A 1984 Song, "Made With AI": How Music Detectors Turn Probabilities Into Verdicts
Updated:
According to one commercial AI-music detector, "Across the Ocean," a 1984 recording by the Swedish post-punk band Twice a Man, was made with a generator called "Treblo_v3." The same tool attributed the band's 1982 track "In My Mind" to "ElevenLabs Music." Both songs predate those products by roughly four decades.
The results come from tests that Digital Music News reported on November 4 and VICE amplified a day later. The tests were not about the old recordings at first. They were run to defend a new album: DECAY / TIME — A Tribute to Twice a Man, a 23-track double album released in June 2026 by the Hamburg musician André Grafert, who records as Coloured Tears. Grafert says he played, programmed, sang and produced every track himself, with no generative AI, and lists his toolchain as Cubase, Sequential's Prophet X, Prophet 12, Pro 3 and Rev2 synthesizers, Native Instruments' Kontakt, and his own voice. Soon after release he faced public accusations of using AI anyway.

Photo by Paul Morigi/WireImage for NARAS (lead image of the VICE article)
This is the uncomfortable twist in the AI-music cleanup: tools built to protect human musicians from a flood of synthetic tracks can also misjudge human musicians — and the people most exposed are the independents who have no label lawyer to call. This site has already covered lawsuit-driven takedowns at DistroKid and YouTube Music's labeling policy; this piece looks one layer down, at the classifier itself: what it measures, how often it is likely to be wrong, and what should happen when it is.
What the tests showed, and what they didn't
The testing was done by Ines Weidemann of International Artist Marketing (IAM), the PR firm working with Grafert. According to DMN's follow-up, SubmitHub's detector rated all 23 album tracks as mostly AI or hybrid, and three other detectors — ACRCloud, Let'sSubmit and TrackVerifier — mostly agreed. Seven original Twice a Man recordings from 1982–1986 also came back as "AI/hybrid" on SubmitHub.
That is the headline. The details are messier, and they matter more:
- The audio source changed the verdict. The seven flagged originals were submitted as YouTube links. Run through SubmitHub via Spotify links, the same seven came back as human — "Decay" at 85%, "Sorrow" at 83%, "In My Mind" at 91%. SubmitHub told DMN that it tested the 1982 album itself and got 85–95% "human" on all ten tracks, and that it discourages low-quality YouTube audio.
- Detectors disagree with each other. On Coloured Tears' latest five-song single, none of 15 results across SubmitHub, Let'sSubmit and TrackVerifier came back as primarily human. But TrackVerifier scored four of the five at 10.5 out of 100 — a "weak AI signal" — while Let'sSubmit called all five AI at 77–95% confidence.
- One detector disagreed with itself. Let'sSubmit, given the same MP3 a week apart, moved "Flat Land" from "Probably Human" (61%) to "Probably AI" (65%).
There are also limits on the evidence. All of these tests come from one side of a dispute: the artist's own publicist. The method has not been published in full, and no independent party has examined Grafert's project files. SubmitHub's representative raised questions about how the vocals were processed, which DMN called fair but unresolved. So the Grafert case is a well-documented anecdote, not a measured error rate. What it does show — and SubmitHub's own pushback confirms this — is that the same song can get opposite verdicts depending on the file you upload and the week you upload it.
What a detector actually measures
The common wording — "AI detectors estimate probabilities rather than examine how a song was made" — is correct, but it skips the more useful point: what the probability is based on.
Deezer's FAQ describes its tool as identifying "specific audio artefacts and signatures associated with generative AI models," signals that are "generally inaudible to listeners." The academic work agrees, and is more blunt about the consequences. In a peer-reviewed study in Transactions of the International Society for Music Information Retrieval (TISMIR), Laura Cros Vila, Bob L. T. Sturm and colleagues trained detectors on 20,000 Suno and Udio tracks and 10,000 human tracks and found the task "easy." But then they took the systems apart:
- Simply resampling audio to 22.05 kHz changed the outputs of every detector tested, including the commercial IRCAM Amplify system.
- High-pass filtering at 8 kHz or above made IRCAM Amplify label all audio as AI, human recordings included.
- The datasets differed in plain technical ways — Suno files at 192 kbps, Udio at 320 kbps, almost all at 48 kHz, while 60% of the human set was 44.1 kHz. A rule using sample rate alone would have reached 83% precision.
- When faced with 50 tracks from Boomy, an AI generator outside its training data, IRCAM Amplify identified only 3 as AI.
The authors' conclusion is the key point of this whole debate: a detector trained this way tells you whether "a given audio file has characteristics consistent with the production pipeline of Suno, Udio, or something else" — not whether a human made it.
Deezer's own researchers have said much the same. Darius Afchar, Gabriel Meseguer-Brocal and Romain Hennequin titled a 2024 paper "Detecting music deepfakes is easy but actually hard": their classifier reached 99.8% test accuracy, but they warned that "a good test score is not the end of the story" and listed calibration, robustness to audio manipulation, generalization to unseen models, interpretability and recourse as open problems. A July 2026 paper co-authored by Afchar and Hennequin says existing detectors' performance "collapses under simple audio manipulations, such as speed modification or pitch shifting."
Why might synth-heavy, retro-sounding records be at higher risk? No study has tested this directly; IAM itself says it knows of no research on niches such as darkwave. What follows is the author's inference from the findings above, not an established result:
- The "human" reference set is narrow. TISMIR's human data came from the Million Song Dataset, mostly mainstream commercial Western music, and DMN notes that human benchmark sets generally lean toward Western pop. A tightly quantized drum machine, a sequenced bass line and heavily processed vocals sit far from that center, and a binary classifier has to put such outliers somewhere.
- Generators were trained on these genres too. If generator output imitates 1980s synth timbres, then the region of "sounds" where AI and human overlap includes exactly these timbres.
- Transcoding distorts the signal. Rips from YouTube, re-encodes, and older masters change the high-frequency band and the sample rate — the same incidental features TISMIR showed can flip a verdict. The Twice a Man results (flagged via YouTube, cleared via Spotify) fit this explanation, though they do not prove it.
When a score becomes a verdict
A probability on its own harms no one. The damage comes when someone in the distribution chain treats it as a pass/fail line.
SubmitHub explicitly says its "93% human"-style results express model confidence, not the share of a song that was machine-made. But its own results page says tracks at 85% or above are not accepted, and its September 2026 AI policy excludes tracks that score at least 85% AI-likely from its promotion services, hybrids included, where writing the lyrics or composition does not exempt a recording. At launch, DMN reported there would be "no appeals process"; in November, DMN wrote that SubmitHub "does offer an appeals process." The two reports conflict, and we could not confirm which describes current practice.
Further along the chain, the stakes grow:
- Platforms with bans. Bandcamp announced in a January 13, 2026 Reddit post that music made "wholly or in substantial part" with generative AI is no longer allowed. (VICE dated the ban to June 2026; Billboard's contemporaneous report puts it in January.) Bandcamp asked users to flag suspected tracks for review by its team, so a public accusation — perhaps based on a screenshot from a free detector — is a direct path to enforcement.
- Streaming services. Deezer removes tracks it detects as AI from algorithmic recommendations and editorial playlists and labels them for listeners. Its tag is applied at the album level, which Deezer says reduces the risk of misclassification but also means that one flagged track labels the whole release.
- Money and accounts. IAM argues that false positives lead to blocked accounts, blocked releases and withheld royalties, mostly for independents who use DIY distributors and face automated filters. That is IAM's assessment; DMN did not document a specific royalty freeze in the Grafert case.
The asymmetry is the core of the indie argument. DMN notes that the industry's new labeling approach relies on voluntary self-disclosure by labels and distributors. A major-label act's AI status is disclosed by the label; an unsigned producer's AI status is inferred by software. In Grafert's words, newcomers are being asked to prove "with individual tracks that they are real people, while no one would ever doubt the big, established acts."
The base-rate arithmetic
Vendor accuracy numbers sound reassuring. Deezer says its system is 99.8% accurate, missing about 2 AI tracks per 1,000 and wrongly flagging "fewer than one genuine human-made track in every 10,000." SubmitHub cites a UMG-commissioned test putting its detector at 99.3% accuracy. Neither benchmark has been published in a form outsiders can reproduce, and the Coloured Tears campaign calls such figures "confidential industry benchmarks."
What a false-positive rate means in practice depends on volume and on how common AI tracks are in what is being screened. The following is an illustrative calculation by the author. The inputs are assumptions, not measurements:
| Scenario | Assumptions | Human tracks wrongly flagged | Share of flags that are wrong |
|---|---|---|---|
| A. Deezer-scale, vendor rate | ~90,000 AI + ~90,000 human uploads/day (Deezer says it received nearly 90,000 AI tracks a day by July 2026, and AI exceeded 50% of daily uploads at its June peak); false-positive rate 0.01% | ~9 per day, ~3,300 per year | ~0.01% |
| B. Same volume, research-measured rate | As above, but 4.7% false-positive rate (IRCAM Amplify on human music in TISMIR — measured on only 150 human samples, a different product from Deezer's) | ~4,200 per day | ~4.5% |
| C. A curator's or small label's inbox | 1,000 submissions, 10% AI (SubmitHub said ~11% of recent uploads scored above 85%); detector catches 95% of AI | at 1% FPR: ~9; at 4.7% FPR: ~42 | ~9%; ~31% |
Two points follow. First, at Deezer's stated rate and Deezer's scale, the base rate actually works in the detector's favor: when half of everything is AI, a flagged track is almost certainly AI — if the stated rate holds. Thousands of wrong flags a year is still thousands of real records, but it is not a crisis.
Second, everything depends on whether the stated rate holds for your music. An average false-positive rate across a benchmark can hide a much higher rate for a subgroup that the benchmark barely includes. If darkwave or EBM records were wrongly flagged at, say, 20% (a hypothetical figure; nobody has measured it), the platform-wide average would barely change, while for the artists in that scene a flag would be closer to a coin toss. That is why the Grafert case, anecdotal as it is, raises a fair question: the people most likely to be misjudged are those least represented in the test sets that produced the reassuring number.
The case for detecting anyway
None of this is an argument for turning detection off, and the counter-arguments deserve their full weight.
The flood is real and largely fraudulent. Deezer received about 10,000 fully AI-generated tracks a day when it launched its tool in January 2025; by July 2026 it was nearly 90,000. Those tracks draw only 1–3% of streams, but Deezer found up to 85% of those streams were fraudulent in 2025 and excludes them from royalty payments. Every fraudulent stream that gets paid is taken from the same pool that pays human artists, including Grafert. A study by CISAC and PMP Strategy, cited by Deezer, estimates nearly 25% of creators' revenues could be at risk from AI by 2028.
Vendors also have reason to object to the framing. SubmitHub calls DMN's claim of "rampant" false positives unsupported, and nothing published so far shows a systematic error rate. One album, one PR firm's tests, and unverified vocal processing do not make a statistic. Some platforms also design their systems for caution: Deezer's tool targets fully AI-generated tracks and "does not typically tag AI-assisted tracks," and Bandcamp's process, according to its announcement, routes reports to human reviewers rather than banning automatically.
The fair summary is that detection is needed, current detectors are useful at scale, and they are least reliable at the margins — which is where a lot of independent music lives.
What due process should look like
If detection is here to stay, the question becomes procedure. Much of what is needed already exists in other areas of content moderation. In the EU, the Digital Services Act requires online platforms to give a statement of reasons when they remove content, restrict its visibility, or suspend payments (Article 17), and to offer an internal complaint system whose decisions are not taken "solely on the basis of automated means" (Article 20). Whether and how this applies to each music service and distributor is a legal question beyond this article, but the principles transfer well (the author's view):
- Notice. Tell the artist that a track was flagged, by which system, and with what consequence — label, demotion, rejection or payment hold.
- Reasons. Report the score, the model version and the audio source analyzed. "We analyzed a YouTube rip" and "we analyzed your WAV master" are different findings.
- A real appeal. Accept evidence detectors can't see — project files, stems, session screenshots, hardware recordings — and have a human with relevant genre knowledge review it.
- Thresholds aren't verdicts. A confidence score should trigger review, not automatic punishment. Naming a specific generator, as in "Treblo_v3" for a 1984 track, should require far more evidence than a probability.
- Publish subgroup error rates. Vendors who sell accuracy figures should report false-positive rates by genre, era and audio format — and fix results that change from one week to the next.
The problem the Grafert case exposes is not that detectors exist, but that a statistical guess is being treated as a judgment with no one accountable for it. Fixing that does not require choosing between human musicians and the fight against AI spam. It requires treating a classifier score as what its own makers say it is: evidence to review, not a ruling.
Sources
- VICE — False Positives from AI Detection Tools Threaten the Indie Musicians They're Supposed to Protect (original article)
- Digital Music News — More AI Detector False Positives Emerge In Rapid Succession (Nov 4, 2026)
- Digital Music News — Human Enough? How AI Detector Scores Put Indie Artists at Risk (Nov 8, 2026)
- Digital Music News — SubmitHub Bans AI Music Submissions (Sep 17, 2026)
- Cros Vila, Sturm, Casini & Dalmazzo — The AI Music Arms Race: On the Detection of AI-Generated Music, TISMIR (2025)
- Afchar, Meseguer-Brocal & Hennequin — Detecting music deepfakes is easy but actually hard (arXiv, 2024)
- Dugelay et al. — Improved Robustness in AI-Generated Music Detection (arXiv, 2026)
- Deezer Newsroom — AI-generated tracks now represent 44% of all new uploaded music (Apr 20, 2026; FAQ figures updated to July 2026)
- Billboard — Bandcamp Bans AI-Generated Songs in a First for Major Online Music Platforms (Jan 14, 2026)
- Regulation (EU) 2022/2065 — Digital Services Act