Ellia Lebleu
AI researcher and doctoral student in AI at the Inria Rennes research center, and a founding member of Wikipedia's WikiProject AI Cleanup, researching AI-generated content detection.
Timeline (last 90 days)
Ellia Lebleu is one of the founding members of WikiProject AI Cleanup and is currently pursuing an AI PhD at the Inria Rennes research center near Toulouse.
Lebleu collaborated with the AI detection company Pangram to quantify the extent to which AI content has infiltrated Wikipedia.
According to Lebleu's estimate, about 20% of new articles and drafts were flagged as AI-generated at the end of 2025.
According to Lebleu's estimate, by early 2026 the proportion of new articles and drafts flagged as AI-generated dropped to 15% to 16%.
Lebleu said it is currently impossible to determine whether the decline in the proportion of AI-generated content is because community policies worked or because models got stronger and detection tools failed to keep up.
The characteristics of AI-generated content in late 2023 included concluding paragraphs, formulaic rhetoric, and a mechanical three-part structure; by 2025, the signals became more structural and more subtle, such as dense stacking of triple parallel structures and negative parallel structures.
AI contamination does not only occur in the body text; false or misleading citations have become a major concern.
Lebleu offers a phased judgment on the relationship between Wikipedia and AI models: yesterday AI was the object needing cleanup, today it is a tool assisting cleanup, and tomorrow it will be the air we breathe.
All Facts (8)
Ellia Lebleu is one of the founding members of WikiProject AI Cleanup and is currently pursuing an AI PhD at the Inria Rennes research center near Toulouse.
50%UnverifiedLebleu collaborated with the AI detection company Pangram to quantify the extent to which AI content has infiltrated Wikipedia.
50%UnverifiedAccording to Lebleu's estimate, about 20% of new articles and drafts were flagged as AI-generated at the end of 2025.
50%UnverifiedAccording to Lebleu's estimate, by early 2026 the proportion of new articles and drafts flagged as AI-generated dropped to 15% to 16%.
50%UnverifiedLebleu said it is currently impossible to determine whether the decline in the proportion of AI-generated content is because community policies worked or because models got stronger and detection tools failed to keep up.
50%UnverifiedThe characteristics of AI-generated content in late 2023 included concluding paragraphs, formulaic rhetoric, and a mechanical three-part structure; by 2025, the signals became more structural and more subtle, such as dense stacking of triple parallel structures and negative parallel structures.
50%UnverifiedAI contamination does not only occur in the body text; false or misleading citations have become a major concern.
50%UnverifiedLebleu offers a phased judgment on the relationship between Wikipedia and AI models: yesterday AI was the object needing cleanup, today it is a tool assisting cleanup, and tomorrow it will be the air we breathe.
50%