Generative AI
From Wiki for Sustainable IT
Generative AI designates systems able to produce new content, text, images, sound, video or code, from patterns learned across a large body of examples.
How it works, and what follows
The model does not consult a knowledge base. It produces the most probable continuation given what it has learned. Two consequences follow directly from that mechanism.
Plausibility is not accuracy. Incorrect output can be perfectly well formed. Verification remains the user's responsibility.
There is no source. By construction the model cannot say where a statement comes from, unless paired with a document retrieval system.
Practical issues
- Confidentiality: what is typed in may be retained or reused by the provider (see Shadow AI) ;
- Rights: over training data and over generated output ;
- Transparency: the AI Act requires disclosure that content was generated ;
- Resources: each generation is an inference, whose cost varies widely with the size of the model used.
