“I’d created 2000 free-text responses and labelled them ‘UK’. Then I copied and pasted the exact same 2000 responses but labelled these ‘US’. Finally, I combined them to create a dataset of 4000 total responses, and jumbled them up.
Despite the responses being identical for the UK and US, Copilot produced a rich, detailed summary of how US and UK respondents differed.”
https://kucharski.substack.com/p/real-signals-or-artificial-stereotypes
@gregeganSF @sinalana.eurosky.social What this basically demonstrates is that AI is a Bayesian filter, taking data in and using it to update its already massive database.
Actually, the updating is not really live, the AI /only/ uses its existing database to produce the results based on (prompted by) the new data. This application of it is just inappropriate.
The problem here is people not understanding what AI means as a technical term.
@khleedril
What part of the "application" was inappropriate?
To give this task to the LLM at all? But that's what they're advertised and pushed for.
To give it (nearly or completely) identical datasets?
In a real situation, you wouldn't know that has happened.
Also, how is it any better if the LLM invisibly skews datasets that are not identical to begin with? The result is wrong nonetheless.
@khleedril
We’re in agreement here.
Another kind of task it’s pushed for that it’s incapable of doing.
@gregeganSF @sinalana.eurosky.social
Greg Egan
Khleedril
Landa 