Research study · AI voice detection
You'll listen to short voice clips and decide whether each is a real human or an AI. Partway through, half of participants get a short lesson on what to listen for. Everyone is tested the same way at the end, including on an AI model no one trains on.
About this study
Purpose. You’re being invited to take part in a research study on whether people can learn to tell real human voices apart from AI-generated ones, and whether that skill carries over to AI voices they haven’t practiced on.
What you’ll do. You’ll listen to a series of short audio clips and decide after each one whether it was spoken by a real person or generated by AI. Partway through, some participants receive a brief lesson on what to listen for; which group you’re in is decided at random. Afterwards you’ll answer a few short questions about yourself. The whole thing takes about 15 minutes and is done in one sitting.
Risks and discomforts. The risks are no greater than those of listening to any audio recording. You control the volume at all times and can lower it or stop whenever you want.
Benefits. There is no direct benefit to you. The results may help show whether teaching people to recognize synthetic voices actually works, which matters because cloned voices are now used in phone scams.
Voluntary participation. Taking part is entirely up to you. You can stop at any point by closing the page, with no penalty and no need to give a reason. If you stop before finishing, your responses are not included in the analysis.
Privacy. We do not collect your name, email address, or IP address. What is recorded: your answers, how long you took on each one, a randomly generated ID, and your answers to the background questions. This is stored in a secure database that only the researcher and his school supervisor can access. Results are reported only as group averages — nothing about an individual participant is ever published.
Who to contact. Hassan Fakhry, hassanfakhrypersonal@gmail.com. Supervising teacher: Brent Buell, Sunset High School, brent_buell@beaverton.k12.or.us.
You must be 18 or older to take part.
This study is publicly pre-registered on the Open Science Framework — osf.io/h8tpr. Our hypotheses, analysis plan, and power analysis were posted before any data was collected.
Before we start
This study is about hearing small differences in voices, so a few conditions really matter for the data to be usable.
Quick background
These help us understand our sample. Neither question affects your results.
How familiar are you with AI-generated voices?
What is your native language?
Headphone check
You'll hear three quiet tones in a row. One of them is softer than the other two. This test only works over headphones — it's how we confirm the audio is set up right.
Trial 1 of 3
Human or AI, and how sure are you?
Play the clip to unlock your answer.
No feedback during tests — just your best judgment.
Phase 2 of 4
You'll practice on real-vs-AI clips with feedback.
Cue
Human or AI, and how sure are you?
Complete
Here's how your ear did across the study. The number that matters most is the last one: how well you told the unseen AI model apart from real voices.
Researcher view
Each dot is one participant. The x-axis is how much training they did; the y-axis is their far-transfer sensitivity (d′) on the strong model they never trained on. The dashed line is chance (d′ = 0). A trained participant who lands well above it was trained only on the weak model, then still discriminated the unseen strong model — that is the transfer signal.