Home · Sep 24, 2026 · Español

LANDR + Mila cut Spotify fraud flags 95%. What their system covers — and what it doesn't.

On September 22, LANDR announced an expansion of the streaming-fraud detection system it built with Mila — Yoshua Bengio's Montreal AI research institute. The headline number: artificial streams flagged by Spotify across the music LANDR distributes fell 95% between Q1 2025 and Q2 2026.

That's a tier-1 signal about the state of the market. The largest indie distributors are now spending on custom fraud infrastructure, in partnership with named academic labs. Streaming fraud is no longer an "eventually" line item — it's a resourced initiative with measurable results.

What LANDR's system actually covers

Per the announcement, two model families do the work:

Profile and behavioral models — assess accounts using profile and behavioral data.
Listening-activity models — examine streams for patterns inconsistent with a real audience (bot-like play distributions, unusual velocity, geographic anomalies).

The new expansion moves detection earlier in the pipeline, adding metadata-layer signals: unusually high release frequency, catalogs that appear overnight, clusters of accounts sharing devices or payout details, and metadata designed to impersonate established artists.

"Finding fraud by hand at our scale was a needle-in-a-haystack problem. The first generation of the model turned that into something we could actually manage. Now we're moving those capabilities earlier in the distribution process, so our team has better signals before suspicious releases reach streaming services." — Tommy Desilets, CTO, LANDR

Three layers. LANDR covers two. The third is where AI-generated fraud lives.

Distributor-side fraud detection actually sits across three distinct layers. Each one catches different things. None of them substitutes for the others.

Layer 1

Behavioral (post-release streams)

Bot play patterns, velocity anomalies, geographic implausibility, coordinated listening clusters. Runs after the release is live and streams start flowing.

Covered by LANDR

Layer 2

Metadata (pre-release)

Release frequency spikes, overnight catalogs, shared devices or payout data across accounts, artist-impersonation naming. Runs on the ingest form before the file is even fingerprinted.

Covered by LANDR (new)

Layer 3

Audio (the file itself)

Is the audio AI-generated? Does the fingerprint already exist in another distributor's catalog under a different artist? Neither behavior nor metadata sees this — it lives in the waveform.

Distinct layer

Layer 3 is where the fraud that clears layers 1 and 2 hides. A track generated with Suno, released once, credited to a plausible-sounding artist with a normal-looking metadata footprint and no bot streams behind it, would pass every LANDR check clean — because nothing in their system listens to the file. Same for a cross-distributor duplicate: same recording uploaded through TuneCore under one artist and through DistroKid under another, with clean metadata on both sides. Only the audio fingerprint matches.

The market implication for the other 100+ indie distributors

LANDR partnered with Mila. That's a Montreal-adjacent advantage — Mila was founded by a Turing Award recipient down the street, and the collaboration produced two production models over an 18-month span. Most indie distributors don't have that option. Building an audio-layer classifier trained on hundreds of thousands of confirmed AI and human tracks is not a "hire one ML engineer" project.

What's available instead is the audio layer as a service. DistroShield takes the file, runs a multi-model AI classifier (trained on a curated corpus that separates Suno, Udio, licensed AI, and unknown generators), checks the acoustic fingerprint against a growing catalog of previously-analyzed recordings across distributors, and returns a signed verdict per track. The output slots in wherever a distributor already has an ingest checkpoint — before the release goes into review, before it hits the DSP handoff.

None of that competes with what LANDR built. It complements it. A distributor running audio-layer detection on top of LANDR-style behavioral + metadata detection catches the fraud that clears the other two layers. A distributor without any of the three catches almost none of it.

If you distribute music and want to see what audio-layer detection looks like on a real track, run one through the free lookup. It's the same engine we expose to B2B customers via API.

Analyze a track →