The AI labelling paradox: Why transparency mandates could backfire for publishers
- Applying labels is not only a question of ethics and trust but a daunting task without an instruction manual.
- Labelling content as AI-enhanced or AI-generated reduces engagement, and the drop is steepest for emotional content.
- Labels also create an "implied authenticity" effect: users trust unlabelled content more, even when it is false.
- The EU's AI Act makes labelling unavoidable, yet some studies find disclosure does not make AI text less persuasive.
Generative AI is now part of routine newsroom work, and regulators want audiences told about it. The European Union's AI Act mandates disclosure for synthetic media, and platforms from Meta to TikTok are rolling out new labels.
The stated goal is transparency: clear signals about how content is made. The research tells a more awkward story. Labels meant to build trust can depress engagement, and they may leave audiences more exposed to unlabelled misinformation, not less.
Another impossible task for the individual journalist
Newsrooms experimenting with generative AI keep running into the same question: where does "AI-assisted" end and "AI-generated" begin? The answer shapes both editorial credibility and the economics underneath it. How an outlet draws that line, and how it explains the line to readers, says a lot about what it thinks its audience is paying for.
The scope of AI use is wide and varied, and drawing a line in the sand is virtually impossible. For instance, if a journalist is researching an article through a search engine like Google, but through the process relies on Google’s “AI overview” should the resulting information be labeled as “AI-assisted” or “AI-generated”, or should it have a label at all?
On the opposite end of the spectrum the same dilemma applies. Let’s suppose that a journalist asks an LLM like ChatGPT to write an entire article for them, while they themselves apply a few changes or add a paragraph. What label should that article have?
Process versus impact: two kinds of label
The first question for any labelling policy is what the label is supposed to say. Researchers at MIT distinguish between process-based labels, such as "AI-generated", which describe how content was made without judging its accuracy, and impact-based labels, such as "Manipulated" or "False", which do make that judgment. Experiments show the two behave very differently.
In the MIT work, an "AI-generated" label had only a modest effect on what users believed. "Manipulated" and "False" labels were far better at reducing belief in misleading images and the willingness to share them.
Blanket process labels carry a cost, though. Applied to everything AI touched, including harmless digital art and accurate summaries, they made people less confident in the content whether it was true or false. Publishers therefore face a choice. Label everything and accept that suspicion falls on accurate work too, or label only harmful content and build a fact-checking operation big enough to defend those judgments. Most organisations do not have one.
The engagement penalty of AI disclosure
A study published in Electronic Markets, which simulated user interaction on Instagram, found that engagement fell as the stated level of AI involvement rose from "human-created" to "AI-enhanced" to "AI-generated".
The drop was not a rational rejection of machine-made content. It ran through what the authors call affective engagement: knowing AI was involved made people feel less connected to the material, which made them less likely to like, comment or share.
The size of the penalty depended on the content. Emotional posts suffered a much larger drop than rational, information-based ones. Human-created emotional appeals generated far higher engagement, and as AI's role increased that advantage disappeared.
This is uncomfortable for publishers, because so much journalism depends on emotional resonance, in narrative and in visuals. Research from the Nuremberg Institute for Market Decisions found the same pattern in advertising: consumers rated identical ad content more negatively, particularly on emotional impact, once it was labelled as AI-made. A newsroom that adopts AI for speed may be trading away the connection that keeps readers loyal.
The unintended consequences of transparency
Labels also change how people read everything else. A study from the CISPA Helmholtz Center for Information Security in Germany found that labels helped users spot flagged AI content but encouraged them to lean on the labels themselves.
Participants were more likely to believe unlabelled content, even when it was false. Because some content carried labels, they assumed anything without one had been vetted. The authors call this "implied authenticity". Instead of becoming more discerning, readers outsourced their judgment to the label and got worse at telling true from false.
For some formats, labels may not do their main job at all. A Stanford HAI study found that labelling persuasive policy arguments as AI-generated did not make them less persuasive. Participants correctly identified the author as AI, but their support for the policy, their assessment of its accuracy and their willingness to share it did not change. For text, the format most journalism works in, a label may deliver transparency without blunting AI-driven influence campaigns.
There is also a familiar risk from advertising: banner blindness. If AI warnings appear everywhere, audiences may stop seeing them.
Navigating the new regulatory reality
None of this changes where regulation is heading. The EU's AI Act requires platforms to label deepfakes and other synthetic media, and it will shape rules elsewhere. In the United States, federal bills such as the proposed REAL Act have moved slowly, but the direction is the same. Meta, YouTube and TikTok already require creators to disclose generative AI in realistic content, so for most publishers compliance is coming regardless of legislation.
That moves the strategic conversation from whether to label to how. The research argues against a generic "AI-generated" tag on everything: it depresses engagement and creates false confidence in whatever goes unlabelled. Labels probably need to carry more context, and they need to sit alongside media literacy work that helps audiences judge content on its merits, whatever its origin.
How would we help your newsroom?
30-min call can show you what is possible and trail with the tool can give you the insight in the benefits for you.
Book a demo and reclaim hours→
For media leaders, the job is to prepare for mandatory disclosure while designing around the documented failure modes. A warning label on its own does not build trust. Explaining how humans and machines work together in the newsroom might.
This article was drafted using AI, namely NewsLabs, on whose website you are reading these lines. It was verified and edited by a human editor.
