If you want to use song lyrics by Ger­man musi­ci­ans to train a lan­guage model, it would pro­ba­b­ly be wise to obtain a GEMA licence. That is the impli­ca­ti­on of the Munich Regio­nal Court ruling from 11 Novem­ber 2025 – a signi­fi­cant set­back for ope­ra­tors of gene­ra­ti­ve AI.

GEMA sued Ope­nAI becau­se ChatGPT used song lyrics from nine well-known Ger­man artists, inclu­ding Her­bert Grö­ne­mey­er and Inga Hum­pe, wit­hout pay­ment. It was undis­pu­ted that the lyrics were used for trai­ning pur­po­ses. The court lar­ge­ly upheld the lawsu­it and orde­red Ope­nAI to cea­se and desist and pay dama­ges (Case ID: 42 O ¹⁴¹³⁹⁄₂₄).

Accor­ding to the court, both the memo­ri­sa­ti­on of the song lyrics in the lan­guage models and their repro­duc­tion in the chat­bo­t’s out­puts con­sti­tu­te an inf­rin­ge­ment of copy­right explo­ita­ti­on rights. The­se inf­rin­ge­ments are not cover­ed by any rest­ric­tions – in par­ti­cu­lar not by the rest­ric­tion on text and data mining. Howe­ver, the court rejec­ted the cla­im that the use of the lyrics vio­la­ted the artists’ gene­ral per­so­nal rights.

The ruling only appli­es to the older ver­si­ons GPT‑4 and GPT-4o. Whe­ther newer ver­si­ons also inf­rin­ge copy­rights remain­ed dis­pu­ted and unre­sol­ved. Both par­ties had reques­ted a refer­ral to the CJEU – the court decli­ned. Ope­nAI announ­ced it is con­side­ring next steps: “We dis­agree with the ruling and are con­side­ring next steps.” GEMA has alre­a­dy laun­ched a second case against Suno AI.

For AI deve­lo­pers, this means: the com­mer­cial use of copy­righ­ted con­tent for trai­ning AI models can trig­ger a licen­sing obli­ga­ti­on. For rights hol­ders, their nego­tia­ting posi­ti­on vis-à-vis AI pro­vi­ders has been signi­fi­cant­ly streng­the­ned. Accor­ding to Ope­nAI, the ruling has no impact on the mil­li­ons of peo­p­le, busi­nesses and deve­lo­pers who use ChatGPT every day.

The Munich ruling adds to a gro­wing wave of copy­right liti­ga­ti­on against gene­ra­ti­ve AI. In the US, The New York Times lawsu­it against Ope­nAI and Micro­soft is pro­cee­ding after a fede­ral judge in March 2025 rejec­ted Ope­nAI’s moti­on to dis­miss. That case – alle­ging that Ope­nAI used mil­li­ons of artic­les wit­hout per­mis­si­on to train ChatGPT – could set a his­to­ric copy­right pre­ce­dent for the AI era.

See: https://www.juve-patent.com/cases/open-ai-must-pay-gema-licence-fee-for-chatgpt/

To what ext­ent can AI such as ChatGPT be used for ques­ti­ons on the inter­pre­ta­ti­on of fea­tures or inven­ti­ve step?

This recent T‑decision (T 1193‍/‍23) of the Euro­pean Patent Office, in which Bös­herz Goe­bel Patent­an­wäl­te were invol­ved as repre­sen­ta­ti­ves, now cla­ri­fies (machi­ne trans­la­ti­on):

“The gene­ral­ly incre­asing spread and use of chat­bots based on lan­guage models (”lar­ge lan­guage models“) and/​or ”arti­fi­ci­al intel­li­gence“ alo­ne does not jus­ti­fy the assump­ti­on that an ans­wer recei­ved […] neces­s­a­ri­ly cor­rect­ly reflects the under­stan­ding of the per­son skil­led in the respec­ti­ve tech­ni­cal field (at the rele­vant time)”

(see decis­i­on text)

The EPO Board of Appeal empha­si­zes that appro­pria­te tech­ni­cal lite­ra­tu­re must be con­sul­ted ins­tead for the cor­rect inter­pre­ta­ti­on of patent claims. In this case, the ques­ti­on was whe­ther ChatGPT ans­wers could be used to inter­pret the term – an approach that the Board cle­ar­ly rejec­ted.

This decis­i­on is a first, important land­mark for the limits of AI use in patent pro­cee­dings and under­lines the con­tin­ued importance of sound tech­ni­cal exper­ti­se.