Can an “AI-generated” label protect the truth?
Reading the EU AI Act through three philosophers
Will the EU’s new rules on labelling AI-generated content put an end to misinformation? The ideas of three philosophers help us see what a label can—and cannot—do.
Imagine finding a video on social media in which a politician appears to make an important statement. A small label on the video says “AI-generated.”
The label tells you that AI was used to create the video. But does it also tell you whether the message is false?
The answer is no.
Information produced by AI can be correct. Information produced by a person can be wrong. An “AI-generated” label tells us how something was made, not whether its content is true.
Why, then, does the EU require disclosures for certain kinds of AI-generated content? We can approach this question through the ideas of Immanuel Kant, Hannah Arendt and Bernard Williams.
What is changing in the EU?
On August 2, 2026, the transparency obligations in Article 50 of the EU AI Act begin to apply.
The rules involve two main kinds of disclosure.
1. Marks that computers can read
Providers of AI systems that generate text, images, audio or video must make those outputs detectable as artificially generated or manipulated. This can involve information embedded in a file, similar to a digital watermark.
2. Disclosures that people can understand
People and organisations using AI professionally must clearly disclose certain deepfakes. They must also disclose AI-generated or manipulated text published to inform the public about matters of public interest.
This does not mean that every use of AI needs a prominent “AI-generated” label.
Text on a matter of public interest is exempt from the disclosure requirement when it has undergone meaningful human review or editorial control and a person or organisation takes editorial responsibility for its publication. The EU is therefore asking more than whether AI was used at all. It is also asking who checked the content and who accepts final responsibility for it.
A label shows how content was made—not whether it is true
An “AI-generated” label is not a warning that says, “This is false.” Nor is it a guarantee that says, “This is correct.”
For example, an AI system may accurately summarise reliable weather data. Meanwhile, a human writer may publish an article based only on an incorrect assumption.
An AI label tells us about provenance: where information came from and how it was produced. Why does knowing that provenance matter?
Kant—A label gives people the conditions to judge for themselves
Immanuel Kant was an eighteenth-century German philosopher. He regarded truthfulness as a basic duty that helps make social life possible.
“Truthfulness in statements which cannot be avoided is the formal duty of an individual to everyone.”
— Immanuel Kant, “On a Supposed Right to Tell Lies from Benevolent Motives”
For Kant, a lie does more than deceive one person. If lying becomes normal, people lose trust in what others say. Promises, agreements and other social relationships then become harder to maintain.
Kant did not write about generative AI. We can, however, apply this line of thought to the present.
If an artificial video is presented as authentic footage, the viewer is denied information needed to judge it fairly. Hiding the use of AI removes part of the basis on which the viewer makes a decision.
On this view, an AI label should not be a mark declaring that an AI-made work is inferior to a human-made one. It should give people information that respects their ability to think and decide for themselves.
Yet disclosing the method of production does not make the content true. Truthfulness also requires showing the evidence behind a claim.
Arendt—Free opinion depends on shared facts
Hannah Arendt was a twentieth-century political thinker. She stressed the difference between facts and opinions in a democracy.
“Freedom of opinion is a farce unless factual information is guaranteed and the facts themselves are not in dispute.”
— Hannah Arendt, “Truth and Politics”
A simple example makes this distinction clearer.
Whether a policy is good or bad is a matter of opinion. People may reasonably disagree. Whether the policy was actually adopted, or whether a politician really made a particular statement, is a matter of fact.
Opinions can differ, but discussion requires at least some shared facts. If events that never happened enter society’s “common reality,” the foundation for public debate begins to disappear.
The danger of a deepfake is therefore not limited to the harm it may cause one person. A deepfake can place a fictional event inside the world that people treat as real.
The EU’s focus on deepfakes and texts about matters of public interest can be read as an effort to protect the factual ground on which people form opinions.
Labels are still not enough. Some accurately labelled AI content will be true, while some unlabelled human content will be false. If people begin to treat everything without a label as authentic, a transparency rule may create a new mistaken assumption.
Williams—Accuracy and sincerity are not the same
The British philosopher Bernard Williams described two basic virtues connected with truth in his book Truth and Truthfulness.
Accuracy
The effort to examine evidence, resist wishful thinking and discover what is actually true.
Sincerity
The effort to communicate honestly what one believes and how one produced the information.
This distinction makes the role of an AI label easier to see. Saying that AI was used supports sincerity because the method of production is not hidden. The label itself, however, does not examine evidence or verify a claim. It cannot guarantee accuracy.
Generative AI also does not hold beliefs or accept legal and editorial responsibility in the way a person or organisation can. We cannot make the AI itself the final responsible speaker.
The people and organisations that publish information must therefore check sources, correct mistakes and take responsibility for what they release.
From this perspective, there is a clear reason for the EU’s exception based on meaningful human review, editorial control and editorial responsibility.
Three things we need to protect the truth
Disclosing AI-generated content is important. But writing “AI-generated” on something cannot protect the truth by itself.
- Transparency about provenance — Explain where and to what extent AI was used.
- Verification — Provide sources and evidence, and check whether the claim is true.
- Clear responsibility — Identify who reviewed the content and who will correct it if it is wrong.
An AI label is not a final verdict on information. It is neither a warning that says, “Do not believe this,” nor a guarantee that says, “This is correct.” It is a starting point for checking.
The success of the EU’s rules will depend on more than the number of labels. What matters is whether people can follow a label to the evidence behind a claim and to the person or organisation responsible for it.
NOW IN QUESTION
Are we asking labels to make the judgment for us?
Do we truly want labels that prevent AI from deceiving us? Or are we asking a label to replace the work of judging truth and falsehood that we must ultimately do ourselves?
References
- Regulation (EU) 2024/1689, Article 50
- European Commission, Guidelines on transparency obligations for providers and deployers of certain AI systems
- European Commission, Transparency obligations under Article 50 of the AI Act
- Immanuel Kant, On a Supposed Right to Tell Lies from Benevolent Motives
- Hannah Arendt, “Truth and Politics”
- Bernard Williams, Truth and Truthfulness: An Essay in Genealogy