Reducing hallucinations

AI model developers use a variety of ways to reduce hallucinations in responses (including images and videos):

• More and better training data
• Models that combine different 'modules' — modules specialise in generating different parts of the response (e.g. hands and text), and the results are merged together by the model as part of the generation
• Models that integrate real-world physics and structures better (like models of human anatomy)

An example of where image models have made large improvements is in the generation of hands. A few years ago, AI image models would often struggle to generate realistic images of human hands. However, this has now been largely fixed. The image below shows an example with six people holding coffee mugs with both their hands, which all look realistic.