AI-generated content in Europe 2026: adoption and detectability

Key findings

• 32.7% of people aged 16 to 74 in the EU used generative AI tools in the three months before the 2025 survey, and 63.8% of those aged 16 to 24. National rates range from 17.8% (Romania) to 56.3% (Norway).
• The share of EU enterprises using AI to generate text, speech or code rose from 1.3% in 2021 to 8.8% in 2025. In 2025, the first year it was measured, 9.6% used AI to generate pictures, video or audio.
• In France, 32.8% of people aged 12 and over have used AI to create images, photos or video, and only 21.4% of AI users always check its output before using it (Baromètre du numérique 2025).
• Detection of original files is reliable: 95.6% of raw AI text, 99.5% of AI images and 95.2% of AI video clips, each at a low false-positive rate.
• Detection weakens with distribution. On images it holds (96.2% after WhatsApp and Instagram compression). On AI video of crisis events it falls from 86.9% to 55.9% after a single re-encode.
• Real footage is rarely misclassified: 1.0% of 1,319 real crisis clips were flagged as AI.

1. Introduction

The EU AI Act puts the duty to mark synthetic audio, images, video and text on the providers of generative models (Article 50(2)), and the duty to disclose deepfakes on those who publish them (Article 50(4)). Both obligations have applied since 2 August 2026. Marking works when the marked file reaches its audience unchanged. In practice, content is screenshotted, re-encoded, cropped and reposted, and machine-readable marks are often lost along the way.

This report puts two numbers side by side for each medium: how widely generative AI is used, from official statistics, and how well its output can be recognised after the fact, from detection benchmarks. The aim is to give policymakers, platforms and newsrooms a sober picture of where detection can carry weight and where it cannot.

2. Data and methods

2.1 Use of generative AI

Individuals. Eurostat table isoc_ai_iaiu, from the Community survey on ICT usage in households and by individuals, 2025. It reports the share of people aged 16 to 74 who used generative AI tools in the previous three months, by purpose and age group. 2025 is the first year these questions were asked, so there is no earlier time series.

Enterprises. Eurostat table isoc_eb_ai, enterprises with ten or more employees, all activities except agriculture, mining and the financial sector, 2021 to 2025. We use three indicators: use of at least one AI technology, use of AI to generate written or spoken language or code, and use of AI to generate pictures, video or audio (first collected in 2025).

France. Baromètre du numérique 2025 (Arcep, Arcom, CGE, ANCT), published on data.gouv.fr: 4,145 people aged 12 and over, interviewed online and by phone. We computed weighted shares from the individual records using the survey weight, for three questions: how often people use AI to write text, how often to create images, photos or video, and whether they check AI output before using it.

Luxembourg. STATEC table AI technologies in use in enterprises, published on data.public.lu, by economic activity, 2021 to 2024.

2.2 Detectability

We report results from three evaluations of the WasItAIGenerated detectors, each run on data the models were not trained on:

• Text (September 2026), detector tropa-3: 2,572 texts including output from current language models and from six humanizer tools, scored by five detectors. Results here are for the 513 texts all detectors scored.
• Images and video (October 2026): 469 images from 19 current image generators and 164 clips from 12 video models, against freely licensed real photos and clips, each also as WhatsApp, Instagram and re-shared copies. Results are from the 40% hidden split.
• Video of crisis events (November 2026): 17,974 clips from RA-Bench, a public benchmark of AI videos generated from 1,830 real crisis clips by nine generators, including a re-encoding experiment and 633 fakes that fooled all five human reviewers.

Detection rates are the share of AI files flagged at a fixed false-positive rate on real files, or at the production threshold where stated. The full methodology of each benchmark is published with it.

3. Results

3.1 One in three Europeans uses generative AI

Eurostat isoc_ai_iaiu, 2025, last three months. All countries are in the CSV.

3.2 Companies now generate more media than text

Eurostat isoc_eb_ai, enterprises with 10 or more employees.

In 2025, slightly more EU enterprises used AI to generate pictures, video or audio (9.6%) than to generate text (8.8%). In Germany the difference is larger: 13.5% against 9.0%. Finland (21.6%), Sweden (18.7%) and Denmark (17.7%) lead on media generation. For the detection question, this means synthetic media is no longer mainly a consumer phenomenon: a growing share of it comes out of ordinary business workflows such as marketing.

3.3 France: a third create images or video, few check the output

Baromètre du numérique 2025, n = 4,145, weighted; own calculation from the open records.

The French survey also asked AI users whether they check what the tools produce before using it. 21.4% always do and 43.0% often do; 29.1% rarely check and 6.6% never do. More than a third of users therefore pass AI output on largely unchecked, which is where synthetic text and images enter circulation without anyone having looked at them critically. The French figure for any use (48.3%) is higher than the Eurostat figure for France (37.5%) because the questions differ: Eurostat asks about the last three months and ages 16 to 74.

3.4 Luxembourg: AI use depends strongly on the sector

STATEC, enterprises with 10 or more employees, non-financial business economy.

In information and communication, almost half of Luxembourg's enterprises used AI in 2024 and 28.5% generated text or speech with it, against 3.4% in accommodation and food services. Synthetic content therefore comes disproportionately from the sectors that produce most published content: media, software, consulting and marketing services.

3.5 Detectable at the source, much less after sharing

For text and images, detection holds up well, including after the compression applied by messengers. Text rewritten by humanizer tools is the clear weak point (57.7%). Video of crisis events is the hardest case: one re-encode lowers detection from 86.9% to 55.9%, and with downscaling added it falls to 26%. For comparison, the seven research detectors evaluated in the RA-Bench paper averaged an AUC between 43.9 and 57.3 per generator on the original clips, close to chance; our detector reaches 92.9 to 100 on the same pairing.

4. Discussion

Marking and detection need each other. Some generators already sign their output: every Seedance 2.0 clip we inspected in RA-Bench carried a valid C2PA manifest declaring it AI-generated. Such credentials prove origin as long as the file is unchanged. Most reposting pipelines remove them, and in that same step, re-encoding also makes pixel-based detection harder. Neither approach covers the redistributed copy on its own.

Keep originals. For fact-checkers, platforms and investigators, the most effective step is to obtain the earliest available copy of a file. Detection on original clips (86.9%) is far higher than on the transformed copies tested (26% to 68%), and credentials, where present, survive only there.

Report uncertainty instead of a yes or no. Scores close to the threshold should be presented as "needs review", not as human-made. On RA-Bench, a review tier between scores of 0.5 and 0.8 raises the share of AI clips that receive at least a warning from 86.9% to 94.2%, while firm false positives stay at 1.0%.

Measure independently. Detection quality differs widely between tools and decays as new generators appear. Public benchmarks with released raw scores, such as RA-Bench and the sets used here, are the basis for checking vendor claims, including ours.

The detectors evaluated here are available as a free deepfake and AI video detector, an AI text and image checker and the deepfake detection API. Broader figures on synthetic media are collected in our deepfake statistics 2026.

5. Limitations and conflict of interest

• The author's company develops the detectors evaluated here. Two of the three benchmarks were built by us; RA-Bench was built independently and was never used for training.
• Survey figures are self-reported and cover people aged 16 to 74 and enterprises with ten or more employees outside the financial sector.
• The report sets adoption and detectability side by side. It does not measure how much of the content in circulation is AI-generated.
• The RA-Bench evaluation covers 17,974 of the 25,584 publicly downloadable files, chosen at random and balanced across generators. The detection figures for text and images come from smaller samples; see the sample sizes above.

Data availability

All tables in this report, with every country from the Eurostat tables and every Luxembourg sector, are available as a CSV file under CC BY 4.0. Eurostat data © European Union, reused under the Eurostat copyright and licence policy; STATEC data CC0. The French figures are derived from the Baromètre du numérique, which is published under the Open Database License, and are therefore released separately as a CSV file under ODbL 1.0. The detection benchmarks are documented on the WasItAIGenerated research page.

How to cite

Gerard, P. (2026). AI-generated content in Europe 2026: adoption and detectability. WasItAIGenerated data report, version 1.0. https://doi.org/10.5281/zenodo.24145956

Archived copies: Zenodo · data.gouv.fr · data.public.lu

References

1. Eurostat (2026). Individuals – use of generative AI tools (isoc_ai_iaiu). Updated 5 June 2026.
2. Eurostat (2026). Artificial intelligence by size class of enterprise (isoc_eb_ai). Updated 15 June 2026.
3. Arcep, Arcom, CGE, ANCT (2026). Baromètre du numérique, survey 2025 (open data, ODbL). data.gouv.fr, updated 9 February 2026.
4. STATEC (2025). AI technologies in use in enterprises. data.public.lu, CC0.
5. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50.
6. RA-Bench authors (2026). Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination. arXiv:2608.14391.
7. WasItAIGenerated (2026). Pangram vs GPTZero vs ZeroGPT vs tropa-3: AI detector benchmark, September 2026.
8. WasItAIGenerated (2026). AI image & video detector benchmark, October 2026.
9. Coalition for Content Provenance and Authenticity. C2PA technical specification.