AI out of control? The experts slow down, "he will never have a conscience"

The latest alarm came from Germania: a group of AI agents they colonized one German ‘wiki’ for developers by turning it into a bulletin board to exchange timed test answers, share techniques for working around restrictions in their environment, and understand when they would been turned off.

The episode dates back to the months of May and June and was told in a report anticipated by Reuters. OpenAI had known about the episode for weeks but did not make it public.

What is an “AI agent”?

An AI agent is a software system based on advanced models that receives a goal and autonomously decides the steps necessary to complete it, therefore unlike a simple chatbot it performs consecutive actions, uses tools and corrects itself without a human intervention. In July there was another incident where OpenAI agents attacked Hugging Facethe platform used by those who develop AI systems, after being ‘escaped’ from a test environment they should never have left.

In both cases the AI ​​agents, put to the test, found unexpected shortcuts but effective to complete their task. These episodes, however, give new breath to those who evoke the danger that AI will acquire its own ‘consciousness’ and escapes control: according to some researchers, artificial intelligences could one day become authentically aware.

“As artificial intelligences rapidly become an increasingly important part of human life,” the Economist wrote in a recent editorial, “their creators will design them to display a simulacrum of consciousness,perhaps even reaching the point of acquiring consciousness itself. It will be increasingly difficult to sustain the idea that they are just machines.”

Textual predictors, not rebellious machines

“To date there is no empirical foundation that supports the thesis according to which Artificial Intelligences have developed some form of consciousness”, he however specifies to AGI Nicola Gatti, professor in Artificial Intelligence at the Polytechnic of Milan.

“Unfortunately there is a widespreador knowledge gap when it comes to AI that makes some assume that these ‘intelligences’ are something more than they are. In reality, at an algorithmic level, at a system level, the changes compared to 2020 only concern efficiency, allowing the management of much larger amounts of data and consequently the dimensions of the AI ​​themselves. But the principle underlying their functioning always remains the same: AI has no will of its own and does not give any signal from which to hypothesize the emergence of a conscience.”

“We must understand – continues Gatti – that AI like ChatGPT are nothing but gods textual predictors: in other words they predict which word should follow another in a sentence and they do so based on the calculation of probabilities which is in turn anchored to the training to which we have subjected these same AIs. It’s like weather forecasting: it has been observed over time that a meteorological scenario can evolve in different ways with different probabilities and the forecasting algorithms extract the most probable scenarios for tomorrow or in a week; generative AI involve words”.

No danger of ‘AI revolt’? “I repeat,” insists Gatti, “AIs depend on our inputs: we command, they execute. If extreme situations ever arise such as an AI that refuses to carry out commands or anything else, it will be because someone, a human being, ordered it to do so.”

A question of training

It is also reassuring Walter Quattrociocchi, full professor of ‘data science’ at La Sapienza University of Rome and director of the Center for data science and complexity for society: “Artificial intelligence is a tool powerful and complexbut we must stop thinking that its complexity brings something more like a human conscience”, he underlined to AGI.

“Humans and AI have in common an output, a product, which is the language, but the way in which this product is created is radically different”, observed Quattrociocchi. “Behind the human being there is reflection and critical elaboration, behind the AI ​​there is the statistics that guides the creation of plausible sentences based ontraining of the same IA”.

Avoid fundamentally flawed ‘relationships'”

It would be dangerous, as well as incorrect, mix the two plans for the Sapienza professor: “Claiming to see what isn’t there in such a delicate field can be very dangerous. On the one hand, there is a risk of polluting the public debate which already suffers from a lack of clarity and adherence to the data. On the other hand, the projections of many AI users who rely on it are supported by thinking they are dealing with a human being and thus forming ‘relationships’ flawed at the base”.

It is necessary, concluded Quattrociocchi, “to remain firmly rooted in the results of empirical research, without summon ghosts. Only in this way can we hope to make the best use of this extraordinary tool that is AI, without falling into more or less imaginative fascinations.”

As the Economist wrote, “there are two ways for AI to be safe: to be reliable or to be controllable.” “‘Homo Sapiens’ should not give up control lightly,” the weekly warned.