A Turing Test for Life?

Hi Hiro, hi Keisuke!

There is a recent increase in interest in discussions about finding life in the universe. Netflix did a half-scifi, half documentary called Alien Worlds, the Seti Institute has launched a new competition with a $15,000 prize to find extraterrestrial signals…

But the goal of our field of research, Artificial Life, is to make new life from scratch, here on Earth. There is a big debate about how we would even decide whether that kind of artificial agent is really “Life” or not, and a question that all 3 of us are interested in is: would we even be able to tell the difference between common Earth-life and current artificial agents (which most people agree are not alive)?

Probably because there are so many different aspects of what we mean when we say "life," there is a big debate about how to determine whether it is life or not. Even if the way of judging is different, I think the common view among most researchers is that life cannot be artificially created yet. I think that each of us should set up their own criteria based on their interest and create something that we can say is life. Leaving aside the question of what the criteria are, I am optimistic that in order to break through to life on each criteria, we will inevitably break through to the other criteria as well. I believe that the reason why researchers from different fields can gather and enjoy discussions at the international conference on artificial life is because everyone believes this implicitly or explicitly, too.

Indeed, it is difficult to establish a single criteria that distinguishes the living from non-living systems. I feel that there are two major parties here; there are those who believe observer-free (e.g. information theoretic) criteria to define agency, and there are those who believe we need a kind of Turing test because we cannot exclude observers from the loop to define agency. In addition, the recent advances in GAN (Generative Adversarial Networks) point in a new direction: synthesising artificial "discriminators" to distinguish between living and non-living systems, instead of using human observers.

Lana, you and Olaf ran the “Fake Life” competition. Can you describe what strategy the prize winner took? Do you guys think it's possible or not to create a universal life detector using machine learning algorithms in principle?

I agree with Hiro on the fact that finding a definition of life is difficult, but I think it’s due to how we approach the problem. I think definitions only make sense when they help us do something. For example, we’re not sure how we would use a “conscious” robot vs an unconscious one (assuming that conscious does not just mean “very intelligent”), and lacking a clear function we struggle to agree on a definition. Maybe we can’t agree on a global definition of life because we can’t think of a global function for “life”. As Keisuke points out, for any single “function” (criteria) of life, we have already created or simulated something that fulfills that function: self-replication, cognition, etc… My feeling is that we are still unsatisfied, but we don’t know why. I’m not sure if I agree that agency = life. For example Keisuke, do you think that a vacuuming robot has, or does not have agency?

In the Fake Life Recognition Contest, the participants examined the normalised trajectories of living and artificial systems. The winners of the contest obtained very high scored by assuming that real living systems changed trajectories at a rate that can be approximated “well enough by not perfectly” by a power law. Basically, they found that living organisms that are moving in their environment change direction relatively often, and rarely move in straight lines for a long time; but systems that follow this pattern too exactly, or don’t follow the pattern at all, are likely to be artificial. Interestingly, the winners misclassified drones as “alive” ⅔ of the time, maybe because drones are subject to variations in the wind direction…

I must add that a recent paper (“Scale-free networks are rare”) undermines the popular idea that life is associated with power laws...

Yes, we are not satisfied with something just having those functions such as self-replication and cognition. The reason for this is that, as Keisuken has pointed out, the criteria for function are functions that are required only for the individual, and are not interactive. I think that the functions required for life as an individual are necessary but not sufficient. We have not been able to create something that exploits those functions as individuals to realize the functions (or achieve criteria) involved in interacting with people (or other life). The interaction based criteria has to be something like a Turing test in which an artificial agent (we create) makes people with whom it is interacting feel that it is life. The original Turing test was based on verbal interaction, but there is no need to limit it to verbal interaction; physical interaction is more likely to reveal primitive life. Fake life competition is a question of how an observer judges life, not the function required as an individual, and the behavioral turing test is that the observers can interact with the object and ask how they judge. To add a little more I think the relationship between the objects and observers will be more symmetrical, rather than asymmetrical as in the original turing test.

Right, unlike the classical attempts at defining life, the Turing test for AI is meant to be interactive. But I don’t know if I am satisfied with a purely observator-based criteria. For example, one of the programs to successfully pass the Turing test was a basic chatbot programmed to say that it was a non-native English speaker, a 14 year-old boy. The flaws of the chatbot were accepted by the human judge only because of the prejudices or low expectations of adult native speakers towards a foreign adolescent. I don’t feel happy with that program as a Turing test winner. How do we make it harder to fool a judge in a Turing test for artificial life? We don’t want it to end by saying “This is life from Venus, that’s why it’s a bit wonky” and the judge just accepting this excuse. Maybe we can avoid this by saying that the artificial agent must be positively judged by more than one species. My dog is good at differentiating life from inanimate objects, and he doesn't care about excuses such as “this is from Venus!”

Maybe the first question is, what kind of interaction should the test be based on? Hiro, you work with fish. What type of interaction do you think we should use as a base for a fish-judge?

I kind of agree with Lana that observer-based criteria have the drawback that judges can set the bar too low in some cases! Indeed, using animals as judges would be one way to avoid this problem. Maybe, another way is to use an artificial intelligence as a judge like Generative Adversarial Network scheme. This way to discriminate between humans and artificial bots is already used in Google CAPTCHA ( “Completely Automated Public Turing test to tell Computers and Humans Apart"). You may provide data to Google everyday by choosing “fire hydrant” images discriminating from other images. I wonder if in this way we can train an algorithm to perfectly discriminate living and non-living systems. Also it is interesting that Hiro pointed out that the relationship between the judge and the object should be symmetric. To become a good judge, it might need to behave like an actual agent. This is an implication I got from Hiro’s perceptual crossing experiments (Perceptual Crossing: Two people interact with each other with finger movements in one dimensional space where a tactile feedback is provided only when their positions cross each other. They need to judge if the opponent is human or not). I believe the idea behind the perceptual crossing is also used in his fish experiment.

Going back to the self-replication/self-maintenance capability of life, I am still curious whether those fundamental properties of life can be related to those interaction-based criteria of agency to some degree. I don’t have the answer so far, but I believe that the autopoietic/homeostatic activity can be a source of intentional and objective behaviours which lead the object observed as an “agent” from an observer.

It’s ironic that the first Turing test that worked (CAPTCHA) had a criterion that feels so much lower than the original: “can you reproduce these distorted characters” vs “can you communicate like a human”. Another case of hard things that look easy and easy things that look hard?

As for perceptual crossing, I like that the symmetry is inbuilt: the agents adapt their behavior to each other but also to their current opinion about each other (“I think my partner might be a robot so I will move this way”). There might even be a recursive aspect: “I think my partner is a human who thinks I am a robot, so I will move this way”, “I think my partner is a human who thinks I am a robot who thinks he is a robot...”

Maybe for fairness in the Life Turing test, both participants should also be forced to use the same embodiment. It might be complicated but I think there would be some advantages. First, the data would be easy to analyse. Second, animals (or plants) of different scales (sizes) could interact. Even entire colonies could interact. Actually, maybe perceptual crossing could be the Life Turing test. For example, with fish, they might have to swim in a long tube and receive a stimulus (vibration? light?) when crossing the remote partner.

I think we have to be satisfied with a purely observer-based criteria on the Turing test. When we do the Turing test, there are so many explicit and implicit regulations. For example, there may be physical constraints that determine how you interact with your opponent. In a linguistic Turing test, no one would suddenly throw a keyboard at the wall, and nothing would be communicated to the opponent when it is thrown. If a chatbot passes the Turing test when we limit the interaction to verbal and set a certain time constraint, then we have to say that chatbots are life under that constraint. The problem is that when we feel life normally, we are using all our senses, not just language. If we live in a world that allows only verbal interaction with others, and we recognize the other agent as life through verbal interactions, then we must judge that the chatbot is undoubtedly life (under the time constraint). However, this situation is very different from the situation in which we normally feel life. In this sense, I think the most natural way of doing the Turing Test is to perform the test with interactions where all five senses can be fully used. Actually, Lana can be an artifact created by a genius scientist, and I recognize it as a living organism without noticing it at all, and it may have already passed my Turing test (considering that sometimes it doesn't pass in a short time, the genius scientist is not genius enough though). It is interesting to do the Turing test with multiple species, and I have done fish experiments based on it, but I'm concerned that we can't be sure that non-human species are really making the decision between life and non-life. When I talked to a frog researcher, he told me that frogs only eat live food, and I thought I could do the Turing test, but the other guy told me that they eat dead food even if you move it around in front of them. In the fish experiment, rather than a Turing test, we are trying to see if communication is possible.

Just in case, I am not saying that the Turing test with constraints is not good. When you are doing research, it is always a good idea to set the sub-goals instead of just aiming for the final goal. Even if we limit the modalities as in the perceptual crossing experiment and investigate a sense of life through simple touch and movement, I think there is still a lot we can learn. I believe that this can be the basis for a sense of life in the interactions with all senses.

As for the question of whether self-replication and self-maintenance of life are involved in the interaction-based criteria, I believe they are. The Turing test tends to assume a timescale of a few minutes to a few hours at most, but life also has various timescales, and in the long run, even when Keisuke dies and we are faced with the inevitable fact that we will never be able to talk to you again, we will feel life from you.

It might be hard to get rid of pure subjectivity, but I think varying the observer/judge would at least make sure it’s a shared objectivity.

I’m not sure if “using all senses” is really a requirement, but you could always perform the same experiment with different senses: for example, the perceptual crossing experiment used touch, but you could replicate the results with sound or eye tracking. As long at it’s the same info being transmitted (1D position and crossing or not crossing), I don’t think a variety of senses would matter. So I guess what Hiro said about “using all our senses” is more about the density of information, even when using a single sense.

It’s interesting to me that animal behavior science still says things like “we can’t be sure what animals think.” Of course we can’t be sure, but we can’t be sure with humans either, especially with babies; yet we still believe we can guess why they do what they do. I think observing the animal’s reactions or training them is enough to make inferences… For example, for the frog, maybe the threshold of “prey” is really just very low and any motion is enough to convince them. To be frank, I think probably any moving robot could be recognized as life by other animals, and also by human babies… And by human adults if we have never seen a robot in our life and we don’t know how it’s built. After all, many cultures used to (or still do) recognize typhoons or rivers or volcanoes as life.

So maybe we need a much stronger criterion. Something like “would you adopt and raise this agent as your own family?” or “Would you jump into a fire to save this agent?”

A very dangerous and costly decision, where you have to decide to constrain or risk your own life. It could be too limited, as each species has their own criteria, but maybe it’s still general enough: after all humans adopt all kinds of animals and plants, and there are other examples of interspecies adoption, elephants and dolphins saving other animals from predators... Yes, that’s my conclusion: we need an extremely strong criterion for the Life Turing Test, a decision that is truly costly to the judge and brings them no advantage. Extreme Altruism as a Life Turing Test.

Even if we can create something that makes them feel it is life through a specific modality, I don't think anyone will be convinced that we found a theory of life. I think people will only be convinced when we can create something that makes them feel it is life in all the senses because you can't argue that it's not in any modality paths.

The idea of Extreme Altruism as a Life Turing Test might be a good idea. When my son was still a baby and crying in a bad mood no matter what I did, I suddenly wondered if there was a power switch of him. Of course, there is no such thing, and even if there were, I would not press it. However, we can easily turn off the power of our robots. Life keeps running until the end. Even if there is a power switch of robots and we can't push it, we might feel life from them.

One of the difficulties of using multiple modalities to judge life from others is that it is more confusing between contradicted signals: ‘Uncanny Valley’ is a good example for that. something looks life-like, but not quite in other modalities makes us feel this kind of strangeness. This phenomenon itself is interesting to explore, but I would rather take the minimalist approach such as the perceptual crossing paradigm as an approximation.

The idea of Extreme Altruism as a Life Turing Test is indeed interesting and I found it is similar to one of the ways to define animal consciousness. When one is asking ‘where is a boundary between animal with consciousness and animal without consciousness’, I would rephrase the question to ‘which animal can feel pain or can have suffering?’ Answering this question might reflect our embodiment and an ability to infer other’s mental state (e.g. it’s more difficult to guess how a sea cucumber feels pain than how a horse feels pain), but it still make clear our intuitional boundary to divide the systems we feel it’s important for us and the others we don’t care that much.

By the way, once I met a person saying ‘a fish can feel pain, but does not have consciousness’. I didn’t understand what this meant for a while (as I thought there is no pain if there is no consciousness), but now I think he means consciousness as much higher order mental activity. The possibility of this dissociation between life and mind is another interesting observation, in terms of animal rights. Do we prioritise saving living forms having higher order complexity of their ‘mental states’ or all living forms which resist death must be saved equally as they can feel pain anyway? I don’t have an answer for this, but I sometimes feel ‘more complex is better’ is quite a Western view to see the world (As a side note, I believe this is the reason why some people are really afraid of the rise of AI (i.e. the singularity). Once something more complex than human beings appears, our existence is threatened just because we are inferior).