Based on the Turing test, what would be the criteria for an agent to be considered smart?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
If an artificial agent (AA) passes the (standard) Turing test (i.e. where you have to imitate a human that speaks), then, on average, the AA should be able to imitate any human in any situation that mainly requires the conversation abilities and common-sense knowledge of a human, without being ever recognized as an AA.
For example, if you want to talk about football, you don't expect the AA only to say "I don't know" or to clearly avoid a topic (e.g. by redirecting you to a search engine) when it doesn't know something (although some humans behave in this way), but you expect it to have common-sense knowledge, such as that Messi, Cristiano Ronaldo, Pelé, Maradona, etc., are among the best players of all time, and this should be precious information that the AA should have in any case, even if it doesn't know much about football.
You also expect it to be emotional and have a personality, given that humans are emotional and have personalities. So, in the example above, maybe the AA could say that Maradona is its favorite player, and then it could explain why in an emotional way (e.g. by changing the tone of the voice).
The AA should also be able to keep track of (almost) everything you and it said, and it should be able to contextualize very well, as humans do. Some personal assistants already take context into account, but they don't do this very well or just do it to a little extent.
The AA should also be able to reason given the current situation. For example, if you explain something to the AA, you expect it to infer or predict something based on the information it has acquired and the common-sense knowledge.
Moreover, when you speak or write to the AA, you don't expect it to regularly hear badly or not understand what you say or ask (and ask you to repeat), but you expect it to understand well what you say almost always, provided you don't talk or write trash. You also don't expect big delays and you expect the AA to at least say something while it searches for a more appropriate answer, although not all humans behave in this way, but I think that interjections or words such as "hm", "well", "let me think...", etc., will be very important to make the AA look or sound like a human. In general, the AA should be as interactive as a human.
These are some traits that the AA absolutely needs to have in order to pass the Turing test (and be considered intelligent according to that test), but there are probably many others.