OPEN LINK ↗
// 1h agoRESEARCH PAPER
Graphite Maps 13,000 Shifting AI Tells
Graphite analyzed 10,000 human articles and 90,000 AI-generated articles across nine models, identifying nearly 13,000 words, phrases, and stylistic patterns associated with AI writing. The study finds that model-specific tells change between releases, while AI writing often resembles other AI writing more than human prose. [Research](https://graphite.io/five-percent/research/ai-tells)
// ANALYSIS
The study makes generic AI-writing checklists look increasingly brittle: detection needs to account for model family, version, topic, and prompting conditions.
- –65% of identified tells are unique to one model family, so “AI vocabulary” is not one universal fingerprint.
- –Between consecutive model versions, 55%–72% of tells are version-specific, weakening rules based on words like “delve” or punctuation like the em dash.
- –Every tested model’s word distribution is closer to at least one other model than to human writing, suggesting an emerging machine-native writing style.
- –The matched-topic methodology improves comparisons, but identical prompting and synthetic generation may still influence the results.
- –For developers, human editing should focus on restoring voice and specificity—not simply deleting a blacklist of “AI-sounding” words.
// TAGS
graphite-ai-tellsllmevaluationresearchdataset
DISCOVERED
1h ago
2026-09-30
PUBLISHED
1h ago
2026-09-30
RELEVANCE
8/ 10
AUTHOR
Better Stack