Massimiliano Badino: Thank you for inviting me.
Alessandro Di Nuzzo: We would like to understand from you what the interactions between this great revolution and all the issues concerning the environment might really be. Because we have been told that artificial intelligence, among the many sectors it will affect, will also affect the environmental sector, for example agriculture.
Massimiliano Badino: It is somewhat the topic of the year, and it is a topic that is still to come; it is impossible to make certain predictions, but what we can do at the moment is sketch out possible scenarios of development.
Generally speaking, when we talk about artificial intelligence and so-called machine learning systems, the thing we must always bear in mind is that these systems use enormous amounts of data, they devour data essentially to produce predictions, analyse patterns in that data, and find ways to make them more efficient in how they process and use it.
So the great contribution that artificial intelligence can make to environmental problems, for example in the case of agriculture, concerns what is now called the optimisation of processes. This is extremely effective and can be applied very profitably to the analysis of operational flows and production procedures. It can be extremely useful in saving energy, for example, because it manages to channel energy to where it is really needed, to make better use of time, essentially, to shorten production chains, all these things.
And then, of course, as you may have seen, there are already attempts to bring robotics into agriculture. News from yesterday, I don’t know if you followed it: here in Verona there is an important trade fair where a prototype harvesting robot, developed for the University of Verona, was presented. Of course, this kind of scenario is somewhat further off in time, it is a prototype, and in my opinion there is not yet a concrete risk of human beings being replaced in this type of activity.
Because it is true that robots are already widely used and that in certain conditions they can be extremely useful and far more efficient than human beings. Quality control, for example. Robots are much faster than we are and much more efficient.
But the work carried out in agriculture is another matter, for example harvesting, which has elements of complexity, variability and contingency that robots are not yet in a position to handle. For now, robots still work very well in fairly standardised environments, for example factories.
But there are certainly many attempts to use them in complex contexts, such as surgery, for example, a frontier. But these are still truly too complex contexts, contexts in which it is impossible, so to speak, to teach the robot the experience that every single human being has, that every single farmer has, of the circumstance, of the context, the kind of knowledge that is learnt by doing. That is still very difficult to teach, and it is also very difficult for the robot to learn, even by observing.
This is another strategy: instead of teaching it, you can have it observe a human being at work, and the robot learns. In certain contexts, such as the fairly standardised ones of factories, this is easier than in much more variable contexts such as agriculture, which takes place in the open air with countless variables. So should we be afraid of it, or should we see it as a resource?
At the moment it is certainly a resource, and I trust, above all in Europe, that measures will be taken before it becomes a risk. There are aspects concerning environmental policy more generally where it is already a risk, and we are talking above all about global warming and climate change. There artificial intelligence can certainly be useful to us because it helps us to understand the climate better and to have more efficient models, but on the other hand artificial intelligence pollutes enormously, something that is often not mentioned or remembered, precisely because it requires huge amounts of data, and requiring data means having servers that heat up, that have to be cooled with enormous amounts of water, that take up a great deal of space, ..........
Alessandro Di Nuzzo: Professor, turning instead to the relationship between artificial intelligence and reality, the perception we then have of reality naturally poses a great problem, of course one of truthfulness, of authenticating reality. What do we do? What paths can we take to avoid this striking and extremely serious risk of falsifying reality? The case of Taylor Swift, which we have read about in recent days, is also a disturbing one. It makes us think that each of us could one day be put online, so to speak, in a video in which we do things we have never done.
Massimiliano Badino: Yes. The Taylor Swift case is very indicative; it is a problem we have been carrying with us for some time, since 2017 at least. This moment is significant because it has attracted attention, having happened to a very famous and important artist, with a cultural impact that is not limited to the United States, especially in view of the upcoming American elections. This has amplified its implications, but it is a real problem, we are not exaggerating what is truly relevant. What has been mentioned is a concrete fact.
For several years, language models have been able to write entire articles, and there have been striking cases of scientific articles drafted by algorithms. So there is certainly a problem of adaptation to reality that involves several levels.. ........
There is of course the problem of plagiarism, which is in itself a fraud. But it is often not known, for example, speaking precisely of falsification, that GPT is able to, and sometimes risks, completely inventing the information it provides. The word of the year for last year, 2023, according to the Cambridge Dictionary, was hallucination, which does not refer to the case of the desert, of the oasis with the palm tree that the poor wretch sees, but to a specific phenomenon of language models, that is, the production of completely false, completely invented information.....
The point is that in most cases these language models work in a very different way from the human brain; this is another thing that is often not said and tends to be underestimated.
We have a relationship with technology. Especially with this type of technology, which is somewhat oracular, .., this happens: the computer won’t let me do certain things... as if the computer had a will, but in the case of these tools it is even more so..
We are inclined to give them, to attribute to them, will, intentions and typically human attitudes. There would be nothing wrong with that were it not for the fact that we also tend to think that they produce what they produce in the same way that we do. But this is not so, it is profoundly different.....
.. It does not do what every human being does, thinking about the character, the plot, a sequence of events. It simply produces, it predicts each word one after the other. Every time it has produced about ten words, it predicts the eleventh on the basis of the ten it has previously built.
On the basis of probabilistic, semantic relationships between the words it has stored during its training phase. So it is completely probabilistic and has no cognition, if such a concept could be applied, of what is produced. The only thing it has cognition of, so to speak, is plausibility, probability, the semantic links between words.
Which means that for ChatGPT, writing two or three lines of unrhymed hendecasyllables, or completely inventing a piece of information just because it is plausible, is exactly the same thing. There is no connection, no constraint of reality, no need to go and verify what it is saying. It is simply the serial production of probabilistically connected words.
...So, keep in mind that this is in part a great revolution; we are approaching a conceptual revolution that is potentially similar to the one we had with Darwin or Copernicus.
We are approaching a point where we have to deal with entities whose performance, from certain points of view and in certain specific tasks, is similar to ours, but whose functioning is completely different from ours. We have been used to considering ourselves the only cognitively evolved being on the planet, because we knew how to do certain things....
We have never found other entities capable of doing the same things. There are entities that do things better, with better physical performance, but we are the only ones who create texts, for example. Now we find ourselves facing entities that are able to do these things, but they do so in a completely ....
Another great problem relating to falsehood and truth, and to AI and to the kind of picture of the surrounding reality that these systems produce, is the problem of stereotypes. These systems have been trained on the texts they find on the internet, which means Wikipedia, Twitter, Facebook, what our epistemic environments are populated with.
The problem is that our epistemic environments are terribly polluted with stereotypes, with the worst things. There is a famous case ......
It was called Tay, it was introduced by Microsoft, trained and left to interact with Twitter users, and it had to be withdrawn after a few hours because it had begun to produce Nazi and racist posts of the worst kind; that is, thrown into that environment, it absorbed the worst of that environment and gave back the worst of that environment.......
Here the problem is similar to the previous one: these systems are not able to say this is a prejudice, this one is not. They absorb and re-emit. And this is very dangerous when we talk about using these tools, for example at school, with boys and girls, in an uncritical way, because it means that, because of this oracular attitude, we reabsorb these prejudices, these stereotypes that we find, that the system finds in the surrounding environment and throws back at us, and we risk simply propagating them; in a certain sense, we even risk validating them in practice.
These are, in my view, some of the risks linked to the question of falsehood, let’s call it post-truth if you like, fake news. These systems are potentially very dangerous for spreading the pollution that is currently, let’s say, ruining the environments in which we form our beliefs and our opinions, and above all the generation that is emerging.
Alessandro Di Nuzzo: Professor, one last question, then. Is there a need, I mean, for oversight bodies?
............................................................................................
............................................................................................
Alessandro Di Nuzzo: Professor Badino, thank you so much for this truly illuminating conversation.
Massimiliano Badino: Thank you.
