A company is developing an AI that is supposed to discover new, innovative ideas in a large archive of texts.
The AI is meant to create creative connections and suggest unusual solutions to promote research and development.
At first, the AI delivers many proposals, but soon it becomes apparent: Most ideas are very conventional and repeatedly based on the same known concepts.
Unusual or truly novel approaches hardly appear.
The company wonders: Why does the AI hardly generate creative or original ideas, but always stays with obvious, familiar patterns?
Question: Why does an AI trained on large amounts of text tend to generate mainly obvious and familiar ideas instead of truly innovative or surprising suggestions?
Solution follows tomorrow.
Solution
The AI was trained on existing texts that predominantly contain known, established ideas and patterns.
It learns to recognize connections and probabilities based on frequent combinations and existing concepts.
Since truly new, creative ideas are rare and statistically unusual, the AI rates them as less likely or less fitting.
The training objective to generate the most plausible and best explainable proposals leads the AI to prefer mainly familiar, obvious links.
Innovations that strongly deviate from existing patterns are often not recognized as meaningful by the AI and therefore hardly generated.
Result: An AI based on existing data can have difficulty showing genuine creativity because it is optimized for probabilities and the known. For creative applications, special strategies and training methods are therefore necessary that specifically promote and reward unusual combinations.