Is Imael Angel's 'Bad Times' AI? We Made an AI Copy to Find Out

The question hanging over 'Bad Times' is whether a person made it. We turned it around. Instead of hunting for signs of a machine in Imael Angel's song, we made a machine version ourselves and looked for the places where the two part company. If the copy matched everywhere, nothing in the sound could count as evidence. Where it failed, the failure would show what a prompt cannot fake.

What we gave the machine

Only numbers. We measured 'Bad Times' at 107.7 beats per minute in F minor, with a groove that stays level after a short intro and a high, chanted vocal. We typed those facts into ACE-Step 1.5, a free, open-source music model, running on an ordinary home computer with no graphics card, and added lyrics we wrote ourselves. No audio, melody, lyric or artist name from the song went in. Each take of 2 minutes 48 seconds took between 34 and 37 minutes, and the model used up to 23 GB of memory while it worked.

We made three takes. The first two came from a prompt that asked for warm bass, soft hi-hats and airy pads. For the third we rewrote it to ask for a crisp, bright top end, bright shakers and a clear vocal. The video above plays the first and third takes with the measurements on screen. These are the two prompts, word for word:

• Takes 1 and 2: Hypnotic deep house. Steady four-on-the-floor kick from the first bar, warm rolling sub bass, soft off-beat hi-hats and shakers, airy pads and one simple repeating synth motif. Confident female vocal, half sung and half spoken, with a short chanted hook repeated many times. Minimal arrangement, no big drop: the groove stays level and trance-like from start to finish. Late-night club mood.
• Take 3: Hypnotic deep house with a bright, crisp top end. Steady four-on-the-floor kick and tight sub bass from the first bar, crisp open hi-hats and bright shakers on every off-beat, shimmering high synth stabs. Clear, bright, forward female vocal, half sung and half spoken, with a short chanted hook repeated many times. Polished modern club mix, sparkling and airy on top. The groove is locked in from the start and stays level, no big drop.

What the copy got right in half an hour

Everything a listener names first. All three takes came out at 107.7 beats per minute and in F minor, the same readings we got from the original. The four-on-the-floor kick, the level groove and a short chanted hook repeated many times arrived on the first try, from one paragraph of text.

That is the first clue, and it cuts against both sides. The song's signature is cheap to make: a home computer produces it in the time it takes to cook dinner. But the same signature proves nothing, because a producer working by hand can hit 108 beats per minute in F minor just as easily. Tempo, key and groove are not evidence either way. Anyone who calls the song AI because of how it sounds is pointing at exactly the parts a person could have made too.

Where the copy fell short

The top of the sound. Over each full take, we measured how much of the sound sits above 8 kHz, the range where cymbals, breath and air live. The original's official 30-second preview has 3.24 percent there. Our first take had 0.44 percent, the second 0.71. Put another way, 99 percent of the original's sound sits below 11,100 Hz, and 99 percent of our first take below about 5,000 Hz. Our copies also took longer to arrive: the original is above 80 percent of its full level by the second eighth of the track, our takes only by the fourth or fifth.

Then we tested the gap instead of admiring it. Changing a few words, warm and soft out, crisp and bright in, lifted the third take to 1.28 percent above 8 kHz and moved its 99 percent line up to about 9,240 Hz. One rewrite closed a large part of the distance: the third take has about 2.9 times the treble of the first, and about 40 percent of the original's.

That is the second clue, and it is a warning. The one clear difference we could measure shrank with a single change of wording. A gap that moves that easily measures effort, not a human hand. We did not run more rounds, so we cannot say how far it closes. We can say that the brightness of 'Bad Times' cannot settle the question by itself.

What the test does say

Pace. One take cost us about 35 minutes on a home computer. Deezer counts 32 Imael Angel releases in the 173 days from 30 March to 19 September 2026. At our speed, a month of that output is an evening's work for a machine. That does not show the catalogue was made this way. It shows that the pace is no obstacle if it was.

Put together, the experiment narrows the question instead of answering it. The parts of 'Bad Times' that people point to, its tempo, its steady, hypnotic groove and its chanted hook, are the parts a free model copies first. The part our copy missed, the bright finished sound, is the part that moves most with effort. Neither is a fingerprint. Imael Angel's other hit shows a third route, an AI idea rebuilt by people, which we traced in how 'Movin' To The Sun' was made.

The one piece of evidence our test cannot produce is the one still missing from the public record on Imael Angel: word from the artist or the label on how the song was made. Until that arrives, our copy leaves one firm result. Nothing you can hear in 'Bad Times' proves who, or what, made it.