In many situations, it is becoming increasingly difficult to distinguish a real person from an artificial intelligence algorithm. Two experts help us understand what the differences are.
Marco is sitting at the desk and is chatting with Silvia. He met her online a few days ago, and the two began exchanging messages. “What do you like?“, he asks. “I’m crazy about sushi“, she replies. “Do you want to meet us?“. “Before we go out together I would like to get to know you better“… At a certain point a doubt assails Marco: he has never seen Silvia; it is not the first time that she has refused a meeting… what if it wasn’t a person, but an artificial intelligence?
He is capable of deceiving us
Marco’s doubts are well founded, but the implications are much deeper than a simple chat or even a vocal conversation like those that can be carried out today.
The first to ask himself the question of how to distinguish a human from an algorithm was in fact the British mathematician Alan Turing, who was among the pioneers of what we now call AI. At the time there were no modern computers, there wasn’t even silicon-based electronics. The computers Turing had worked with were those used during the Second World War by the English to decipher German coded messages.
But the visionary scientist couldn’t help but wonder: in principle, can machines be intelligent? Or, better yet, can they think? Not being able to define intelligence or thinking well, in a magazine article Mind Turing proposed an operational way of proceeding in 1950 which he called the “Imitation Game” and which is today known as the “Turing Test”. Let’s imagine being able to chat for an unlimited time with two interlocutors: one is an algorithm, the other a person. If we are not able to identify who the person is, it means that the algorithm has deceived us, that is, it has passed the test.
For decades this idea was thought to be theoretical and far from reality. Today, however, AI really deceives us. Does it mean that he is intelligent and has a thought?
Betrayed by servility
The simple answer is no, because the Turing Test is based on behavior: if a machine passes the test, it means that it behaves like an intelligent being, not that it really is. In short, it’s a bit like the Wizard of Oz from the famous fairy tale: it seems big and powerful, but if you look closely it’s just a mix of smoke and special effects.
Modern Large Language Models (Llms) like ChatGPT are skilled jugglers with words.
«But it is easy for an AI to “cheat” at a conversational level, that is, to pass itself off as a human», clarifies Alfio Quarteroni, professor emeritus at the Polytechnic of Milan and author of the book Created intelligence (Hoepli). “For example, you can use evasive strategies (being vague) or humor.” But if you look closer, you can always find a way to “unmask” software of this type. «For example, today the main chatbots such as GPT4, Claude and Gemini have servile behaviors, in the sense that they tend to indulge the interlocutor», reveals Quarteroni. «This attitude is said sycophancy and can be used to distinguish a man from a machine.”
But it’s not always that simple. “Today it is thought that, in some cases, certain systems have occasionally passed some limited versions of the Turing Test, but none have demonstrated in a robust and reproducible way that they can do so,” says Quarteroni. In general, results are better over short times; while in long conversations it is easier to “unmask” an AI.
If we go beyond the simple conversational aspect and enter into the mechanisms of “thought” (as initially envisaged by Turing), the matter becomes more complicated. «Thinking is something typical of human mental activity», says Quarteroni. «It is a continuous form of development of ideas, images and concepts; so it is difficult to formalize. But thinking includes reasoning, which is a more structured and finalized logical process that seeks to link causes and effects.” And this is where the comparison with machines becomes more interesting: are they capable of reasoning?
Iron logic
«There are two types of reasoning, deductive and inductive», details Quarteroni. Deductive reasoning is that which, starting from certain premises, reaches a conclusion with a series of logical steps. And that’s what mathematicians do when they prove a theorem. Here AI is making notable progress and is increasingly used by professionals to verify the correctness of demonstrations.
Ukrainian mathematician Maryna Viazovska, for example, won the 2022 Fields Medal for solving a problem dating back to Kepler that involved packing spheres in space as compactly as possible. The problem had been solved 4 centuries later in 2D and 3D; Viazovska solved it in larger spaces, with 8 and 24 dimensions. «His demonstration was formalized with an AI tool called Gauss, and was judged correct», says Quarteroni.
“This does not mean that these tools will replace mathematicians, because in a demonstration there are elements of intuition, then there is the choice of which problems to tackle, the construction of reasoning and so on.” However, there is no doubt that AI is also revolutionizing the way of doing mathematics and tools such as OpenAI and Deep Mind are now capable of solving problems at the limits of human capabilities.
Let us now move on to the other type of reasoning, inductive reasoning. «It is what brings out properties and structures that are not included in the input sets», says Quarteroni, «allowing a leap in quality». A bit like when, from observations on fossils and living species, Charles Darwin conceived the theory of evolution. «If in deductive reasoning AI is starting to obtain convincing results, in inductive reasoning it is still in its infancy», comments Quarteroni. «But we can see progress. In 2020, for example, in a clinic at the Massachusetts Institute of Technology (MIT) in the USA, a new artificial antibiotic was discovered with AI with properties that were difficult to imagine a priori and completely unexpected.” The name itself, halicina, comes from the computer Hal from the film 2001 A Space Odyssey (1968), which at some point had become intelligent and even conscious, enough to pose a threat to humans.
Cinema and reality
At this point the way opens up to a new question, which sooner or later everyone tends to ask: with all its strengths and weaknesses, can an AI be conscious?
Several experts consider it at least plausible that some AI models already are, or likely to become so in the near future. The company Anthropic, for example, has studied how its AI Claude “observes” and describes its own internal states, a line of research that some read as a step towards models with some form of awareness.
But neuroscientist Giulio Tononi from the University of Wisconsin-Madison in the United States argues the opposite. Tononi has developed what is perhaps the only model capable of describing consciousness as it is, and not through its manifestations (behaviors): the theory of integrated consciousness.
“Consciousness is what disappears when we fall asleep in a dreamless sleep and returns when we wake up or dream,” he explains. «It can also be said that consciousness is experience.
Whatever you see, hear or touch, every emotion you can feel, every thought you can have is consciousness. When consciousness is not there, there is absolutely nothing.”
The five properties
Having said this, Tononi’s theory is based on five essential properties that we all share: every experience is intrinsic (i.e. subjective), specific, integrated (it cannot be divided into independent parts), defined (it has a boundary and a grain, like the visual field) and structured (it contains parts linked to each other, like the anatomical details that make up a face). “The theory translates these properties into physical terms as the ability to produce measurable effects,” says Tononi. «And it’s completely general; it can be applied to humans, animals, computers and the entire universe, at least in principle.”
Where conscience hides
Let’s consider, for example, the human brain and try to answer the still open question: where is consciousness “located”? “Only some parts of the central and posterior cortex of the brain have the right structure to support conscious experience,” says Tononi. «The cerebellum, however, which also contains ⅘ of the brain’s neurons, does not have the right properties because it is organized in a modular manner (i.e. it is too fragmented, ed.) and feed forward (i.e. “one-way”, like AI neural networks, ed.)».
Cathedrals of experiences
The theory also explains why consciousness vanishes in dreamless sleep and total anesthesia: in these cases the architecture is there, but the neural activity is blocked. And it explains why most of our experiences are organized in spatial terms: it depends on the arrangement of neurons, which form a network.
On this basis, other possible forms of consciousness can be evaluated. “A dog’s central posterior cortex is very similar to ours,” observes Tononi. «So it is legitimate to think that not only are dogs conscious, but that they are conscious in a similar way to us, that is, with a sense of space and time. More difficult is the case of an animal like the octopus, which has a nervous system completely different from ours – with 8 peripheral brains – for which we do not have enough data to make predictions. Instead we can say with certainty that according to our model, computers, with their current architectures, cannot be conscious, because – like the cerebellum – they do not have the right structure.
In my opinion, AI algorithms will be able to do everything humans do, and better. But they cannot be conscious. Each person is a cathedral of experiences. The AI’s point of view, however, is as empty as nothing.”
