The use of generative AI in education and beyond remains controversial, particularly regarding authenticity and academic integrity. Some universities already have policies requiring the disclosure of all AI use, but without a control mechanism, one must hope for the honesty of the students.
The AI detectors sometimes do not work reliably (AI Detection Tools: When You Turn It In It’s Too Late!). Generative AI market leader OpenAI initially developed its own tool for analyzing the origin of text, but soon discontinued it due to a lack of accuracy (New AI classifier for indicating AI-written text).

It is therefore not surprising that attempts have been made to give the AI-generated texts a “watermark” for easy recognition and classification. The language model integrates a small bias in word selection (or technically correct “token selection”), i.e. the use of certain words to introduce a statistically specific, recognizable pattern to the text (A Watermark for Large Language Models). As a result, the text with a “watermark” will differ slightly from that without a watermark, as the words are not selected entirely freely – the quality of the text will generally be somewhat monotonous and rather poorer as a result.

OpenAI recently reported that this process works with a high recognition rate (Understanding the source of what we see and hear online). However, the same article also explains why this procedure will not be used on a regular basis. There are three ways to change the watermark pattern and thus prevent classification:

• Translate the text into another language and back again using translation software (a simple trick that fraudsters have already used successfully with Copy&Paste plagiarism).
• Have the text reworded by another language model.
• The language model itself can be ‘tricked’ when generating the text by instructing it via a prompt to insert specific words or characters between each word and then remove them afterwards with a simple ‘search & replace’. Therefore, bypassing the watermarking process is “trivial for malicious actors”.

Apart from that, a unilateral introduction of OpenAI would probably be seen as a competitive disadvantage compared to other language models such as Google’s Gemini or Anthropic’s Claude. After all, texts outside the academic context could then also be classified as AI-generated without the knowledge of the creators – which would possibly be interpreted to their disadvantage.
Furthermore, it is also not possible to differentiate between the type of use – for example, whether a text was developed and written by the author himself with arguments and ideas, and AI only provided the linguistic finishing touches at the end, or whether it comes from AI completely without a basis. This could put non-native speakers at a disadvantage, for example, if it were then assumed that there was no personal contribution.

Therefore, documenting the research and writing process still seems to be a much more reliable way of transparently and fairly assessing the extent to which the author contributed to the creation of the text.