Rhein-Main Universitäten (RMU)

Providers of systems that were already on the market before August 2, 2026, have a transitional period until December 2 to comply with the new requirements. Since Anthropic announced a new watermark for AI-generated text (see box), at the latest, the industry has once again been abuzz with efforts to find new technical solutions. Meanwhile, a study raises a more fundamental question: Could making the use of generative AI visible actually weaken the very human creative contribution that greater transparency is intended to make more apparent?

By Me, Myself and Machine?

Previous research has focused primarily on how mandatory AI disclosures affect audience evaluations. A research team led by Professor Ekaterina Jussupow of TU Darmstadt, with the participation of Professor Kevin Bauer of Goethe University Frankfurt, shifts the attention to the creators themselves: How does creative collaboration between humans and machines change when people know from the outset that their use of generative AI will be disclosed? The study “The Indirect Disclosure Effect: How Disclosing Generative AI Use Impacts Human Creative Collaboration with AI” thus approaches the issue from a different perspective.

Drawing on Erving Goffman’s concept of impression management, the researchers suggest that when disclosure is expected, creators may worry that audiences will underestimate their own contribution to the final result It is therefore not only the actual label that changes how the finished product is perceived by its audience. The expectation of that label can already influence, on the production side, how much human creativity goes into the process at all. The authors refer to this mechanism as the “indirect disclosure effect”.

For their study, the researchers conducted two related mixed-methods experiments. Participants used a text-to-image generator to create images under different disclosure conditions. Central to the experiments was whether participants expected the people evaluating their work to be told about their use of generative AI. The researchers also examined whether it made a difference if participants expected their work to be judged by laypeople or by experts.

The results reveal a striking effect: when creators expected a lay audience to be told about their use of AI, they withdrew more from the creative process and gave the generative AI greater latitude. Their prompts contained 31.9 percent fewer words, while their self-reported effort fell by 13.9 percent. The researchers identify a key reason for this behavior: participants were concerned that audiences would fail to recognize their own creative contribution to the final result.

The researchers argue that the effect is not simply a consequence of the disclosure itself. What matters is creators’ expectation that audiences may fail to recognize or give due credit to their own creative contribution. Interestingly, the effect largely disappeared when participants expected their work to be judged by an expert audience: with an expert jury, the observed effects were small and no longer statistically significant.

The change in the creative process ultimately affects the end result as well. Images created in anticipation of disclosure were rated as more novel, but also as less visually appealing. Disclosure therefore does not only influence how audiences evaluate AI-assisted work – an effect examined in other studies. It can already change the kind of work that gets created in the first place.

This is precisely the tension highlighted by the study: transparency rules are intended to make human authorship and accountability more traceable. Yet the label “AI-generated” can encompass very different forms of human-AI collaboration – from an almost entirely automated image to a work in which a person played a substantial role in both its conception and execution, but which nevertheless receives the designation “AI-generated” as a result of the legal framework and the way its provisions are interpreted.

The authors therefore warn of a paradox: if the use of generative AI is disclosed through a simple label alone, regulation could inadvertently weaken the very human creative contribution it is intended to make visible and protect – before the audience even sees the label.

With this study, TU Darmstadt and Goethe University Frankfurt – both part of the Rhine-Main Universities (RMU) – are pooling their research strengths in the RMU’s ‘Data’ profile area. The RMU Alliance, which comprises TU Darmstadt, Goethe University Frankfurt and Johannes Gutenberg University Mainz, is applying for the status of an ‘University Consortia of Excellence’ under the federal and state governments’ Excellence Strategy.

Background

The EU AI Transparency Act: How Anthropic and OpenAI Are Responding

On August 10, 2026, eight days after the EU Code of Practice on Transparency of AI-Generated Content took effect, Anthropic, the US company behind the AI assistant Claude, unveiled its technical response to the new legal framework: Text generated by Claude receives, for models that already support the technology (a transitional period applies until early December 2026 – see box), an invisible watermark embedded directly in the text. For file formats such as .svg, .png and .jpg, the company additionally uses digitally signed provenance metadata embedded in the properties of the file. This approach is not new.

OpenAI, the operator of ChatGPT, published details of its multi-layered provenance architecture on June 11, 2026. It includes C2PA metadata – so-called Content Credentials that attach machine-readable provenance information to digital content – SynthID watermarks for images, which embed a human-invisible marker directly into AI-generated content and are more likely than metadata alone to survive cropping, filtering or compression, as well as a publicly accessible verification tool for specifically checking for C2PA metadata and OpenAI's SynthID watermark. OpenAI has used C2PA metadata for images generated with DALL·E 3 since 2024. In this process, the text-to-image model generates the content, while C2PA and newer provenance and watermarking systems are intended to make its AI origin technically traceable.

Anthropic has so far released only limited technical details about its text watermark. The company describes it as a barely perceptible pattern created directly during text generation. Put simply, the statistical signature can be imagined as a trail of digital breadcrumbs: invisible to humans, but detectable by machines. Because the marker is part of the text itself, it travels with the content when it is copied and pasted and can survive lighter edits. A verification tool such as OpenAI's can then allow the machine to make those digitally deposited breadcrumbs visible to humans again.

The approach does, however, have limitations – at what point does a particular verification system decide that content was “AI-generated”? Developers point out that a human-written text may still carry traces of a watermark even if Claude was used only for editing, formatting or translation. Conversely, a watermark becomes harder to detect when a text is heavily edited or rewritten, combined with other passages or substantially shortened.

The EU Code of Practice on Transparency of AI-Generated Content

The transparency obligations under Article 50 of the European Union’s AI Act have been in force since August 2, 2026. The Code of Practice on the Transparency of AI-Generated Content is neither a separate EU directive nor the legal basis for these disclosure requirements. Rather, it is a voluntary instrument designed to help providers and deployers of generative AI systems meet the binding requirements of the AI Act and demonstrate compliance.

The European Commission and the European AI Board have recognized the Code as an appropriate voluntary tool for meeting these transparency obligations.

At the heart of the rules is Article 50(2) of the AI Act. It requires providers of AI systems that generate synthetic audio, image, video or text content to ensure that their outputs are marked in a machine-readable format and can be identified as artificially generated or manipulated. As far as technically feasible, the solutions used must be effective, interoperable, robust and reliable.

The requirement does not apply, among other things, where an AI system merely performs an assistive function for standard editing or does not substantially alter the input data provided by the user or its meaning. This distinction reflects the fact that using AI for certain editorial support tasks is not the same as having AI generate content in its own right.

For certain providers of systems that were already on the market before August 2, 2026, and generate synthetic content, a transitional period runs until December 2, 2026, to comply with the requirements of Article 50(2).

Text: Leonie Schultens (Goethe University Frankfurt)