Humans, in aggregate, behave consistently when presented with structured prompts. The trick with new technology is to figure out what presentation of the information results in the desired behavior of checking for AI blind spots. An example with Flock is that they don’t present partial matches any differently than complete matches - in that case the improvement of describing partial matches as incomplete is obvious, but other triggers for human common reactions are more challenging to tease out. But it can be done!
I believe if regulators set up a system to encourage research and implementation of responsible data presentation - meaning outputs that make it easy for humans to have appropriate trust and add checking as needed - that would have more positive and lasting impacts than a ban.
How about don’t present partial matches, demand more data first to raise certainty of the results. Especially partial matches against open ended data sets is wildly unsafe in almost every case (eg. not already having suspects).
Responsible presentation is sometimes not at all, if the user isn’t an expert.
Humans, in aggregate, behave consistently when presented with structured prompts. The trick with new technology is to figure out what presentation of the information results in the desired behavior of checking for AI blind spots. An example with Flock is that they don’t present partial matches any differently than complete matches - in that case the improvement of describing partial matches as incomplete is obvious, but other triggers for human common reactions are more challenging to tease out. But it can be done!
I believe if regulators set up a system to encourage research and implementation of responsible data presentation - meaning outputs that make it easy for humans to have appropriate trust and add checking as needed - that would have more positive and lasting impacts than a ban.
How about don’t present partial matches, demand more data first to raise certainty of the results. Especially partial matches against open ended data sets is wildly unsafe in almost every case (eg. not already having suspects).
Responsible presentation is sometimes not at all, if the user isn’t an expert.