Computing & AI · the experiment
Unmask the judge
A Turing test turned around, humans and GPT-5 in the interrogator's seat, clues that do not go away. Four questions, from beginner to expert, to find out who is really failing.
1 · Read a negative score
Beginner
Asked to find the human answer, GPT-4 scores d′ = −1.11. What does that mean?
2 · Find the level of chance
The statistical model guesses the source of each answer from 12 features. Its balanced accuracy averages its success over both answer orders.
Intermediate
If GPT-5's answers carried no clue at all, what accuracy would it reach?
3 · Name the humans' failure
In the "find the AI" task, humans slip from d′ = 1.36 against GPT-4 to 0.77 against GPT-5. The statistical model still identifies the source 92.5% of the time.
Expert
Which failure is it, for the humans?
4 · Take the judge at its word
GPT-5 writes that an analytical tone, high readability and plenty of detail give the AI away. Its choices, though, link these three clues to human answers.
Expert
Which level fails here?
Judge unmasked
—