AI and OER: How a legally compliant application can succeed

Do arti­fi­cial intel­li­gence (AI) and open edu­ca­tional resources (OER) go together? Does their rela­tion­ship have a future? In this blog post, we show why AI and OER are a match and how AI tools can be used for OER cre­ation in a legally com­pli­ant way.

This is the sec­ond post in our blog series “AI in the uni­ver­sity”. After the first post was about AI detec­tors in the first post, in this post we looked at AI and OER from a legal per­spec­tive. This will soon be fol­lowed by an arti­cle on prompt tips for high-qual­ity OER. We will then con­clude the series with manda­tory AI skills in higher edu­ca­tion.

Recently, in the atten­tion econ­omy, the topic of “AI” has almost sup­planted the topic of “OER” in the uni­ver­sity con­text. Every­one is talk­ing about AI: AI projects are being funded, AI work­shops are fully booked. OER has moved down the list of pri­or­i­ties. These devel­op­ments are under­stand­able. And yet it should not be for­got­ten that not only AI, but also OER has great poten­tial for teach­ing. So why not com­bine the two top­ics to make the most of their poten­tial?

The inter­ac­tion between AI and OER can bring con­sid­er­able added value for teach­ing: AI can pre­pare open edu­ca­tional mate­ri­als in a didac­ti­cally clever and appeal­ing way, struc­ture them and much more. In turn, OER can help to ensure that AI pro­grams are trained with high-qual­ity mate­ri­als and thus deliver more reli­able results.

And what does the law say about this rela­tion­ship? To what extent is a legally com­pli­ant use of AI pro­grams in OER even pos­si­ble? We will try to answer this ques­tion here. We will only address OER-rel­e­vant aspects (exclud­ing data pro­tec­tion aspects). Please also note that legal devel­op­ments in the field of AI are very dynamic. The expla­na­tions in this blog post refer to the cur­rent legal sit­u­a­tion (Octo­ber 2024).

Is AI output protected by copyright?

AI out­put (e.g. images, texts, info­graph­ics, etc.) is in the pub­lic domain under Ger­man copy­right law, i.e. free from copy­right. This is because, accord­ing to the clear word­ing of Sec­tion 2 (2) UrhG, only humans can be the cre­ators of an intel­lec­tual out­put. Con­se­quently, the AI can­not have any copy­rights to the out­put.

Oper­a­tors of AI pro­grams as legal enti­ties are not con­sid­ered authors either. The author­ship of the AI out­put does not lie with the pro­gram­mers of the AI pro­gram either, because AI is gen­er­ally not con­trol­lable. It func­tions autonomously. The same applies to users of the pro­gram. It is the­o­ret­i­cally con­ceiv­able to manip­u­late AI out­put with very clever prompts (com­mands). How­ever, it will gen­er­ally not be pos­si­ble to con­trol the AI entirely with these. As a result, although the smart prompt is pro­tected by copy­right as the user’s work, the out­put remains in the pub­lic domain.

Some­thing else may apply if the AI is only used as a tool in the cre­ative process (i.e. only as a sup­port), but the actual per­for­mance can be objec­tively attrib­uted to a human being. This can be com­pared to the use of an image edit­ing pro­gram. In this case, the copy­right lies with the cre­ator of the work.

Copy­right pro­tec­tion can also arise if the AI prod­uct is exten­sively cre­atively reworked by a human. The result of the adap­ta­tion must be suf­fi­ciently indi­vid­ual, i.e. take on the per­sonal char­ac­ter­is­tics of the cre­ator of the work. This is not the case if only col­ors or sizes are adjusted or addi­tions are made. Whether the required indi­vid­u­al­ity and cre­ativ­ity (level of cre­ativ­ity) are present always depends on the spe­cific indi­vid­ual case. You can find out more about this in the arti­cle AI and OER: How well do they go together? by Georg Fis­cher.

Can AI-generated output be used in OER? What applies to prompts?

As the pure AI out­put is in the pub­lic domain, it can gen­er­ally be used in OER mate­ri­als. How­ever, this requires that no third-party rights (e.g. copy­right, per­sonal rights, trade­mark rights, etc.) are infringed by the out­put.

The pub­li­ca­tion of infring­ing con­tent under an open license can be warned. This is because copy­right law does not pro­tect the good faith belief that the con­tent used does not infringe the prop­erty rights of third par­ties. Rather, users have a duty to check the con­tent used for infringe­ments, which is not easy with AI-gen­er­ated con­tent. Users can­not know which sources are used to train the AI and whether the rights of third par­ties have been infringed. The­o­ret­i­cally, it is even pos­si­ble for the AI to repro­duce third-party works with which it has been trained in its out­put one-to-one or in part, but this is more of an excep­tion.

Most AI pro­grams do not copy other peo­ple’s works, but cre­ate pat­terns based on train­ing data, sim­i­lar to the human brain. The prob­a­bil­ity of being warned is there­fore not par­tic­u­larly high. Nev­er­the­less, it is impor­tant not to accept the gen­er­ated result with­out reflec­tion, but to check for legal infringe­ments, e.g. with pla­gia­rism soft­ware
or a reverse search in the trade­mark reg­is­ter. If pos­si­ble, the AI-gen­er­ated result should be revised sev­eral times with addi­tional prompts or other pro­grams.

The Euro­pean AI Reg­u­la­tion(AI Act), which came into force on August 1, 2024, requires oper­a­tors of gen­er­a­tive AI pro­grams to develop a copy­right com­pli­ance strat­egy and pub­lish detailed sum­maries of train­ing data (recitals 106 and 107). How­ever, it will take some time for AI oper­a­tors to ful­fill this oblig­a­tion, as the reg­u­la­tion is being imple­mented grad­u­ally.

Cau­tion is also required when for­mu­lat­ing prompts. No prompts should be pub­lished that con­tain copy­righted works or parts thereof, third-party brands, per­sonal pho­tos or other per­sonal data. This can also result in a warn­ing.

Can the AI be trained with copyrighted works? Can I prevent the AI from being trained with my works?

In order for the AI to be able to gen­er­ate con­tent itself, it must first be trained with open sources (i.e. also with OER). This is prob­a­bly legally per­mis­si­ble in Ger­many. It is pre­dom­i­nantly argued that the copy­right bar­rier for text and data min­ing in Sec­tion 44 b UrhG jus­ti­fies the train­ing.

Train­ing AI with third-party con­tent is not so easy to pre­vent. Many read­ers will remem­ber the case involv­ing Meta in the sum­mer of 2024. The NRW con­sumer advice cen­ter had issued a warn­ing to Meta because the com­pany wanted to use user data on Insta­gram and Face­book to train AI with­out the users’ con­sent. Ulti­mately, the users were allowed to object to the train­ing. How­ever, the way to object was some­what com­plex.

For con­tent on web­sites and blogs, it is pos­si­ble to declare a reser­va­tion of use against the train­ing, the so-called opt-out from data min­ing. This must be done by stor­ing a file in the root direc­tory of the domain, i.e. in machine-read­able form, so that it can be read by web crawlers. Web crawlers are pro­grams that search the Inter­net and ana­lyze web­sites. How exactly this works is explained, for exam­ple, in the tuto­r­ial How to block Ope­nAI’s Chat­GPT, Google’s Gem­ini and other bots that want to use your texts for their AI by Kai Spriesters­bach. How­ever, an opt-out can also result in web­site con­tent not being dis­played as search results.

Do I have to label AI-generated content as such?

Since AI-gen­er­ated con­tent is in the pub­lic domain, there is cur­rently no legal oblig­a­tion to cite the source. Nev­er­the­less, mate­ri­als cre­ated with AI should always be labeled as such. This makes it eas­ier for sub­se­quent users to under­stand the cre­ation process and the ori­gin of the mate­r­ial. It is also made clear that the mate­r­ial is in the pub­lic domain, i.e. not sub­ject to any copy­right restric­tions. The label­ing could, for exam­ple, be as fol­lows: “AI-gen­er­ated. Pub­lic domain. Cre­ated with the pro­gram XY. Prompt: XY. Edits: XY”.

The OER mate­r­ial in which AI-gen­er­ated con­tent is used can be openly licensed. Licenses should be cho­sen that are the least restric­tive for sub­se­quent use, i.e. CC Zero release, CC BY and CC BY SA. The pub­lic domain AI con­tent must be excluded from the license: “This OER mate­r­ial is licensed under license XY. The marked AI-gen­er­ated images are not cov­ered by the license. These are in the pub­lic domain.”

Some­thing dif­fer­ent applies to the so-called remix­ing (merg­ing) of mate­ri­als into a new work. If AI-gen­er­ated con­tent and the OER mate­r­ial are merged in such a way that a new work is cre­ated (e.g. an image col­lage, a video), the new over­all work may be openly licensed. Till Kreutzer explains this as an exam­ple in the video Open Edu­ca­tional Resources, Copy­right and AI.

Conclusion

AI pro­grams can be used to cre­ate OER in com­pli­ance with the law if the above rules are observed. If you would like to find out more about the topic or would like to deepen what you have read, visit our work­shop AI & OER in use: Enhanc­ing OER in a diverse and legally com­pli­ant way with AI. For reg­is­tra­tion and fur­ther ques­tions, please con­tact info@twillo.de.