I'm drawn to note 5 "Truth is often important, but it can be subjective, and it is not always a requirement for an AI system (e.g. one intended to produce creative writing or art). Where relevant, truth can be a defined requirement of an AI system, to a specified degree of certainty (e.g. in terms of bias and variance)." There is so much to unpack there!
Agreed! Truth is so often a matter of politics or faith (particularly at this challenging moment). My goal here is to be technocratic, and propose an approach that can 'work' regardless of such issues. Perhaps a topic for another blog!
truth v Truth v facts? Then creative writing or art - as an expression of what I wonder? A feeling about something? So much embedded in language - or perhaps I'm too simplistic. 😄
To take a political example, a core problem (perhaps the main one) with the deportation of Kilmar Abrego Garcia is that he did not receive due process. Failure of due process is the basis of the court decisions against the Trump Administration. The right process would address the competing claims of truth, and hopefully generate an outcome that most can trust.
Likewise, for AI, my point is that trust does not always have to involve truth. For an AI that generates creative stories or art, what is truth and is it relevant at all? We do care, however, about process -- e.g. whether the AI has without permission copied the works of human artists.
In other cases, truth may indeed be relevant and/or crucial for AI, and we should specify that as a requirement.
Yes, ideally. It is a gaping hole if there are no agreed processes, to ferret out facts, that people trust. In AI with "creative" works idea, I'm thinking of a different question. At least for me, it feels odd to describe the work of AI as creative when the mechanism itself can't distinguish truth from fantasy. But again, that's a different question. I'm appreciating this sensible approach you're highlighting and discussing re: building a process framework with necessary grounded flexibility.
Another technical initiative related to trust in AI: https://arxiv.org/abs/2504.15499. About using hypervisors to restrict potential malicious AIs.
I'm drawn to note 5 "Truth is often important, but it can be subjective, and it is not always a requirement for an AI system (e.g. one intended to produce creative writing or art). Where relevant, truth can be a defined requirement of an AI system, to a specified degree of certainty (e.g. in terms of bias and variance)." There is so much to unpack there!
Agreed! Truth is so often a matter of politics or faith (particularly at this challenging moment). My goal here is to be technocratic, and propose an approach that can 'work' regardless of such issues. Perhaps a topic for another blog!
truth v Truth v facts? Then creative writing or art - as an expression of what I wonder? A feeling about something? So much embedded in language - or perhaps I'm too simplistic. 😄
Not exactly. Truth v process.
To take a political example, a core problem (perhaps the main one) with the deportation of Kilmar Abrego Garcia is that he did not receive due process. Failure of due process is the basis of the court decisions against the Trump Administration. The right process would address the competing claims of truth, and hopefully generate an outcome that most can trust.
Likewise, for AI, my point is that trust does not always have to involve truth. For an AI that generates creative stories or art, what is truth and is it relevant at all? We do care, however, about process -- e.g. whether the AI has without permission copied the works of human artists.
In other cases, truth may indeed be relevant and/or crucial for AI, and we should specify that as a requirement.
Yes, ideally. It is a gaping hole if there are no agreed processes, to ferret out facts, that people trust. In AI with "creative" works idea, I'm thinking of a different question. At least for me, it feels odd to describe the work of AI as creative when the mechanism itself can't distinguish truth from fantasy. But again, that's a different question. I'm appreciating this sensible approach you're highlighting and discussing re: building a process framework with necessary grounded flexibility.