Why ChatGPT Can't Tell You Who Owns a Song's Rights

The short answer

You can ask ChatGPT or Google who owns a song, and you will get an answer. It will sound confident. It might even be partly right. But a confident answer is not a cleared song, and the distance between the two is exactly where licensing deals fall apart.

Here is the problem in one line: an AI model tells you the most likely publisher for a title, while clearance requires every rights holder for a specific recording, down to the last fraction of a writer's share. Those are different questions. The first is a guess. The second is a reconciliation you have to get right, because one missed co-writer can undo the whole thing.

What actually happens when you ask an AI

A language model does not look anything up. It predicts text. When you type "who publishes [song]", it returns the name that most often appeared near that title in its training data. That is a reasonable way to answer trivia and a dangerous way to plan a clearance, for a few concrete reasons.

It answers about the title, not your recording. "Hallelujah" has hundreds of recordings. The Leonard Cohen master, the Jeff Buckley master, and a wedding singer's cover are three completely different clearances that share one composition. Ask a model "who owns Hallelujah" and it cannot know which audio is sitting in your edit. Only the recording itself, identified by its ISRC, tells you that.

It flattens master and publishing into one answer. Every commercial song is two separate assets: the master recording (controlled by a label or the artist) and the composition (controlled by the songwriters' publishers). You clear both, usually as two deals with two fees. A chat answer that names "the label" or "the publisher" has quietly hidden half the deal from you. If that split is new to you, start with master vs publishing: who you actually pay.

It cannot count the writers. A modern chart record routinely has four to eight co-writers, each with a separate publishing deal, each share possibly administered by a different company. A model might name the one publisher that shows up most in press coverage and silently drop the other five. You would never know they were missing, which is the most expensive kind of error.

It does not age well. Catalogs get sold. Administration deals expire and move. A writer signs to a new publisher. The web article the model learned from was accurate in 2019 and wrong today. Registries get updated; a training snapshot does not.

The mistake that actually costs money: the missing share

This is the part I want you to sit with, because it is the one that turns a good project into a legal headache months after delivery.

When a song has several co-writers, their shares are supposed to add up to exactly 100%. Say a track has five writers: 40%, 25%, 20%, 10%, and 5%. Each of those five sits under a publisher, and every one of those publishers has to say yes for the song to be cleared. Not the big three. All five.

Now imagine your research (or an AI answer, or a single database lookup) surfaces four of them and quietly misses the writer holding 5%. On paper you feel done. You have "the publishers". You sign, you pay, you deliver, the film goes out. Then, sometime later, the administrator for that last 5% notices the use and gets in touch. They did not approve it. They can demand a fee on their terms now, when you have zero leverage, or they can force the music out of a finished piece. A 5% share carries a 100% veto.

That is why the working standard is not "did I find some publishers", it is "do the writer shares I have found add up to 100%, and can I account for every name". If the shares total 97%, you are not almost done. You are missing someone, and finding out who is the job. A model cannot do this reconciliation for you, because it never had the share percentages in the first place. Even a small share you cannot see is a party who can stop you.

How a registry-first lookup works instead

The professional path does not start with a name. It starts with the recording and works outward, confirming each layer against a source you can point to.

1. Start from the master. The ISRC (International Standard Recording Code) pins the exact recording you are using, so you are never clearing the wrong version.
2. Pin the composition. From the recording, resolve the ISWC (International Standard Musical Work Code), the composition's ID in the global works registry under the CISAC standard. This anchors identity to the work itself rather than a spelling of the title.
3. Reconcile the writers and publishers. Cross-check the writer and publisher picture against public registries. In the US that means the MLC (the Mechanical Licensing Collective) for publisher and writer registration; internationally the CISAC ISWC network ties works across societies. The goal is a writer list whose shares reconcile toward 100%, with a publisher attached to each.
4. Confirm the master side separately. Release metadata and label credits identify who controls the recording, which is a different question from who controls the song.
5. Label the confidence and the gaps. Every layer gets a source, and anything thin or unmatched is flagged rather than smoothed over.

The point is not that registries are magic. It is that they are checkable. When a source says a publisher controls a share, you can go look, and you can see what is still missing. An AI answer gives you a name with no share, no source, and no way to know what it left out.

A note on databases people reach for: community sources like MusicBrainz are genuinely useful as an ISRC-to-work bridge, but they are community-maintained, not an official society registry, so they are a lead to confirm rather than the final word. The same caution applies to any single registry on its own. The confidence comes from the cross-check, from names that agree across ISWC, the MLC, and the release credits, and from shares that add up.

What a registry lookup honestly cannot promise

I would rather you trust the tool because it tells you the truth about its limits than because it oversells.

• It does not hand you a signed deal. Identifying the controlling parties is step one. Someone still has to negotiate and paper it. See how to clear a song for a film or commercial for the full sequence.
• It shows the controlling publisher, not every territorial sub-publisher. No database reliably maps sync rights territory by territory. For a clearance in a specific country you confirm the local administrator before signing.
• It is a starting point for the fee, not a quote. Rights holders quote their own numbers based on your exact use. A realistic range tells you whether a song is even in your budget; the 2026 sync fee guide explains what moves the number.
• Thin data stays flagged as thin. When a recording cannot be matched to an ISWC, the publishing contacts are provisional, and the right move is a human re-check, not blind outreach.

Honesty about these limits is the whole trust proposition. A chatbot will never tell you which parts of its answer to distrust. A good lookup will.

When is Google or ChatGPT actually enough?

Sometimes it genuinely is, and pretending otherwise would be its own kind of dishonesty.

If you are using a fully independent, artist-owned track where one person controls both the master and the publishing, a quick search plus a direct message can be the whole clearance. One party, one yes, nothing hidden. For that case, do not overthink it.

The stakes rise with the song. The moment you are dealing with a well-known recording, a multi-writer chart record, a major-label master, or any use going out for broadcast or paid media, the gap between "a plausible name" and "every party accounted for" becomes the gap between a deal and a dispute. And it rises again outside your home territory, where the administrator you would contact locally is often not the one a US database shows.

The rule of thumb: the more famous the song, the more writers on it, and the wider your release, the less an AI answer is worth and the more the reconciliation matters.

Find your rights holders, then verify them

That reconciliation is exactly what Music Oracle was built to do. Paste a track and it returns the master side, the publishing side with the publishers it can identify, a complexity flag telling you whether you are looking at a clean one-stop or an eight-approval marathon, and a realistic fee range for your use. Every result carries its sources, and the gaps are labelled rather than hidden.

Then, when the picture is complicated and the deadline is real, a Quick Check (€150) puts a supervisor's eyes on the full shortlist: which tracks are cleanly clearable, which will fight you, and where the writer shares still need chasing so nothing is missed. Already locked on one song and do not want to hunt down five publishers yourself? Single Track Clearance (from €300) covers the outreach and negotiation for one track, any territory. Book a Quick Check first and the €150 credits toward clearance within 60 days.

Key takeaways

• AI and Google answer about a title; clearance is about a specific recording and every party attached to it.
• A model gives you a name with no share and no source. It cannot tell you who it left out.
• The real risk is the missing share: a co-writer you never found holds a full veto over your use, even at a few percent.
• The working standard is shares that add up to 100% with a publisher for each, verified against the registries, not a single confident guess.
• A registry-first lookup (ISRC to ISWC, cross-checked against the MLC and the CISAC ISWC network) is checkable; an AI answer is not.
• Use a plain search only for simple one-stop indie tracks. The more famous the song and the wider the release, the more the reconciliation matters.
• Whatever you find online, confirm it in the official registries before you email anyone, especially the writers and their publishers.