Human or AI
Machines write music now, and it shows less and less. One clip, one call, one mistake allowed.
Five games drawn at random, one try each →Five games drawn at random.
Loading the game…
How to tell AI generated music from the real thing
Music generation models have taken a leap in two years. On a short clip, plenty of people can no longer tell, and that is exactly what this mini-game measures. There are still tells, though, and they can be learnt.
Listen to the voice first. Consonants are where machines are weakest: the s, the t and the k are often smeared, as if chewed. Breaths are missing, or land in the wrong place. Then listen to the cymbals and the top end, which easily turn into a shimmering mush instead of clean hits.
One last tell, subtler: does the track go anywhere? A real recording breathes, builds, falls back, lets an instrument take over. A generated one tends to hold the same intensity from start to finish, never risking anything. If everything is perfectly smooth for fifteen seconds, be suspicious.
Can you really tell the difference?
Yes, but less and less easily. On a clip of a few seconds, even musicians get it wrong regularly. That is what makes the exercise interesting rather than discouraging.
Which tells should you listen for first?
The consonants in the voice, how clean the cymbals are, and whether the track goes anywhere. Do not try to recognise the title: the real tracks are often obscure, and recognising them is not the question.
How many mistakes do you get?
One in practice mode, and it ends the run. That is what gives every listen its weight, and what makes you take the time to doubt.
Five games drawn at random, one try each →Five games drawn at random.