Why Alan Turing Wanted AI Agents to Make Mistakes | Hacker News Reader

Why Alan Turing Wanted AI Agents to Make Mistakes

spectrum.ieee.org

77 pointsheadalgorithm18 comments

I think what Turing is getting at here is actually less a commentary on what AI should do, than a shortcoming in the test. When you're looking for genuine intelligence you should be looking for intelligence. However, if your only sign of intelligence is human intelligence, suddenly getting a basic sum wrong is counter-intuitively a sign of intelligence.

In the end getting the sum wrong is an emergent property of human intelligence, we're good at things but we're not deterministically perfect like computers are. It's not better for a computer to start to share the shortcomings of humans, but it is more likely to pass the 'AI' test. Indicating a shortcoming in the test.

I tend to agree with the conclusion in the article that the Turing test wasn't really meant as a blueprint for general AI or whatever other (later) concept we're now familiar with but rather just a though experiment on whether it would even be possible to attribute intelligence to a computer. Or perhaps whether it would be possible not to attribute intelligence to it.

Or the more unnerving and less useful personality trait of lying. The Turing test is after all a game of deceit. Whether the human believes the AI to be human or the AI believes itself to be categorically human begins to blur the lines.

That is, to pass for a human, the AI should imitate humans. This includes imitating their mistakes. AI's genuine mistakes would likely be different enough to help humans tell an AI from a human.

That is, the best strategy for an AI to pass the test is to be perfect, and perfectly imitate a perfect amount of human mistakes.

This definitely requires an ability to lie, to knowingly tell untrue things.

This ability is also required by much simpler circumstances: when directly asked, an AI should be able to lie and say "No, I'm a human", else it will fail the test instantly.

This likely means that a robot following the Asimov's three laws would never pass a direct Turing test.

> This likely means that a robot following the Asimov's three laws would never pass a direct Turing test.

Not necessarily true. An AI could pretend that it doesn't follow Asimov's laws while in fact following them. For example, the AI could pretend that, to execute the given order (second law), it first needs to understand the author psychology. At that point the AI can change the author's mind by using persuasive tactics. If the AI successfully changed author's mind, it successfully didn't follow the initial order by simply cancelling it with consent of the author

The first and third laws are also applicable to humans, and cannot be used to detect the AI

I'm not sure it is a a shortcoming though. It's an entirely defensible proposition that under some circumstances, accuracy and reply speed is inversely correlated with intelligence.

Consider: you ask two colleagues a deep and subtle yes/no question. One immediately replies 'nah'. The other thinks for a long time and then finally replies "I don't think so, but I'm not totally sure..." Which of your colleagues has demonstrated greater intelligence?

Accuracy is not certainty. The one that's most informative is the one that's demonstrating greater intelligence.

I think there is a tendency to confuse human-like with intelligent mainly because humans are the only human level intelligences we encounter. This leads us to associate human-like behavior with intelligence.

For example take a look at the Terminator movies. As this xkcd illustrates (https://xkcd.com/652/), a real killer robot would probably look more like an autonomous flying drone than a humanoid.

One of the big reason we want AI is so we do not want mistakes. We do not want human foibles in our AI. There is no one trying to make self-driving cars that lose focus by texting or getting drunk just so it would seem more human. No one is trying to make AI medical systems that give incorrect diagnoses because they have been running for 30 hours.

As for me, I want AI to get around the limitations of being human, not to pretend to be human by having human like errors.

Humans would be useless at that point. While the machines probably won't start killing us indescriminatly, we'd be marginalized within at most years and self destruct as we wouldn't be able to afford living.

Thankfully, there is very little evidence that general purpose ai is possible at the moment, so I doubt I'll have to worry about that while I live

I'd say the idea that we need AGI for all the "problems" that come with AI is a problem itself.

Humans aren't AGI, they usually focus on one area of knowledge.

Myriad AI's each doing non generalized thinking but still all the things humans do today is the same problem.

And it is inevitable. It cannot be stopped. We do not need to achieve that which some believe is impossible for society to provide for itself without needing humans.

Not to quibble, but if you watch the films, SkyNet uses mostly drones and tanks. It only develops the Terminators later in order to infiltrate the human ranks.

Also, we absolutely want human foible in our AI agents. To use your driving example, the self driving car must be able to negotiate with a kid who wants to cross the street.

Right. The humanoid found Sarah Connor by looking her up in the phone book and making cold calls. How would a flying drone do it?

But in your driving example, is negotiating with a kid really a foible? Or is it a skill?

I believe Turing's argument strictly is that 'can imitate humans' -> possesses intelligence, not that all intelligent agents are capable of passing as humans, although this distinction does get muddled quite a bit.

(Note: Disregarding the possible negative outcomes of general AI on humanity for the sake of this comment)

While I completely agree with you that we don't want machines to make these human-like mistakes, wouldn't you agree that it would be ideal for a machine, when it makes mistakes, to do so in a human-like manner, rather than not?

One thing that appears to separate humans from AI is that, even as an AI may increasingly become better on average than humans at problems, games, etc., an AI seems to always occasionally make some absolutely terrible decision that we can hardly even understand the rationale behind. This is a huge flaw. In the rare instance an AI makes a mistake, it's completely out of its mind (at least to our knowledge) and off-base. When a human makes a mistake, it's one that's closer to the mark.

Is this not a valid reason to make AI more human-like in its mistakes, assuming it will have mistakes of some kind?

This is discussed in the introductory chapter of "Artificial Intelligence: A Modern Approach"[1]. There are four schools of thought about building intelligent agents:

1. Thinking Rationally: Use logic

2. Thinking Humanly: Solve problems like humans would

3. Acting Rationally: Attempt to maximize performance

4. Acting Humanly: Turing test

I agree with their conclusion that rational agents are what we should design. Humans after all try to be rational agents and we still fail to be rational a lot of the time.

[1]: http://aima.cs.berkeley.edu/

> an AI seems to always occasionally make some absolutely terrible decision that we can hardly even understand the rationale behind. This is a huge flaw. In the rare instance an AI makes a mistake, it's completely out of its mind (at least to our knowledge) and off-base.

Some of the game-playing bots have made moves which seem like garbage in the moment, only for a novel strategy to unfold in later moves. Is it possible to achieve the novelty without the true garbage?

So I guess it depends on the work being automated. AI cars should not have novel strategies, "just" follow the rules, while a strategic battle AI working with a human might be better with out-of-the-box thinking (and transparency of reasoning for the human to check).

If the AI agent were quick and precise to answer everything, wouldn't that by itself allow a judge to tell a computer apart from humans?

I believe that faithfully passing the Turing test requires the machines to copy human behaviors feature by feature, bug by bug. Likewise for the machines to demonstrate a good human language understanding.

I'm guessing perfect "AI" will simulate every possibility with the current input sources and to find the best solution. Perhaps similar as our life by the universe being a deterministic system and having slightly small edits compared to other universes functioning with keeping track of one another. What we call "mistakes or failures" are just a summation of events to witness a certain outcome. All the outcomes of all the other universes with maybe our own, hold one sacred outcome. Maybe you're experiencing it now or not.

Reply on news.ycombinator.com