IMO, it's OK for AI vendors to provide some censorship of the AI responses, as long as they're clearly stating that the response was censored, and provide an Opt-out of censorship option.
It will be also useful to clearly state what exactly is censored, and which bias that censorship has. Plus the ethical guidelines used for this censorship.
In case the opt-out isn't technically possible because of the way the model was built, the Censorship disclose should be provided with every response.
Likewise, when a user shares the vendor's AI responses publicly, the censorship (including the absence of it) disclosure should be clearly shown.
Most people are capable of logical inference just fine if they have all the relevant facts readily available. LLMs are useful because they can potentially recall a much larger set of relevant facts much faster than you could look them up on the internet, albeit at some reduction in accuracy.
7 comments
[ 3.0 ms ] story [ 33.0 ms ] threadIt will be also useful to clearly state what exactly is censored, and which bias that censorship has. Plus the ethical guidelines used for this censorship.
In case the opt-out isn't technically possible because of the way the model was built, the Censorship disclose should be provided with every response.
Likewise, when a user shares the vendor's AI responses publicly, the censorship (including the absence of it) disclosure should be clearly shown.
Wouldn't it be much more useful to concentrate on AI's intelligence: its inference capability?
(hmm, how precise is a collection of phrases like "red sky at night, sailor's delight?")