All else equal, AIs are slightly more likely to generate true facts than false ones
AIs are much more likely to generate false facts than true ones.
For example, AI-generated references to sources are almost invariably non-existent. Case in point: those LinkedIn links.
Bret Devereaux had a post about using ChatGPT to write an essay on a historical subject. The result looked superficially like an essay on the historical subject, but it was useless.
Given the likely genesis of the Sol Voltaics web page, reading it should produce no update about the energy cost of making a solar panel.
All else equal, AIs are slightly more likely to generate true facts than false ones
AIs are much more likely to generate false facts than true ones.
I think there might be some English ambiguity here. Suppose there are 81 names that the AI might come up with when writing the sentence “$NAME invented plastic”, and maybe the correct name has a 20% chance of being picked, and each of the incorrect names has a 1% chance of being picked. Then it’s simultaneously true that:
The correct name is 20x as likely to be picked as any individual incorrect name.
It is much more likely that the name picked will be incorrect than that it will be correct.
AIs are much more likely to generate false facts than true ones.
For example, AI-generated references to sources are almost invariably non-existent. Case in point: those LinkedIn links.
Bret Devereaux had a post about using ChatGPT to write an essay on a historical subject. The result looked superficially like an essay on the historical subject, but it was useless.
Given the likely genesis of the Sol Voltaics web page, reading it should produce no update about the energy cost of making a solar panel.
I think there might be some English ambiguity here. Suppose there are 81 names that the AI might come up with when writing the sentence “$NAME invented plastic”, and maybe the correct name has a 20% chance of being picked, and each of the incorrect names has a 1% chance of being picked. Then it’s simultaneously true that:
The correct name is 20x as likely to be picked as any individual incorrect name.
It is much more likely that the name picked will be incorrect than that it will be correct.
Are you taking into account the simulated exams? It doesn’t look like it mostly generates false facts?