5 ms·
"The Turing Test" is a very clever way to define intelligence. BUT... the glory of the goal of creating a machine to pass "The Turing Test" which we have corpo
by gregw2 4y ago
"The Turing Test" is a very clever way to define intelligence.
BUT... the glory of the goal of creating a machine to pass "The Turing Test" which we have corporately tacitly adopted is subtly deeply flawed, tragic, and not well thought through.
To get a computer to pass the Turing Test, we have to teach computers to deceive humans.
Think about that!
Deception is fundamental to our current goals!
How can that possibly go wrong?
How can that possibly be misused for profit easily once built?
If Satan is the "father of lies", what does that make us, make the creator of something that passes the Turing Test?!
So...
How could we reconceptualize our goal of enabling intelligence without putting deception (The Turing Test) at the heart of it?
Sure, deception is a sign of intelligence but does it need to be a requirement?
If we really want to avoid the perils of the singularity, we should do some hard thinking about modifying the type of Test we want our computers (and our AI researchers) to pass.
Or we will get just the deception we asked for, and recognize it too late.
Your ideas are hearby solicited.
- gregw2 4y agoI'm an amateur, but let me offer some possibilities... How about a "Socratic Test" where the computer teaches humans by asking them questions that guide the human to truth? If the human conversing with the bot/AI claims improved enlightenment at higher rates than if a human were asking questions, the AI passes the "Socratic Test". (Variations on the theme are if course possible... an analogous "Zen Master Test" would up the "cool" factor (and difficulty!) but the rate of enlightenment experiences might be too low with either humans or AIs to make a good measurable criteria.) Or how about a "Jesus test"? If a human conversing with the AI is convinced by the truth and goodness and humility from the AI enough to say "surely you are the Son of God!", we have created something worthwhile and benevolent. (Ok, this has the "zen master" problem too and is problematic for other reasons, but perhaps you get my point.) What about the "Teacher/Larry Page Test"? (Larry Page, Google cofounder was known for asking the interview question "teach me something!" to get insight into job candidate's expertise/thinking... and at least some value from an otherwise unfruitful interview ). If in a conversation between human1 and an AI-or-human2, if human1 asks "teach me about X" questions, where X can be an unbounded set of questions, and rates the answers as "likely to be from someone of intelligence they would want to talk with/hire/etc" more often with the AI than with human2, isn't that sufficient to form an alternate "Turing Test" but without the deception-based goal? Alternative suggestions welcome.
- NickRandom 4y agoHow about a mandate that any and all AI's self-identify as such if directly asked? Kind of like a 1st Law of AI