All articles

How Do Platforms Detect AI Songs?

Recently, I tested some of my songs through three AI music detectors to see what machines will tell me about the music created by a human. One of them marked an old live garage track that I recorded many years ago, which reflects a lot about where this technology works and where it is still failing.

Here is the summary before we get into more details. Platforms detect AI songs based on analysis of the audio itself for tiny artifacts left by generative models, watermark checking, metadata disclosure, and behavioral signals of suspicious upload and streaming activities. None of these techniques is enough on its own, and each platform combines these tools in its own way. Knowing how platforms detect AI songs is important now, regardless of whether you are using the AI tools or not, as the detection mechanism is a part of a larger picture that influences recommendations, monetization, and even content disappearance.

The Numbers That Forced Everyone’s Hand

It became a problem in late 2025. According to Deezer, in July 2026, the amount of fully AI-generated tracks reached 50 percent of all new music uploads for the first time. Deezer reported that the number was 90,000 tracks in June 2026. The Decoder reported that this number was 10,000 in July 2025, while Deezer’s April 2026 release showed that the ratio was 44 percent and increased to 50 in June. According to RouteNote, the share of AI in Deezer’s catalog was already near 28 percent when Spotify introduced its policy in fall 2025, so the process was gradual.

Spotify has its own statistics. The company reported in September 2025 that in the last year, it had removed more than 75 million spammy tracks. Music Business Worldwide reported that number, and it refers to spam, not necessarily the AI-generated tracks. But generative models have made spamming at such a scale cheaper.

Here is the reason why platforms care about AI songs. As Deezer found, up to 85 percent of streams on fully AI-generated tracks in 2025 were fraudulent—bots and manipulations that were trying to steal money from the royalty pool. For the entire catalog, Deezer reports 8 percent of streams fraud. The CISAC and PMP Strategy report quoted by Deezer estimates that nearly a quarter of creators’ revenues may be at risk by 2028, potentially reaching €4 billion. Detection is not about aesthetic principles. It is about the protection of royalties.

Artifact Detection Is the Workhorse

This core technique, in my opinion, the only scalable one, involves training classifiers to spot the fingerprints left by generative models in the audio signal itself. Deezer implemented the detector since January 2025. The tool is trained on the outputs of major AI tools like Suno, Udio, and Riffusion. Deezer’s business FAQ is brutally honest about its philosophy. The system analyzes audio directly, without metadata and without declarative information submitted by the uploader, because anything he claims can be falsified.

The science behind the approach is really interesting. Deezer’s research team has released a paper at ISMIR 2025 proving mathematically that the deconvolution modules inside many generative models create systematic frequency artifacts. They are small but distinctive spectral peaks that are related to the famous checkerboard artifact from the field of image generation. ISMIR (International Conference on Music Information Retrieval) is an academic conference studying the ways how computers analyze music. The crucial conclusion made by Darius Afchar and his colleagues is that these artifacts are the result of the model architecture, not the training data, and therefore it is impossible for a generator to remove them accidentally.

Since every model leaves its own fingerprint, a quality classifier can not only identify the fact that the track is AI-generated but also name the particular tool that has generated it. This is forensically useful. In case of rights disputes, saying that the track was 0.97 probably generated by one particular tool is much more convincing than a vague guess about AI usage.

Claims about the accuracy are strong, though I would treat them with skepticism. Deezer reports that its production system operates at 99.8 percent accuracy—two out of a thousand AI tracks remain undetected and less than one track in ten thousand is falsely recognized as AI with the AI tag applied to the whole album in order to reduce the effect of mistakes. Deezer filed two patents on the methods in December 2024, published by the EU and US patent offices in June 2026, and licensed the technology to the rest of the industry since January 2026 (Sacem, the French rights society was among the test customers). Billboard reported about the Deezer’s technology used to screen chart entries.

Watermarks Help, but Only Where They Exist

The second detection layer is audio watermarking, and here people usually overestimate what they cover. A watermark is a hidden signal that is deliberately added to the audio during its generation. SynthID by Google DeepMind is a well-known example. This signal is embedded in everything produced by Google’s Lyria music model and the podcasts generated by NotebookLM. Google reports that SynthID can survive MP3 compression, added noise, and speed changes. The SynthID Detector portal run by Google counted more than 10 billion pieces of content that already carried the signal when the portal was launched.

Now the catch, and it is a big one. SynthID is a proof of Google provenance only. Two biggest music generators—Suno and Udio—do not embed SynthID. It was pointed out by Eyesift: neither SynthID nor C2PA content credentials can be used to detect the presence of those files. C2PA is a metadata standard that adds the provenance information signed by the provider to the file. The metadata of this type can be removed when someone re-encodes the audio.

So clean watermark scanning is almost meaningless by itself. This is exactly why the artifact classifiers are so important: they work without any cooperation from the generator. The recent licensing agreements, Universal and Udio in October 2025 and Warner and Suno in November 2025, may eventually force the big generators to start implementing proper provenance marking. Until that happens, consider watermarking as a supporting witness, not the whole proof.

Metadata and the New AI Credits Paper Trail

The third detection layer is metadata, and 2026 is the year it was really built. Along with the announcement in September 2025, Spotify backed the new industry standard for metadata of the AI-generated tracks developed through DDEX (Digital Data Exchange), the consortium maintaining data formats for delivering music by labels and distributors. As Music Ally reported in the press briefing, Spotify’s policy lead argued that the controversy with The Velvet Sundown—the AI band that started the storm in 2025—would have unfolded in a completely different way if these credits existed back then.

The system was launched in beta on April 16, 2026. Chartlex’s breakdown of the AI credits rollout explained the process well: you declare the AI involvement at your distributor, the declaration travels inside the DDEX delivery according to the MEAD specification, and Spotify applies the AI tag in the song credits on mobile. The disclosure is detailed, not binary: you can apply the AI tag to vocals without applying it to instrumentals and post-production. According to Dynamoi’s guide to the policy, distributors like Amuse, Believe, CD Baby, DistroKid, Empire, FUGA, and IDOL were developing the rollout as of March 2026.

Distributors became the first filter gate of detection, and this is something most of the artists overlook. Chartlex notes that DistroKid runs its own AI scan before delivering the track to the streaming platforms, and if the scanner detects the AI artifacts in the track where you did not check the AI disclosure box, the upload will be sent to manual verification. Believe has introduced the AI Radar detection software back in November 2023, claiming 98 percent accuracy per Music Business Worldwide, and this system is applied by TuneCore as well. Your music is scanned before being seen by the streaming platform.

One honest caveat: declared metadata is the weakest signal in the whole stack, as liars exist. Platforms understand that, which is why Deezer explicitly refuses to rely on declarations and the disclosure systems are combined with the audio scans to prove that the declaration is true. Think of the credits as of the trust signal that you give, with the audio scan as of the referee in case of contradictions.

Behavior Gives Bad Actors Away

The fourth layer has nothing to do with the sound itself. Spotify’s music spam filter, which was rolled out since fall 2025, targets the behavioral indicators of royalty farming: mass uploads, duplicate tracks, ultra-short filler tracks trimmed to just cross the minimum threshold, and titles packed with search bait. RouteNote describes the policy well: the uploaders and tracks flagged for spam get excluded from the recommendations across Spotify’s programming, which is fatal for the spammers as long as they do not get takedown notices.

Streaming behavior is monitored just as carefully as uploading. Deezer’s fraud department has discovered that the vast majority of streams on AI tracks are generated artificially, and it removes those streams from royalty calculations. As of July 2026 announcement, the platform goes even further: AI tracks used for streaming fraud are removed, AI tracks that have not been streamed for six months are removed, and since April 2026 Deezer stops storing high-resolution versions of AI uploads to save on infrastructure expenses. The end result is quite striking: despite the flood of uploads, fully AI-generated music accounts for about 1 to 3 percent of streams on the platform.

In my opinion, behavioral detection is underestimated in these discussions. Audio classifiers can be fooled at the margins, but a farm uploading four hundred nearly identical lo-fi tracks a week, all of them streamed by the suspicious accounts, flashes like a flare regardless of the audio score.

Voices, Likenesses, and Impersonation Checks

There is another detection problem alongside the bigger one. AI that imitates a particular real person. Spotify’s impersonation policy, effective since September 2025, says that unauthorized vocal clones and deepfakes will be removed, and provides artists with the claims process for removing the content uploaded using the artist’s voice or likeness without permission. Note the framing: the trigger is the impersonation, not the AI. Any clone made with any technology violates the policy.

YouTube developed the most elaborate machinery in this area. Its official AI page describes the technology of synthetic-singing identification inside Content ID, which is piloted now with partners, and can automatically detect AI-generated content mimicking an artist’s singing voice. Content ID is the long-term fingerprinting system of YouTube that matches the uploaded tracks with the database of the registered works. Besides, there is the likeness detection system, which was introduced to a selected group of creators in October 2025, expanded to celebrities and talent agencies like CAA, UTA, and WME in April 2026, and finally extended to the eligible adult creators.

The bigger shift happened in May 2026 when YouTube started automatically detecting and labeling AI videos even if creators skip the self-disclosure required since late 2023. YouTube says explicitly that the disclosure label by itself does not affect the recommendations and the monetization eligibility. What does bite, as Dynamoi’s YouTube guide states, is the July 2025 monetization update that made mass-produced content with no original human input ineligible for revenue.

What Each Big Platform Actually Runs

Deezer runs the whole stack, and it is the only service that labels AI music for the listeners: artifact detection on the upload, the AI tag applied to albums since June 2025, the exclusion from the algorithmic and editorial recommendations, the demonetization of streams, and finally the removal of the fraudulent or dormant AI tracks. In 2025 alone, over 13.4 million AI tracks were detected and tagged in Deezer.

Spotify, and this might be a surprise to you, has not announced any Deezer-like audio classifier that labels the tracks as AI. Its public tooling attacks spam, impersonation, and disclosure, and the company has said explicitly that the goal is not to punish responsible AI usage. We will see whether this position will survive the upload trends of 2026.

YouTube relies on Content ID, the synthetic-singing detection, likeness scanning, and automatic AI labels. Bandcamp went the opposite direction and, as Digital Music News reported in its platform roundup, just banned the music made entirely or mainly by AI, relying on the policy and user reports rather than the public detection system. Apple Music and Amazon Music said less publicly, so they are relying on distributors and metadata at the upload stage.

Want to Test a Track Yourself

You can perform the same checks as the platforms themselves, in a small-scale way. The easiest free option is the Deezer’s own AI music detector for playlists, launched in June 2026, which scans the playlist and tells you what part of it is AI-generated using the same technology as the platform’s upload scanner. Google’s SynthID Detector portal is free as well, and it answers only one question: whether the file carries the Google watermark, so treat the negative result as meaningless for anything outside the Google tools. Google’s SynthID page says that now you can check it in the Gemini app by uploading the clip and asking the question.

On the professional side, there are IRCAM Amplify’s AI Music Detector, launched in May 2024 with the claimed 98.5 percent accuracy per Music Business Worldwide, that now boasts about 99 percent accuracy with under 1 percent of false positives and throughput above 250,000 tracks an hour, though the price is behind the sales conversation. Authio detector from Forward Digital claims 99.42 percent accuracy with the ensemble of twelve models, with the prices starting around 12 euros a month and a free trial. These specifics come from the 2026 detector comparison of Forward Digital, so please note that the company ranks authio detector first because they are the producers of it. ACRCloud launched its detector in January 2026 and, as Rolling Stone India reported in its roundup, it analyzes vocals and instrumentals separately and tries to name the platform that generated the content.

A note from the person who had waded through this swamp: there is a lot of junk in the detector-review space. During my research, I stumbled upon a review site claiming that the rival product does not perform the AI detection at all, while the rival’s documentation and independent industry press claim the opposite. Several of these sites are run by the competing detector companies. Trust the vendor documentation, peer-reviewed research, and your own testing over the affiliate blogs.

And test properly. Create a small collection of tracks that you know are human, add a few tracks generated by yourself, run both of them through the detector you are considering, and estimate the rate of false positives with your own ears. The tool that flags your acoustic demos is useless.

Where Detection Still Breaks Down

The time for an honest reality check, because marketing numbers hide real weaknesses. First, there is a problem of generalization. Classifiers learn the artifacts of the models they were trained on, so a brand-new generator with the different architecture can pass until the community gathers the samples and retrains the classifier. Deezer says that it made some progress towards the detection of AI content without the model-specific training data, but the careful wording shows that it is hard.

Hybrid tracks are the second weakness, and, in my opinion, the one that is the most important. The song with AI-generated instrumentals with the genuinely human vocals or human production polishing the AI-generated content sits in the gray area that current technologies handle poorly. Hybrid detection for partially AI-generated tracks is listed by Deezer’s business FAQ as the active research, which is a polite way of saying that it is not yet solved.

Third, there is the arms race. Heavy mastering, re-encoding, and deliberate processing can mask the artifacts the detectors look for, and the cottage industry now offers the services that promise to make AI tracks undetectable. Most of the time it works worse than it is promised, but it works often enough to matter. Spotify’s cautious rollout of the spam filter exists precisely because everyone remembers the legitimate artists caught in the past crackdowns.

Which brings us to false positives, the failure mode nobody likes to advertise. The spectro’s technical explainer puts it nicely: detection is the probabilistic screening, not the proof, and things like vocoder-heavy human vocals or the lossy re-encoding can trigger false alarms on the totally human music. My garage recording flagged by one of the detectors? It had the distorted room microphone, the crushed dynamics, the cheap encode, and the artifact detection picked synthesis artifacts where there were none.

Please note that the human ears are not the backstop. In Deezer’s Ipsos survey of 9,000 people across eight countries in November 2025, 97 percent could not distinguish the fully AI-generated music from the human recordings in a blind test, while 80 percent still wanted AI music clearly labeled. Machines are not helping our judgments here. They are replacing it.

The same survey found that 73 percent of streaming users want to know when the service recommends AI music, and 52 percent do not think that fully AI songs belong in the main charts next to the human work. Listener sentiment, not just the fraud math, is what drives platforms towards the detection and labeling of the AI content.

What I Would Do Before My Next Release

If you release music, a few actions keep you on the right side of this topic. Declare the AI usage in the credits honestly, even for the partial AI usage like AI-assisted stems, since the checkbox exists now in DistroKid and other major upload workflows and inconsistencies between your declaration and the scan trigger the hold. Never use the voice that imitates a real artist without the written authorization, because that is the category every platform removes on sight.

Avoid anything looking like spam even if your intention is pure. Do not bulk-upload the near-duplicates, do not pad your release with the filler trimmed to just cross the payout threshold, and do not stuff your titles with the search bait, because behavioral filters do not read your heart. Keep your session files, stems, and project timestamps for everything you create, because provenance evidence is exactly what will win your appeal in case of the false flag. And before releasing something important, run your masters through the free Deezer detector or the trial of one of the paid tools, so the surprise comes from your dashboard and not the distributor’s rejection letter.

The bigger picture is worth to reflect on. How platforms detect AI songs went from the academic curiosity to the core infrastructure in about eighteen months, and the direction is clear: audio forensics is the core of detection, credits and watermarks add the provenance where cooperation is available, and behavioral filters catch the fraud which is driving the flood. The detectors will get better as long as billions of euros in royalties are at stake. Make honest music, document how you made it, and the machinery will work for you.

Sources

Common questions

How do AI music detectors actually work?

They are classifiers trained on the outputs of tools like Suno, Udio and Riffusion to recognise fingerprints in the audio itself. Deezer's research team showed at ISMIR 2025 that the deconvolution modules inside many generative models create systematic spectral peaks, similar to checkerboard artifacts in image generation, and that these come from the model architecture rather than the training data. Because each model leaves its own fingerprint, a good classifier can often name the specific tool that generated a track.

How accurate is AI music detection?

Deezer reports 99.8 percent accuracy for its production system, meaning about two in a thousand AI tracks slip through and fewer than one in ten thousand human tracks is wrongly flagged, with tags applied to whole albums to limit mistakes. IRCAM Amplify claims around 99 percent with under 1 percent false positives, and Authio claims 99.42 percent. The article advises treating vendor numbers with scepticism and testing detectors on your own known human and AI tracks.

Do Suno and Udio songs have a watermark?

Not one that standard scanners can read. Google's SynthID is embedded in everything from its Lyria model and NotebookLM podcasts and survives MP3 compression and speed changes, but it only proves Google provenance. Suno and Udio do not embed SynthID, and C2PA metadata can be stripped by re encoding, so watermark scanning alone tells you little about the two biggest generators. Artifact classifiers matter precisely because they work without the generator's cooperation.

Does my distributor scan my music for AI before Spotify sees it?

Yes, and this is the gate most artists overlook. DistroKid runs its own AI scan before delivery, and a track showing AI artifacts without the disclosure box ticked goes to manual verification. Believe's AI Radar, claiming 98 percent accuracy, is used by TuneCore as well. Your music is checked before any streaming platform ever receives it.

What are Spotify's AI credits?

A disclosure system launched in beta on April 16, 2026, built on the DDEX metadata standard. You declare AI involvement at your distributor, it travels inside the delivery according to the MEAD specification, and Spotify shows an AI tag in the song credits on mobile. The disclosure is granular, so you can tag AI vocals without tagging instrumentals or post production. Declared metadata is the weakest signal in the stack, which is why platforms pair it with audio scans.

Can Spotify detect AI music from the audio?

Spotify has not announced any Deezer style audio classifier that labels tracks as AI. Its public tooling targets spam, impersonation and disclosure, and the company has said its goal is not to punish responsible AI use. Deezer runs the full stack, from artifact detection at upload through labelling, recommendation exclusion, demonetisation and removal, and tagged over 13.4 million AI tracks in 2025 alone.

How do platforms catch AI music spam without analysing the audio?

Through behaviour. Spotify's spam filter targets mass uploads, duplicate tracks, ultra short filler trimmed to cross the payout threshold and titles stuffed with search bait, and flagged uploaders lose recommendations across the service. Deezer strips artificial streams from royalty calculations, removes AI tracks used for fraud or dormant for six months, and found that up to 85 percent of streams on fully AI tracks in 2025 were fraudulent against 8 percent for the whole catalogue.

How does YouTube detect AI singing and voice clones?

YouTube runs synthetic singing identification inside Content ID, piloted with partners, to spot AI mimicking an artist's voice, plus a likeness detection system extended to talent agencies in April 2026. Since May 2026 it automatically detects and labels AI videos even when creators skip self disclosure. The label alone does not affect recommendations or monetisation, but the July 2025 policy makes mass produced content with no original human input ineligible for revenue.

Can I test whether my song will be flagged as AI?

Yes. Deezer's free playlist detector, launched June 2026, uses the same technology as its upload scanner. Google's SynthID Detector only answers whether a file carries Google's watermark. Professional options include IRCAM Amplify, Authio from about 12 euros a month with a free trial, and ACRCloud, which analyses vocals and instrumentals separately. Be wary of detector review sites, several of which are run by competing vendors.

Where does AI music detection still fail?

New generators with different architectures can pass until classifiers are retrained, hybrid tracks mixing AI instrumentals with human vocals sit in a grey area Deezer still lists as active research, and heavy mastering or re encoding can mask artifacts. False positives are real too: the author's old garage recording with a distorted room mic and cheap encode was flagged as AI. Detection is probabilistic screening, not proof, which is why keeping session files and stems matters for appeals.