Trained to Imagine
Rendering
Creative Intelligence
AI Production / Hollywood / Legal

German court rules against Suno in landmark GEMA training-data case

Source: Variety ·

The Munich Regional Court found that Suno breached German and US copyright law by training on GEMA's repertoire, and held that AI companies must license material for systematic use and commercial exploitation.

TTI Takeaway

For anyone producing for the European market this is the most consequential ruling of the year. It moves training-data provenance from a reputational question to a licensing obligation, and German and EU clients will start asking which models were used and on what basis. Answering that requires records kept during production, not reconstructed afterwards.

Analysis

What makes the Munich decision consequential is where it lands and how it frames the problem. In the United States, the training-data fight is still argued as fair use, one case at a time. In Munich, a regional court treated the systematic use of a protected repertoire to build and run a commercial model as something that has to be licensed, and according to GEMA the decision covers both model training and output generation. That moves the question from whether training is permitted to what it costs.

The case itself was narrow. GEMA, which has around 100,000 members, built it on six songs, among them “Atemlos,” “Rasputin,” “Forever Young” and “Mambo No. 5,” which it said it could essentially reproduce on Suno through basic prompting. The court found Suno in breach of German and US copyright law by training on GEMA’s repertoire and by storing and reproducing the works. The full reasoning has not been published; most of what is known comes from GEMA’s own statement. Suno told Variety it disagrees with the ruling, calls it a fundamental mischaracterisation of how its technology works and how US law applies, and is evaluating its options, including an appeal. The win follows GEMA’s case against OpenAI over unlicensed lyrics roughly nine months earlier, and GEMA has since published a proposed licensing system for AI use of music.

In practice, the ruling touches two stages of production. Upstream, the question of which model generated an asset, and on what licensing basis it was trained, stops being a footnote in the brief and becomes something a German or EU client’s legal and procurement teams will ask before they accept delivery. Downstream, the reproduction finding matters as much as the training finding. A generated temp cue or sound-alike dropped into an edit is not a harmless placeholder if the model can return recognisable protected material on request. Our view is that generated music should be handled the way library music already is: a documented source, known terms, and a note of the tool and version, recorded when the cue enters the timeline rather than reconstructed when someone asks.

The American track runs in parallel. In August, Universal and Sony were allowed to add stream-ripping claims to their case against Suno, though the judge refused to widen it beyond the 560 named tracks; Variety notes Warner Music has already settled with Suno on licensing terms. In October, Presley rights holders filed suit against Anthropic alleging that recordings were used without permission to train Claude; Anthropic has said it disagrees and will defend the case. Taken together with the EU AI Act labelling rules that took effect in August, European productions now face a licensing expectation on the way in and a disclosure expectation on the way out.

The limits are real. This is a first-instance judgment, an appeal is possible, and the reasoning is not public, so how far it generalises to image and video models trained on other kinds of work is not something the ruling, as reported, settles. Two things are worth watching: whether Suno appeals and on what grounds, and whether GEMA’s licensing proposal attracts signatories. If it does, that becomes the first working price list for training data in Europe, and other rights holders will have a template to point to.