondary details. They form part of the evidence.
But knowing the origin is not the same as using origin as a substitute for verification.
The fact that an image is synthetic tells us that it cannot automatically serve as photographic evidence of a real event. The fact that an analytical essay was written with AI assistance does not automatically tell us that its arguments are wrong. The significance of origin depends on the use, the risk and the claim being made.
Regulation therefore cannot be merely horizontal and symbolic. It must be proportionate to actual risk.
From labels to responsibility
If we want a framework that reflects reality, we need more than a general disclosure that AI was used.
First, there must be clear human or organisational responsibility. The essential question is not only which tool was used, but who decided that the output was ready to be published or acted upon. Who signs it? Who had the authority to reject it? Who is accountable if it causes harm?
Second, we need provenance of the data, not only provenance of the text. A small icon declaring “AI” has limited value if we cannot establish which sources support an important claim, whether those sources are authentic and whether the underlying information is still current.
Third, verification procedures must reflect the level of risk. A creative essay, an advertising image, a medical recommendation and a decision affecting someone’s access to employment or credit cannot be governed through the same compliance logic.
Fourth, where consequences are significant, we need audit trails that record the meaningful decisions—not every irrelevant keystroke, but which data entered the process, which system version was used, what the human changed, who approved the outcome and according to which criteria.
Fifth, the label must be presented for what it actually is: provenance metadata, not a reliability score.
When legal correctness is not enough
This issue reveals a broader weakness in the way institutions regulate complex technological systems.
A legal text can be coherent, careful and institutionally sound. It can define obligations, exemptions and penalties with precision. Yet the model of reality on which it rests may still be inadequate.
This would not be the first time that a European regulation with a legitimate purpose risks equating compliance with the desired outcome. In the case of the GDPR, the right to control personal data has often appeared in everyday life as banners, consent declarations and legal notices, while citizens still cannot always know which data organisations actually retain about them. Similarly, under the AI Act, the presence of a label may prove that a transparency obligation was followed; it does not prove that the validity of the content was assessed.
When regulation is designed mainly through a legal lens, it tends to treat technology as an object that can be classified. Modern information systems, however, are dynamic environments. They contain data of uncertain quality, supplier chains, changing models, people who use systems in unexpected ways, commercial incentives and legacy infrastructure that does not become compliant merely because another paragraph has been added to a regulation.
Technology therefore cannot be regulated only by lawyers, just as it cannot be regulated only by engineers. It requires lawyers, engineers, system architects, security specialists, data scientists, economists, operators and professionals who understand what happens when a rule meets real data and a production environment.
The law can be legally correct and operationally inadequate at the same time.
AI labelling is a characteristic example. It creates a clear obligation and a visible act of compliance. It is easy to verify whether a label is present. It is much harder to determine whether the content is accurate, whether its sources are valid and whether the person publishing it exercised genuine judgement.
The danger is that we measure what is easy to measure and assume that we have solved what is difficult.
The question that must change
The next stage of the debate cannot be limited to, “Was this written by a person or by AI?”
That question assumes a clean separation that no longer exists in many modern workflows. Even if we could identify the boundary, it would not tell us what we actually need to know.
The meaningful questions are different:
Can the claim be verified? What data support it? Are the sources available? Was there substantive human review? Who bears final responsibility? Is there a record of the critical decisions? What are the consequences if the output is wrong?
Transparency about AI use remains useful and, in many situations, absolutely necessary. But it should not be confused with truth, nor should it become a socially acceptable way to dismiss an argument without examining it.
The label describes the origin of the content. It does not certify its validity.
Ultimately, the critical question is not whether a machine participated in expressing a thought. It is whether the thought survives scrutiny—and whether a person or an organisation is prepared to stand behind it and accept responsibility for it.