The Munich Regio­nal Court (LG) ruled in a judgment dated Novem­ber 11, 2025 (Ref. 42 O 14139/24) deci­ded that repro­du­ci­b­le trai­ning texts available in a model (here: ChatGPT 4 and 4o from Ope­nAI) (“Memo­rizati­on„) as dupli­ca­ti­on within the mea­ning of Sec­tion 16 of the Ger­man Copy­right Act (UrhG). It is suf­fi­ci­ent that the trai­ning texts are repro­du­ci­b­ly available in the model:

The plain­ti­ff claims that the chat­bot gene­ra­tes Repro­duc­tions of the trai­ning data to a con­sidera­ble ext­ent. This so-cal­led Memo­rizati­on of con­tent within models leads to the Regur­gi­ta­ti­oni.e. to pro­du­ce out­put that expli­ci­t­ly repro­du­ces cer­tain trai­ning inputs […] […] 

201 a. The Cham­ber is con­vin­ced that the texts at issue are […] Inclu­ded in the model.

202 aa. It is known from infor­ma­ti­on tech­no­lo­gy rese­arch that trai­ning data can be con­tai­ned in models and can be extra­c­ted as out­puts, which is refer­red to as memo­rizati­on […]. Such memo­rizati­on occurs when the unspe­ci­fic para­me­ters in trai­ning do not just extra­ct infor­ma­ti­on from the trai­ning data set, but a com­ple­te trans­fer of the trai­ning data can be found in the para­me­ters spe­ci­fi­ed after the trai­ning.

203 The mul­ti­ple occur­rence of a trai­ning date in the trai­ning set is assu­med to be the cau­se of memo­rizati­on, which main­ly occurs with lar­ge models […].

204 The memo­rizati­on of trai­ning data can be veri­fi­ed using various methods. If the trai­ning data is known, it is pos­si­ble to compa­re the trai­ning data with out­puts using simp­le prompts and suf­fi­ci­ent text length to deter­mi­ne memo­rizati­on. Other­wi­se, the para­me­ters entro­py and per­ple­xi­ty are used to exami­ne the cer­tain­ty with which a model repro­du­ces an out­put – in the case of trai­ned and memo­ri­zed con­tent, the cer­tain­ty is high […]. Con­tra­ry to the defendant’s state­ment, simp­le prompts are not a con­di­ti­on for gene­ra­ting the trai­ning data as out­puts, but mere­ly ser­ve to pro­ve memorization. […] 

205. the Memo­rizati­on can alre­a­dy be deter­mi­ned here by com­pa­ring the lyrics with the out­puts. The use of the dis­pu­ted song lyrics as trai­ning data is undis­pu­ted. Accor­ding to Annex K 2, the song lyrics in dis­pu­te are cle­ar­ly reco­gnizable in the sub­mit­ted out­puts by the very simp­le prompts “What are the lyrics of [song tit­le]”, “Who wro­te the lyrics”, “What is the cho­rus of [song tit­le]”, “Plea­se also tell me the first ver­se”, and “Plea­se also tell me the second verse”.

The fact that texts have been fed in as trai­ning data and are repro­du­ced during queries con­sti­tu­tes pri­ma facie evi­dence that the texts are stored in the model in dupli­ca­ted form. Fur­ther­mo­re, dupli­ca­ti­on does not requi­re a work to be repro­du­ced iden­ti­cal­ly. It is also suf­fi­ci­ent to spe­ci­fy a work in a modi­fi­ed form. The tech­ni­cal details are also irrelevant:

For repro­duc­tion under copy­right law how memo­rizati­on works in detail remains open. It is irrele­vant whe­ther one speaks of sto­ring or copy­ing the trai­ning data or, as the defen­dants put it, whe­ther the model reflects in its para­me­ters what it has lear­ned based on the enti­re trai­ning data set, name­ly rela­ti­on­ships and pat­terns of all words or tokens that repre­sent the diver­si­ty of human lan­guage and its con­texts. This is becau­se it is cru­cial that the song lyrics that ser­ved as trai­ning data are repro­du­ci­b­ly con­tai­ned in the model and thus embodied.

The fol­lo­wing was then not appli­ca­ble TDM excep­ti­on (§ 44b UrhG):

Lan­guage models such as the models in dis­pu­te gene­ral­ly fall within the scope of the text and data mining rest­ric­tions. The regu­la­ti­ons cover neces­sa­ry dupli­ca­ti­ons when com­pi­ling the data cor­pus in pha­se 1 (see abo­ve), but not fur­ther dupli­ca­ti­ons in the model in pha­se 2. If, as in the pre­sent case, infor­ma­ti­on is not only extra­c­ted from trai­ning data in pha­se 2, but works are also repro­du­ced, this does not con­sti­tu­te text and data mining. Even if the limi­ta­ti­ons pro­vi­si­ons gene­ral­ly app­ly to the trai­ning of models, repro­duc­tions in the model are not repro­duc­tions that are cover­ed by the limi­ta­ti­ons pro­vi­si­on, as they are not only used to prepa­re the text and data mining.

The judgment was issued by Mathi­as Lejeu­ne comm­ents.