Nicky Molina

Squint

Checks whether writing was likely AI-generated, and shows you why, word by word.

Role
Builder
With
Solo
When
2026 Age 17
Status
Private. Friends get in with a password.
Visit site squint.nickymolina.com (opens in a new tab)

What I did

  • Started with Binoculars, a method that compares how two language models read the same text, running on my own Mac mini.
  • Tested it on 1,638 texts it had never seen. It caught 0 of 288 ChatGPT texts, so I moved the verdict to a trained classifier.
  • Fine-tuned ModernBERT on human, ChatGPT and Claude writing, and kept Binoculars only for the word-by-word heat map.
  • Made Squint answer "Too short to tell" under 50 words, because short human texts were wrongly flagged 2.5% of the time.
  • Squint stores nothing people paste in.

Results

  • Wrongly flagged 4 of 1,340 human texts, and 1 of 512 essays by non-native English writers
  • Caught 744 of 747 ChatGPT texts (99.6%)
  • Caught 674 of 731 Claude texts (92%)

As of October 18, 2026, on texts it never trained on

What was hard

My first method failed. Binoculars couldn't tell 2026 AI writing from human writing, so a week in I had to replace the core of the app.

It still has gaps. It misses older models like 2022 ChatGPT, and it called Claude human 6 of 20 times when Claude imitated an English learner. I also chose the 50-word minimum after seeing results, which flatters the numbers. A fresh test on new sources helps, but its AI side is only 195 texts.

All work