New AI detection tool measures how “surprising” word choices are (www.freethink.com)
TLDR; Summary
A team of researchers at the University of Maryland has developed an AI detection tool called Binoculars that can accurately identify over 90% of writing samples generated by AI.
Binoculars uses two "lenses" to analyze text: an "Observer" LLM that measures the unpredictability of the text, and a "Performer" LLM that predicts the next word based on the previous words.
If there is little difference in perplexity scores between the two LLMs, Binoculars predicts that the text was likely written by AI.
The tool has a false-positive rate of just 0.01%, meaning it incorrectly flags one out of every 10,000 samples as AI-generated.
However, schools may be hesitant to use the tool due to the risk of any false positives, even though the risk is lower than with other AI detection tools.
The researchers believe Binoculars could be valuable for detecting AI-written content on websites and social media platforms.