FAQ GenAI

The system is not 'intelligent¡¯ in the human sense, but it has been shown to perform well in a number of standardised tests.

For example, OpenAI has tested the first version of GPT-4 on a variety of external page Benchmarks. Or there is an attempt to create an external page Rating of LLMs, on the capabilities of the various models. Experts agree that the external page capabilities will continue to increase in order to achieve good results in various benchmark tests.

The models are trained using large amounts of data. Each model available in this way has a last status on which it was trained. After this, training is usually no longer carried out, but the content is enriched with current data.

For example, after a user inputs information in the form of a question, an internet serach is triggered. The results obtained are fed back into the LLM (lare language model) as an additional prompt. Or after an input, a comparison is made with currently stored data before this is passed on to the LLM as a whole.

It is important to emphasise that although the models do not continue to train, any data that is passed on to an LLM can be used for future training purposes unless this is explicitly excluded in the terms of use.

The data consists of publicly available data sources, but certainly also content from the internet. The companies do not currently disclose exactly which data was used for training. However, it is known that the amount of data required is so huge that as many data sources as possible are used.

In addition, unless explicitly restricted, all current access data (requests, uploaded files) can be used for future training sessions. Caution is required here and attention must be paid to the terms of use.

Technically and legally speaking: no. Plagiarism is the unrecognised copying or use of another person's intellectual property. As long as artificial intelligence is not assigned a personality, using this tool is technically no more plagiarism than doing a Google Scholar search, using Grammarly or DeepL, or using R or Python to perform large calculations.

However, the outputs of generative AI need to be scrutinised very closely for possible plagiarism. In addition, depending on the use of GenAI, the usage should be correctly cited.

No, not reliably. GenAI answers are the result of a probabilistic algorithm and are therefore generally not reproducible. The answers are not simply fragments of the text corpus used for training and therefore the answer passages cannot be found using Google.

In any case, it is not allowed at ETH Zurich to assign grades completely automatically; AI could help with scoring, but the grade must always be assigned by a human. GenAI tools can create plausible-sounding grades and justifications for given rubrics and grading scales. It is recommended that the chosen procedure is always discussed with the students and disclosed at all times.