news

BACK TO LIST

Generative models

29/03/2023

Generative models are the basis of machine learning computer programs that, after some training and generally a further refinement, can generate new content, such as text or images, from an initial input, usually provided by the user, either text or an image.

 

StableDiffusion, Midjourney, DALL-E for image generation, or ChatGPT, LLaMA, Bard for text generation are some examples of popular generative models and applications based on them that allow obtaining new contents related to an initial input. The statistical processes performed by these computer programs enable new contents to be obtained that are coherent with the input provided, and that are not simply a combination of the elements used for their training, incorporating a random component (generically called "temperature") that prevents the output from being deterministic, that is, always the same. Thus, a user interacting with these applications has the impression of dealing with an artificial intelligence that answers questions or follows instructions, although in some cases inaccuracies, such as imprecisions, falsehoods, or hallucinations can arise.

 

Given the rapid evolution in the quality of the models, it is to be expected that these errors will be solved and, eventually, it will be impossible to distinguish content generated by a generative model from content generated by a human being.

 

Today, being able to use some of these generative models is only a few clicks away, but it is very likely that we will find them more and more embedded in our daily lives, even without being aware of it.

 

The novel ease of content creation by means of these generative models is causing the different administrative bodies to react quickly. Regarding the authorship of creations using generative models, the U.S. Copyright Office decision of 02/21/2023 in relation to the work "Zarya of the Dawn" is noteworthy, in which, although protection was initially granted for the whole work, protection was finally denied for the images because they were generated using a generative model -Midjourney-. This has prompted the U.S. Copyright Office to issue on 16/03/2023 guidelines for works that include material generated through the use of artificial intelligence technology.  

 

Whereas so far one of the main interests related to generative models was about their patentability, including their training and use as computer-implemented procedures, nowadays issues derived from their use, such as the authorship of the generated contents, are also gaining relevance. 

Xavier Montalvo

Telecommunication Engineer.

European Patent Attorney (EPO).

Senior Associate.