What $150,000 a Song Means for Suno AI Music Generation

One hundred fifty thousand dollars is the ceiling US copyright law sets for statutory damages on a single work infringed willfully, where the work was registered in time. A plaintiff who gets there does not have to prove a lost sale or a displaced stream. They need the registration, the work, and a finding of willfulness. Then they multiply by the catalog.

That multiplication is why every docket entry in the Suno AI music generation cases gets read twice by people who have never opened a DAW. The reading is often wrong, though, because the headline outcome at any given stage says very little about where the exposure actually sits. A copyright case against an AI music tool is not one question. It is a sequence, and each stage in it decides something narrow and specific. Read in order, the mechanism is legible.

The first act is a copy nobody hears

Before a prompt is typed, there is an ingestion pipeline: acquiring audio, decoding it, resampling it, storing it, and passing it through training in some resident form. Plaintiffs plead each of those steps as a reproduction under Section 106(1), and it is the easiest allegation in the case to make plausible, because it does not depend on what the model later produces. If the recordings were in the training set, copies were made.

The defense to that act is not "no copy." It is fair use, drawing on the intermediate-copying line that runs through Sega v. Accolade and Sony v. Connectix, where an unlicensed copy made in service of building something new was excused. Running alongside it is a provenance fight that is factual rather than doctrinal: how the audio was obtained. Downloading from a licensed source, scraping, and stream-ripping are three different stories, and only one of them keeps the argument confined to fair use.

The second act is an output, and it does not sound like the briefing

Here the case gets harder for plaintiffs, and the reason is technical. A generation request returns a mixdown — commonly two to four minutes of stereo audio at 44.1 or 48 kHz, sometimes with stems, with no seam showing where an interpolated melody would be. Musical infringement analysis was built for lead sheets: melodic contour, harmonic rhythm, lyrics, a recognizable hook. Most AI renders sit in a genre-average zone that is stubbornly hard to litigate — a four-chord loop at 120 BPM under a vocal timbre that evokes a named artist without reproducing a protected phrase. Style is not protectable. A hook is. Output claims live in the gap between them.

That gap narrows when a model regurgitates. Prompt at a specific artist, era, or song and generators sometimes return something recognizable, and near-verbatim outputs have been the most quotable exhibits in this litigation. But regurgitation is intermittent, prompt-dependent, and awkward to reproduce on demand. Which brings the mechanism to its most misread stage.

Does a plaintiff have to name a specific infringing song to sue?

No — not at the pleading stage. Federal pleading requires a plausible claim, not proof. A plaintiff who alleges ownership of registered works and alleges copying can proceed to discovery without attaching a side-by-side comparison for every work in the schedule. "They have not shown us the infringing song" is a summary-judgment argument wearing pleading-stage clothes, and as of writing, courts handling the US suits against AI music developers have treated it that way at exactly the moment the defense most needed otherwise.

The consequence is structural. A defendant can win real narrowing at this stage — theories trimmed, some claims preempted, categories of relief cut away — and still be precisely where it did not want to be. Narrowing removes arguments. It does not remove the works, does not shrink the discovery obligation, and does not change per-work exposure on the claims that survive. A ruling can be honestly reported as a win and still function as a scheduling order for the expensive part.

Discovery is where the balance sheet changes

What gets produced next is not audio. It is dataset manifests, scraper configurations and logs, vendor and hosting agreements, engineering decisions about deduplication and filtering, retention policies for source files, internal memorization testing, and the thread in which somebody asked whether a license was needed and somebody answered.

Willfulness is proved out of that record, not out of a waveform. This is the stage that sets the multiplier on the number this piece opened with, and it is also where the cost asymmetry bites. The plaintiffs are institutions with catalogs and standing litigation budgets. The defendant is a company with payroll, a product roadmap, and a fundraising calendar that does not pause for a custodial collection.

The CMI claim keeps its own clock

The DMCA claim under Section 1202(b) runs in parallel and on different elements. It asks whether copyright management information — the identifying data attached to a file — was removed or altered, and whether the defendant knew or had reason to know that removal would conceal infringement. It does not require any output to be substantially similar to anything. The statute's damages provision runs from $2,500 to $25,000 per violation, and what counts as a single violation is contested, which is one reason this claim tends to survive early motions with its arithmetic intact.

For anyone modeling exposure, the CMI track deserves its own line. It can be worth more than the output claims, and it turns entirely on facts the defendant controls.

Fair use arrives last, applied to a record already written

Fair use is the defense the whole field is arguing about, and it is resolved at summary judgment or trial — after ingestion facts, provenance facts, and willfulness documents are fixed. The Supreme Court's 2023 decision in Warhol v. Goldsmith made commercial substitution harder to write off as transformative under the first factor, which pushes weight onto the fourth, market effect. There, rightsholders make an argument that is circular and effective: a market for training licenses exists, they have licensed into it, and unlicensed training displaces it.

So the doctrinal question that will eventually be reported as the outcome of these cases gets decided against evidence produced two stages earlier by a defendant who had no dismissal available to stop the production.

Diligence worth doing before it is your problem

If your client uses these tools rather than builds them, the exposure is different and the questions are narrower.

That last row is the one clients underestimate. A game studio with 400 adaptive loops baked into a shipped build is not exposed to the lawsuit. It is exposed to the remedy.

Back to the number

One hundred fifty thousand dollars per work was never a prediction. No stage in the first years of these cases tests whether it applies: pleading does not, narrowing does not, and discovery does not decide it — discovery prices it. The number's function is to make settlement rational long before any court reaches the merits, which is why the likely end of the AI music copyright fight is not a fair use ruling at all. It is a rate card.