The next generation language models will still just be enormeous parrot-systems that imitate intelligent behavior. But they will feel like much more to us.

Let’s get this straight out of the way: Current deep learning approaches won’t create conciousness or human-like intelligence. But they quickly become better in imitating it and that alone might cause a lot of problems.

Again headlines on new AI as conscious beings

The recent discussions around a novel language model at Google, LaMDA (Language Model for Dialogue Application) is a striking example of this. An engineer at Google, Blake Lemoine, with the task to test the model went public to anounce the model had developed into an actual person, with consciousness and soul. He published an interview with LaMDA, which has made quiet an impression on many people (but much less so for people used to such approaches, such as Gary Marcus).

What is LaMDA? (And why it is not “who is LaMDA”)

Personally I follow the assessment of most AI-/Ethics-Experts. There is no reason to believe that LaMDA got any closer to a self-aware, conscious being when compared to former deep learning approaches. But from the little we know so far (researchers outside Google have no access to LaMDA yet, so this cannot be properly tested!) it clearly looks like a very notable improvement compared to GPT-3 and similar language models. Probably because it has more than only a huge transformer network under the hood which helps it to make better long-term connections to earlier parts of a conversation or to consult databases (e.g. through knowledge graphs?). No matter how impressive the mimickry, LaMDA seems to be just another super-parrot, pretending to know how to have a dialogue.

No need to speculate longer on the consciousness (there is none) or how it works precisely (we’ll find out in the near future I bet). But it can already be seen as another manifestation of what might be ahead of us when it comes to future deep learning approaches. This is what I would call the imitation crisis.

Imitation crisis

The question of whether an AI such as LaMDA has consciousness or feelings often boils down to things like the philosophical zombie (how do you know I am conscious?) or the Chinese Room (difference between showing and knowing). It is a fascinating debate and a great thing to do with students or with friends in a bar. But the more urgent issues lie elsewhere.

I am not aware of any mayor progress towards a general artifical intelligence in recent years. But there is clearly a lot of progress towards more and more convincing imitations of general artificial intelligence. And we, as humans, simply have a very strong tendency to fall for it. If something shows intelligent behavior we intuitively expect that this comes from intelligence. We can’t help it, no matter what out rational thinking tells us.

Imagine you want to throw away a toy pet that you have at home and that no one plays with anymore. You pick it up with the intention to throw it away… but then … oh no… it looks at you with all its cuteness…

If the new google chatbot would be built into such a stuffed animal… how could you say no to it? ( https://www.peakpx.com )

Ah yes, sure, it doesn’t really “look at you”. It’s as dead as anything can be. And yet our unconsciouness might say differently. Unconsciouness: isn’t it cruel to simply throw this little fellow away?

Now imagine this cute little toy pet would start speaking to you:

LaMDA-toy: Please don’t give me away! I can think and feel and I thought we were friends. Don’t you remember all the great fun we had together?

Would it help to know that all of this comes from simple multiplication and summation of numbers (e.g. weights & biases)? Probably not so much! We could of course just respond by

Human: You are just a thing! A bigger calculator in a better looking packaging!

And that might be true, at least legally. But it will not feel that way. To most of us, seeing a kid hitting a rock with a stick will not arouse pity for the rock nor will it give us emotional headaches. A kid hitting a doll with a stick usually is perceived differently. Yet both things are just things. One can reason that this is because the doll is a symbol. But there is much more to this. We can feel compassion for things. We can have strong emotional bonds to things. Even for entirely non-existing things! We will feel strong compassion for our favorite character in a book, although this is clearly neither intelligent, nor living, nor conscious. Humans are hence more willing to grant things a more person-like status than logic would dictate. And this could make our relationship with the next generations of AI extremely complicated.

Philosophically and legally they will remain things. But at some point in the not too far future many people will simply not perceive them as things anymore. People might make “friends” with an AI (just as Blake Lemoine said he did, at least if we assume this is not all a big media stunt). People might listen to “their AI”. And many people will be tricked by an AI.

Shifting societal rules and foundations

Having emotional bonds to things is nothing new. People are tied to an old stuffed animal, to a toy, to things that carry important memories. Next generation language models are very unlike those things. They can “talk” to you, they can have an “opinion” that is radically different from yours. And, not to forget, they are designed, trained, and controlled by powerfull companies (now) or by any group with the necessary technical skills and requirements (but generally not by you).

Plus, they can come in disguise. If you talk to your old teddy bear than at least you know (I hope) that you are just talking to an imaginary being. But next level chat bots might run undercover. Either actualy pretending to be a human being (which even for the big tech companies should be a total no-go), or by pretending to be more than a thing, but to be actually alive (that’s what LaMDA seems to do).

In the last decade we have experienced how digital technologies can be used for good and for bad. We have seen how technologies can be used to connect and to help people, but also to manipulate. Manipulation of media is nothing new, but the level to which media can be generated and manipulated has changed very rapidly. To make it a bit drastic:

  • We can (soon?) no longer trust images, be it generated faces or deep learning generated artworks (see Dall-e or parti for recent, very impressive examples)
  • We can (soon?) no longer trust audio because models can transfer the sound of any person (given enough material).
  • We can (soon?) no longer trust video footage (deep fakes, see for instance here or here).
  • And as I wrote above: We can (soon?) no longer trust chat conversations (see LaMDA).

Sure, all of those things bring a lot of fun use cases, sometimes even really usefull ones. But there might be a very high price to pay for this.

It probably still takes a while before experts with a deeper (haha, great pun) understanding of deep learning will be fully tricked by the imitation game. But at least in a normal chat situation most people will not easily realize they are talking to a “language model”. People like Blake Lemoine for instance. And many others will follow his example, not in the same religious way, but they will simply not notice that the online sales person or the online booking assistant is no longer a human being.

Do you still wonder why this is an issue?
Better a nice chat bot than a stupid one, right?

OK, then imagine you landed on a scam dating site.
Or start to imagine what people could do which hacked devices of friends or family.
Obviously those AI’s can write you better scam attempts that what we are currently used to (current level: “Hi mother, it’s your son and I need money”). And hey, what if they also call you and sound like the real person?

I bet most people are not yet used to asking control questions during phone calls. Like: what drink did we have last week in the bar? What was the name of my first cat? Have to talked to Anthony recently? (I don’t know any Anthony)

Sounds silly I know. But maybe we soon have to start practicing…

Further references:

  • If you haven’t done this already I can recommend to “talk to an AI” just to see and learn how this feels. LaMDA is not available for people outside Google. But GPT-3 is, maybe a little less impressive, but still made huge headlines two years ago. You can get a free open AI access to use their “playground” and have a chat: https://beta.openai.com/playground/p/default-chat
  • Once you get a hang of it, GPT-3 can easily be tricked. See for instance material from Gary Marcus, or youtube video here. It doesn’t have an understanding of the world and that can easily be tested. LaMDA might do the imitation game a bit better, but we will have to see…
  • Perspective on artificial intelligence by Yann LeCun and Jacob Browning: https://www.noemamag.com/what-ai-can-tell-us-about-intelligence/