David Sterling: Megan, quick question before we get into it — did anything about the Suno announcement this week actually surprise you, or was it the expected move?
Megan Skiendel: Honestly? The timing surprised me more than the tool itself. Because announcing Musixmatch Sentinel — real-time copyright screening for prompts and AI outputs — while Universal Music Group and Sony Music are in active litigation against you is a very specific kind of bold.
David Sterling: Bold or calculated.
Megan Skiendel: Calculated, almost certainly. And here's what the coverage keeps missing — Sentinel and the RIAA litigation are not solving the same problem. Sentinel is an inference-time tool. It screens what users generate today. The lawsuits are about training-time copyright: what Suno fed the model before any of us ever typed a prompt. Those are structurally separate crises, and Mikey Shulman's framing — artists get disclosure options, platforms get transparency — that framing only addresses one of them.
David Sterling: Wait, so when he says artists get options — he means options over outputs, not over whether their catalog was used in training.
Megan Skiendel: Right, and the labels know that. Which is probably why Warner Music Group took a different path — settled November 2025, entered a licensing deal — while Universal and Sony kept litigating. Warner read the situation and decided the fight over training-data liability wasn't worth holding out for. The other two apparently disagree.
David Sterling: So the real question isn't whether Sentinel works. It's whether it changes anything for the cases that are actually in front of judges right now.
Megan Skiendel: And my read is — it doesn't. It might be very good technology. It might even be sincere. But a post-hoc screening tool cannot retroactively consent to the copyrighted works already baked into the model's parameters. That gap is where this whole story lives.
David Sterling: Let me try to make that gap concrete, because I think it's actually simpler than people are letting on. Imagine a chef who learned to cook by secretly reading thousands of cookbooks — none of them hers, none of them licensed. Then she installs a plagiarism checker on the menu printer. The checker catches if a dish description copies a recipe word-for-word. But it cannot un-teach her everything those cookbooks put in her head.
Megan Skiendel: That's Sentinel exactly.
David Sterling: That's Sentinel. It's a menu printer checker. The training-time problem is the chef's brain.
Megan Skiendel: And what the GEMA court actually found — the memorization finding — is that the chef can reproduce a recipe almost verbatim from a prompt. Not paraphrase it. Reproduce it. That's not inference-time behavior. That's evidence of what was absorbed at training time.
David Sterling: Which means Sentinel catches the symptom. Not the cause.
Megan Skiendel: And the watermarking announcement has the same blind spot — actually, wait, it's worse in one specific way. Suno named no technology provider. No rollout timeline. This is early August 2026 and it's still vaporware. Sentinel at least is real and running. The fingerprinting is a press release.
David Sterling: Hold on — no named provider at all?
Megan Skiendel: None. And that's not a minor detail. Because if you're trying to signal to Universal and Sony that you're structurally serious, you name the vendor. The fact that they didn't tells you the watermarking announcement was aimed at judges and partners, not engineers.
David Sterling: This is the 'train first, license later' pattern — the whole sequence. Build on unlicensed material, deploy, then roll out the compliance infrastructure once the suits are filed. And courts are now explicitly scrutinizing that order of operations.
Megan Skiendel: And that sequencing gets worse once you read what GEMA actually put in front of the court. The ruling — reported August 5, 2026 — wasn't a fuzzy inference about aesthetic similarity. The court found memorization. Near-complete copies of specific songs, stored in the model, reproducible almost verbatim from a simple prompt. That's forensic, not theoretical.
David Sterling: Hold on — verbatim? From a prompt?
Megan Skiendel: Almost verbatim, yes. Which detonates Suno's entire 'new sounds' defense. Because that defense only works if the model is synthesizing something genuinely novel. If the court can show the model is — I mean, not interpolating, not referencing — actually retrieving stored copies of protected works, then the output isn't a new sound. It's a playback.
David Sterling: That's the mechanism Sentinel can't touch. Sentinel is downstream of the retrieval.
Megan Skiendel: Picture a session musician in Hamburg. Friday night, she opens Suno, prompts it for a cinematic orchestral swell — nothing lyrical, no words, no title reference. Sentinel screens it. Clean. She uploads to Spotify. But the harmonic structure she got was statistically shaped by a memorized training work — no lyric match, no audio fingerprint hit, zero disclosure. She has no mechanism to know. Suno has no mechanism to tell her. That's transparency asymmetry made real.
David Sterling: And that's before you get to the jurisdictional claim, which — frankly — nobody is talking about enough. A German court asserting authority over U.S.-based AI training because outputs were served to EU users. That's not just a GEMA problem.
Megan Skiendel: Every AI music company with European users is technically exposed by that precedent. Every single one.
David Sterling: The load-bearing assumption the industry's been running on — train in the U.S., export globally, litigate under U.S. fair use doctrine — that assumption now has a crack in it.
Megan Skiendel: And wait until we get to what Sentinel's adoption mid-litigation actually signals to a judge — because the 'we're fixing it now' move has a very specific legal risk that makes the whole transparency argument flip on Suno.
David Sterling: That flip is the part I want to stay on, because it's not hypothetical. Walk through the logic with me. Sentinel exists. A court finds training was infringement — and the fair use question in the U.S. is still genuinely unresolved as of right now, August 2026, no clean precedent. If that ruling comes down, what does Sentinel look like in retrospect?
Megan Skiendel: It looks like acknowledgment. That's the problem. You built a tool to screen outputs for copyright violations — a court can read that as: you understood the outputs were potentially tainted.
David Sterling: And chose output management over retraining.
Megan Skiendel: Instead of licensing upfront. Which is the 'train first, license later' charge — procedural compliance instead of a structural fix. Sentinel becomes the exhibit, not the defense.
David Sterling: I mean — courts read contemporaneous remediation exactly that way when underlying liability is established. The remedy implies the company understood there was something to remediate.
Megan Skiendel: And Warner settling in November 2025 and licensing — that's one version of the endgame. Retroactive licensing. Universal and Sony not moving there yet means the cost of that train-first sequence is still being calculated in front of judges. Nobody's closed the book.
David Sterling: So Mikey Shulman is — wait, actually, is there any reading of the Sentinel announcement where it genuinely reduces exposure? Or is that incoherent once you accept the training-time framing?
Megan Skiendel: Good faith. That's the only live argument. It signals to a judge — and to potential licensing partners — that Suno isn't reckless. But good faith on outputs doesn't cure bad faith, or even negligence, on training. Those are separate findings. The procedural compliance doesn't reach the structural remedy.
David Sterling: The fact is, the asymmetry cuts both ways then. Warner moved. Universal and Sony haven't. And the tool that was supposed to close the gap might be the document that proves Suno knew it existed.
Megan Skiendel: When you started this — did anything surprise you — I said the timing surprised me more than the tool. But honestly, sitting here now, it's almost the inverse. The tool is what should have surprised me. Because Sentinel can tell you exactly what a user prompted. It can flag a lyric match, trace a composition back to a rightsholder through Musixmatch's database. That part is real. What it cannot do — what no inference-time tool can do — is tell you which copyrighted works are already living inside the model's weights. Suno can't disclose that. They may not even fully know it.
David Sterling: The transparency asymmetry is structural, not a gap they can engineer around.
Megan Skiendel: And until a U.S. court actually rules on fair use for AI training — which hasn't happened, not cleanly, not in the music context, not as of right now — that asymmetry is the whole ballgame. The GEMA finding, Warner settling, Universal and Sony still in the room with judges — none of it resolves the foundational question. Sentinel just watches the door. The thing courts are actually arguing about was baked into the walls before Mikey Shulman ever announced anything.
