Research
Open, applied computational linguistics from the HumanizeMyAI lab: rigorous, transparent and useful. We study how language reveals the line between human and machine writing, and we put the same tools to work on real-world questions, in partnership with the organisations that need the answers.
Covered by TechRadar Pro, DigitalToday.
The Fingerprint Is Leaking Into the Record
The AI vocabulary (delve, intricate, underscore, showcase) didn't stay in AI text. We tracked it across 18,989 arXiv abstracts over eight years and watched these words leak into human writing: up about 7× in computational linguistics after ChatGPT, but essentially flat in mathematics, where there's no topical reason to use them. That math control is the proof it's stylistic diffusion, and it means the human baseline lexical detectors rely on is now contaminated.
Every Model Has an Accent
We gave five AI models the same 102 prompts and measured only their style. The vendor is almost perfectly separable (OpenAI versus Anthropic at 96% accuracy), while the exact model lands at just 50%, with errors trapped inside each family. And the famous 'em-dash = AI' tell is really an 'em-dash = Claude' tell: Claude uses roughly five times more em-dashes than GPT.
What Actually Changes as English Proficiency Grows
Across 3,300 learner essays graded from beginner to near-native, the clearest mark of progress is accuracy: errors fall almost five-fold, from about 21 to 4 per 100 words, converging on native writing. Complexity grows too, but phrasally: richer noun phrases, not piled-up subordinate clauses. And punctuation, stubbornly, becomes the single largest remaining error for advanced writers.
There Is No Aesthetic Surprisal Arc
A great piece of music shapes your surprise into an arc: tension that rises and resolves. We tested whether writing does the same, and whether that arc is what gives AI text away. It isn't. On three public benchmarks, the shape of a text's surprisal trajectory predicts neither human-vs-AI authorship nor writing quality. What actually separates human from machine prose is far more boring, and one half of it survives a paraphrase attack while the other collapses.
Estetik Bir Şaşırtıcılık Yayı Yok
İyi bir müzik parçası şaşkınlığınızı bir yaya dönüştürür: yükselip çözülen bir gerilim. Yazının da aynısını yapıp yapmadığını ve bu yayın yapay zekâ metnini ele verip vermediğini test ettik. Vermiyor. Üç herkese açık kıyaslama kümesinde, bir metnin şaşırtıcılık eğrisinin biçimi ne insan-karşısında-yapay zekâ yazarlığını ne de yazı kalitesini öngörüyor. İnsan nesrini makine nesrinden gerçekte ayıran şey çok daha sıradan ve bunun bir yarısı bir başka deyişle yazma saldırısını atlatırken diğeri çöküyor.
It Isn't Delve: The Real Lexical Signature of Machine-Generated Text
Everyone knows the AI tells: delve, tapestry, intricate, realm. We measured them against 11 million words of real human and ChatGPT answers, and the famous words barely register. What actually separates machine writing from human writing is duller than any vocabulary list: a hedged, list-building textbook register that leans on important, overall, including and may by three to nine times. And the specific words shift with every model, which is exactly why a word list ages so fast.
What 12,000 Tourists Say About Alanya
We used AI to read 12,222 public reviews of 49 Alanya businesses (in a dozen languages, over nine years) for the Alanya City Council. The surprise: every nationality describes a different Alanya, and the one thing they all find overpriced is the beach.
12.000 Turist Alanya Hakkında Ne Diyor?
Alanya Kent Konseyi için 49 işletmenin 12.222 herkese açık yorumunu (onlarca dilde, dokuz yılda) yapay zekâ ile okuduk. Sürpriz: her milliyet farklı bir Alanya anlatıyor ve hepsinin pahalı bulduğu tek şey plaj.