Contents
There is no single answer, because the field produces two kinds of numbers and both are real. On the benchmarks they were built for, the best 2026 detectors are close to perfect: F1 scores above 0.99, error rates under 5 percent. On unseen generators, edited files, and real broadcast audio, the same detector families drop to the sixties, or below coin-flip. Neither number is a lie. The conditions decide which one applies to your track, and the headline number is almost always the first kind.
The lab numbers are genuinely excellent
Start with what the field can do at its best. ArtifactNet, a 2026 forensic detector that reads the residual left by the neural audio codecs inside most music generators, reports an F1 score of 0.9993 with a 0.09 percent false-positive rate on the SONICS benchmark, a corpus of 97,164 songs built for exactly this task (Oh, 2026, arXiv:2604.16254). MusicDET, which models the statistical density of real music rather than any generator’s fingerprint, reaches a 4.51 percent average equal error rate zero-shot, meaning on generators it never saw in training, and 2.89 percent on SONICS (Han, Wang, Gui, 2026). Closed-set attribution on FakeMusicCaps, where the candidate generators are all known in advance, reaches an F1 near 1.00. These are not cherry-picked outliers; they are the current state of the art, properly measured. Inside their evaluation conditions, modern music detectors work.
The out-of-distribution numbers are a different world
Move any distance from those conditions and the floor drops. The SONICS paper itself reports that its detector’s F1 falls by 20 to 64 percent across harder and unseen partitions of its own benchmark (Rahman, Paul, Hakim, Sarker, Fattah, 2025, ICLR). CLAM, a detector that reads vocal-to-instrumental coherence, posts a 0.925 F1 on its home benchmark; evaluated on another lab’s benchmark it fell to 0.7576 with a 69.26 percent false-positive rate, and a separate study measured its detection rate on held-out Udio tracks at 2.6 percent (Batra et al., 2025; Park et al., 2026). Park, Kim, Koh and Saito, probing detectors across generator shift, put the diagnosis plainly: standard benchmark splits are “nearly saturated” while held-out generators expose the instability underneath (arXiv:2606.08663). A number earned on one benchmark, in other words, routinely fails to transfer to the next one.
Real-world audio is harsher still
The newest evidence tests something closer to deployment. AI-OpenBMAT, a 2026 broadcast benchmark from López-Ayala and colleagues, evaluates detection under real broadcast conditions: short excerpts, music under speech, low signal-to-noise mixes (arXiv:2602.06823). The best detector tested, the same SpecTTTra family that scores around 93 percent on clean audio, managed an overall F1 of 61.1 percent, falling to 33.2 percent on low-level background music, while a CNN baseline reached 27.6 percent. That is not an adversary at work; it is ordinary radio and streaming conditions. The lab-to-wild gap has its own article in why music detectors fail in the real world; the broadcast numbers say the gap is widest exactly where detection is most wanted.
Accuracy depends on what the detector reads
Different detectors read different traces, and they do not fail the same way, which is why one accuracy figure never covers the field. Some read a high-frequency codec artifact, some read long-range song structure, some model the statistical density of real music, and some transcribe the lyrics and classify the words. The fragile family is the pure artifact reader: excellent on the generator version it knows, then collapsing on a version shift, as when the Deezer Fourier detector fell from 100 on seen Suno v3.5 and Udio-130 to 39.83 on unseen Udio-32 (Afchar, Meseguer-Brocal, Akesbi, Hennequin, 2025). The more durable families are not magic either. MusicDET is strong across generators yet fragile to editing, its 4.51 percent error rate rising toward 44 percent under a pitch shift; ArtifactNet is codec-aware, cutting cross-codec drift by 83 percent from 0.95 to 0.16, but it still leans on the neural-codec residual it was built to read. Before trusting any score, how AI music detectors work is what tells you which reader produced it.
Why the spread is so large
Accuracy in this field is conditional, and the conditions compound. A detector is accurate on generators in its training data, on clean unedited audio, at durations it was built for, in the foreground. Each departure, a new Suno or Udio version, a platform re-encode, a 5 second clip, a mix with speech over it, moves the operating point from the first set of numbers toward the second, the fragility examined in can AI music detectors be fooled?. Commercial tools sit under an extra layer of uncertainty: their accuracy claims are vendor-published without independent benchmarks, and where one was checked, against IRCAM Amplify’s advertised “99% accuracy,” an independent study measured a 4.7 percent false-positive rate on human music, well above the advertised under 1 percent (Cros Vila et al., 2025, TISMIR), the gap unpacked in do AI music detectors falsely flag real music?.
So what accuracy should you expect?
Ask three questions of your track before trusting any published figure. Is the generator mainstream and older than the detector’s training data? Is the file clean, or has it passed through platform compression and editing? Is it full-length foreground music? Three yes answers put you near the lab numbers, high 90s. Each no moves you down through the SONICS cross-partition drops, past the broadcast 61.1 percent, toward the held-out-generator collapse. The safest way to read any result is tiered: a high-confidence score on a clean original from a known generator is useful evidence; a score on a compressed or edited file is weaker; a score on a short broadcast clip is weaker still; and a score from a tool that will not disclose its threshold or false-positive rate is not a verdict. Detection is close to solved for yesterday’s generators on clean audio, and unsolved for new generators, processed files, and real deployment audio, so what a score means depends on how far your track sits from the conditions the number was earned under. The cross-media version of this question is covered in are AI detectors accurate?.
Sources
- Oh (2026). ArtifactNet: Forensic Codec-Residual Detection of AI-Generated Music. arXiv:2604.16254.
- Han, Wang, Gui (2026). MusicDET: Zero-Shot AI-Music Detection via Frequency-Guided Normalizing Flows. arXiv:2605.18072.
- Rahman, Paul, Hakim, Sarker, Fattah (2025). SONICS: Synthetic Or Not - Identifying Counterfeit Songs. ICLR 2025. arXiv:2408.14080.
- Batra, Sharma, Thukral, Bhatia, Batra (2025). Melody or Machine: Detecting AI-Generated Music. arXiv:2512.00621.
- Park, Kim, Koh, Saito (2026). Probing Token Spaces for AI-Music Attribution under Generator Shift. arXiv:2606.08663.
- Afchar, Meseguer-Brocal, Akesbi, Hennequin (2025). A Fourier Explanation of AI-Music Artifacts. ISMIR 2025. arXiv:2506.19108.
- López-Ayala, Cabello, Zinemanas, Molina, Rocamora (2026). AI-Generated Music Detection for Broadcast Monitoring. ICASSP 2026. arXiv:2602.06823.
- Cros Vila, Sturm, Casini, Dalmazzo (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. TISMIR 8(1). DOI:10.5334/tismir.254.