Do EU AI Act labels lower purchase intent?

What the experiment shows

Prof. Dr. Fabian Lang of Hochschule Hannover and Lukas Ting showed 163 people product images of coffee, chocolate, a jacket and a backpack, some conventional photos, some made with AI. Half of them carried the German notice “KI-generiert” (AI-generated), regardless of how they had been made. This separates the effect of the AI image from the effect of the label. The researchers measured product ratings and purchase intent.

According to the report by IT-Recht Kanzlei:

• Whether an image was actually made with AI had no noticeable negative effect on product ratings. On purchase intent, AI images even did slightly better.
• As soon as the “AI-generated” notice appeared, ratings and purchase intent dropped.
• This also applied to images that were in fact conventional photos.
• Participants' age, gender and AI literacy made no difference.

The limits

The experiment measured stated purchase intent for four products, not actual purchases. It tested a single wording, “AI-generated”, with no explanation of what the AI had done. Our source gives no effect sizes, and the full study had not been published when the report appeared.

Do not hide the label

For deep fakes, the label is mandatory (Article 50(4) of the EU AI Act). It must be clear and distinguishable and appear at the latest when someone first sees the image (Article 50(5)). Anything easily overlooked under normal conditions, such as a notice only in menus or terms of use, is not enough (guidelines para. 142). The Code of Practice requires the mark to be recognisable immediately, without interaction or sustained attention. A tiny, half-transparent or hover-only label does not meet that. More in label design.

1. Label only what needs a label

Many uses of AI in a store do not turn an image into a deep fake under the guidelines, for example standard editing such as cropping, colour correction, dust removal or cutting out a background (para. 92), and insignificant changes such as a background for clearly aesthetic purposes behind a real, unchanged product (para. 116). For such images, a label is generally not required under the guidelines. Check each image individually, for example with the labelling check, instead of labelling everything. When in doubt, the label remains the safer choice.

2. Say precisely what the AI did

The Code of Practice encourages a second layer that says whether content was generated or modified, and what was modified. According to Annex 1 of the Code, variants with a clear text label such as “modified” performed significantly better in user tests for noticeability and clarity. An example for a modified photo:

This photo was modified with artificial intelligence (AI). Modified: the background. Unchanged: the product.

Only write “Unchanged: the product” if that is true. For a photo modified with AI, the visible text is “AI-modified”, not “AI-generated”. The disclosure text generator creates matching wording.

3. Real photos where the product matters

The main image shows what the customer will receive and should be a real photo of the product. AI scenes work for mood and setting. If an AI image shows the product nicer, bigger or of higher quality than in reality, a label does not help, because it does not make a misleading image lawful (recital 137 AI Act, guidelines para. 129). Under German competition law, for example, images count as statements about the goods (Section 5(4) UWG).

4. Do not over-label

In the experiment, the label cost trust on real photos too. Legally, “AI-generated” on a real photo is an untrue statement about how the image was made. Whether it is misleading under competition law has not been decided.

• A blanket sentence such as “We use AI” for the whole store does not replace the label on the image and casts doubt on real photos.
• Do not put a label “just in case” on images made without AI.
• If you disclose standard editing voluntarily, write “Edited with AI”, not “AI-generated”.

Under the Code of Practice, labelling is “not a check-box exercise”.

5. Test, but with a label

In your own store, only test variants that all carry a label, such as different text labels or a second layer with and without “Unchanged”. A variant without a label is not a permissible test for a deep fake. A page explaining how you label can also build trust (see AI disclosure statement).