Artificial intelligence causes interest and fear in many people. Some fear that AI could outwit humanity and threaten its existence, as shown in popular sci-fi films. However, such a scenario is extremely unlikely, because the current iterations of AI are still not self-aware, but only able to predict the text based on the studied data. Although it often looks like very believable .

In order to understand how AI works, you do not need to try to penetrate its psychology. It is much more efficient to analyze its output in terms of probabilities. This approach has already borne fruit: a group of scientists from the University of California at Berkeley was able to figure out more product information Open AIthan disclosed by the company itself.

The AI ​​decides what to output based on what it receives as input. According to the researchers, the way AI responds to user requests can be used to draw conclusions about what data it used for training. After some analysis of the data, the scientists came to the conclusion that modern generative models, in particular, ChatGPT, are indeed trained on copyrighted works. This has been discussed before, but now there is quite definite evidence for this.

The scientists found that the OpenAI models were fed mostly science fiction and fantasy. Experts gradually agree that it is better to use open data sets for training neural networks so that the final model is less biased in its answers. But also the ban on the use of copyrighted literature relates to the fact that chatbots, although not intentionally, can infringe copyright by copying the original parts of specific works in their responses.

If the original of a work is protected by copyright, you can do with it only what the copyright owner allows. But training a neural network creates a mathematical set of connections for the selected work, mixing it with other similar data. And even if the final result will not be very similar to the original work, can it be called unique, not infringing copyright? Maybe. However, copyright holders don’t think so.

On the other hand, modern LLM– models are quite creative, they learn and improve quickly, and all sorts of legal delays can only slow down their development. It is quite possible that over the past six months people have become so accustomed to the fact that in their lives there are such smart assistants as chatbots that capture the context of communication and can really speed up routine work, that humanity will rather choose the development of neural networks, giving up copyright. After all, they are the future.