Dear Felienne!

You won’t believe it. This time, I actually did the homework you assigned instead of going off course. I am a bit scared to comment on this Turing test paper, though! First, because I don’t want to repeat Alan Turing’s mistakes and say all kinds of things without sufficiently engaging with other domains. And second, because I make my points with a story about Alice and Bob having great sex. I am not sure the Computer Science community (who probably only knows Alice and Bob from cryptography lessons where they send each other rather innocent encrypted text messages!) is ready for that. But ok, here we go…

Fake love, great sex and big tech

My biggest concern with the “Computing Machinery and Intelligence” paper is its assumption that imitating intelligent behaviour is essentially the same as truly behaving intelligently if the two cannot be told apart. But this is crazy…

Let’s just for a second try out this line of reasoning with emotions rather than intelligence. Assume Alice is rich and Alice loves Bob. Bob does not love Alice, but he would really like to inherit her money. So he imitates loving Alice. He is a master manipulator and does all the right things! So during their entire life together, Alice never has any suspicion that something might be off. She believes Bob loves her and lives a happy life. Then she dies, and Bob inherits all her money. Does it matter that Bob never loved her and was just pretending?

Of course it matters!!! And to me, it matters not only for emotions but also for intelligence. Let me spell it out: In this metaphor, Bob is Big Tech and all the AI companies. They are just in it for the money. While we cannot know for certain whether their products can think, there is a very high chance they are faking it for the money.

Of course, there are many smart people who’ll now say: ”I know AI is not really thinking, and I always double-check all the answers for hallucinations. I know big tech is bad and all that. They are not fooling me. I am using it responsibly and I still benefit from it.”. That might just as well be true… In the Alice and Bob story, this would mean Alice is aware that Bob is faking his feelings, but Alice decides to stay in the relationship nonetheless, for the great sex. That’s Alice’s choice to make, of course. I am not here to tell anyone how to live their life, or that they should not use AI. But I’d say: just keep in mind that while machines may not be able to think, they definitely are able to manipulate. There is no doubt about that!

I guess most readers will agree that Bob’s pretending to love Alice to exploit her is unethical. (I hope so!) And the same point applies to AI: Machines pretending to be human, to be thinking, or to be feeling something they are not really feeling for some company’s benefit is fundamentally wrong! And while I personally can’t believe the sex with Bob is all that great in the first place, I respect Alice’s choice to stay in the relationship. We just should not call this love. Likewise, I get that people are using genAI, but we really should not call AI’s behaviour thinking.

One more detail about the word “thinking”

I have one more thing about the word “thinking”. A small prediction by Turing about how this word will be used in the future stuck out to me. So he writes:

Nevertheless I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.

I think we are at a turning point here. For some people, speaking of machines that way has become normal, while others still are bothered by this. In fact, my partner came home from work recently, completely worked up and in disbelief about something a colleague had asked on Slack about an online search that did not provide the expected results. Specifically, this colleague had asked: “Did you enable thinking?”…

What has happened? As you and our readers probably have guessed, this was around the time when Google Gemini had introduced a new feature to its AI mode, called “thinking”. People have subsequently started using this term as if it were normal to speak of machines thinking. Hence, Turing was right that people are talking about machines thinking, and they are probably not expecting to be contradicted — but I am glad some of us are still contradicting it! (Of course, it is not just Gemini; other models also introduced so-called Reasoning, Extended Thinking, or ‘think flags’ that one can set).

Autonomy as a necessary condition for intelligence

I have another, more fundamental problem with the paper and its view of thinking! For me personally, thinking can only be intelligent if it is also autonomous. So I see autonomy as a necessary condition for intelligence. This is, for instance, the idea behind how we evaluate our PhD candidates. We want to know if they are autonomous researchers and if their thesis is the result of their thinking.

Of course, I certainly don’t want to pretend that I know what autonomy means. This is an entire key concept in itself, where we need philosophers. What I mean here, simply put, is that for someone or something to be intelligent, the smart thoughts they express also need to be their own; they need to be authentic, autonomous thoughts. And that is where it goes wrong for me when we look at Turning’s paper. It is very much focused on the ‘smart part’ and does not sufficiently consider the necessity of autonomy.

This is also where it goes wrong when we look at current big tech LLMs. The output might appear intelligent, but it is not the result of an autonomous thought process. I actually asked an LLM whether it agrees, and it confirmed: ”No, Large Language Model (LLM) output is not inherently authentic” and ”LLM output is not the result of autonomous thought”. No surprise here!

I believe we are allowed to do this little trick that Turing also does in his paper, where he just replaces one question with another question. So I want to replace the question ”Can machiens think?” with the new question ”Is machine output authentic?”. And in this case, I agree with the LLMs above: No, it is not!

Imitating human looks

A detail I found fascinating about the paper is that imitating human looks is not interesting to Turing at all. Read for yourself:

No engineer or chemist claims to be able to produce a material which is indistinguishable from the human skin. It is possible that at some time this might be done, but even supposing this invention available we should feel there was little point in trying to make a “thinking machine” more human by dressing it up in such artificial flesh.

This is actually something the computer science field also looked into, even though Turing said there is little point in doing so. In this regard, I just want to point out to our listeners/readers Hiroshi Ishiguro’s robots. He is a Japanese roboticist who has been creating robot clones of himself that look like him, and they were, for example, used to explore questions around transmitting human presence to remote locations.

Another thought experiment

Like Turing, I also have a thought experiment for you. Let’s, just for a brief moment, go along with Turing and accept that there is indeed this weird party game where you have an interrogator talking to a man and a woman, trying to find out who is who while the man pretends to be a woman. And then, like Turing, you want to replace one of these people with a machine. I just wonder why you would replace the man in this scenario? Why not replace the interrogator instead? I would assume that to do well in this game, the interrogator must be somewhat intelligent. At the very least, they must ask sensible questions, draw sound conclusions, understand the strategies that the man might apply, and so on. So even if this is a stupid game, the interrogator needs to be somewhat smart. Hence, I wonder if it occurred to Alan Turing to replace the interrogator in this scenario with the machine. What would have happened then to the rest of the paper and to our field of computer science?

Actually, if we look at how AI is used nowadays, I believe it often does play the role of the interrogator! Decision-support systems are used to decide “Is this cancer?” or “Is this a good hiring candidate?”, or “Is this piece of information true or fake news?”! Metaphorically speaking, the AI decides who is the woman in Turing’s weird party game!

Something we can ask ourselves in this age of AI, thus, is: aren’t we increasingly playing this weird party game where both the interrogator and the other player are replaced by machines? And maybe more importantly, why are we still playing the game, and to whose benefit and amusement are we playing it? I fear we are playing at the amusement of the tech bros; we are basically the jokes at their parties.

Celebrating Alan Turing

And to wrap this up, a final question: “Can we still celebrate Alan Turing?”… I have been thinking about that and the fact that people forgive LLMs for being wrong and saying stupid stuff all the time. There was this BBC study that found 45% of AI results to have errors. When I asked an LLM how often it is wrong, it admitted to being wrong 10% of the time (and, based on this answer, we should take that number with a grain of salt!). But people just accept such error rates from LLMs. So I think we can certainly also be forgiving with humans and with Alan Turing. He did a lot of super weird things in this paper, and some things I absolutely do not endorse. But he was human, and I think he did infinitely better than any machine could have done. So I think we can still celebrate him, just preferably not for this particular paper.

Cheers,

Hanna