The authors are university professors, experts at Google or OpenAI, or researchers in renowned institutions such as CNRS or Inria. Their research attempts to assess the possible slippages of the multiple forms of present and future artificial intelligence (AI), and the consequences of a loss of control on the part of the creators of these models. Let’s get rid of a fantasy right away: there is no question in the following examples of any awareness of machines, which remains in the domain of science fiction.
Researchers in computer science, however, are very interested in the situational awareness, which is generally translated as “situational awareness” and sums up with the acronym APS a series of intriguing concepts. APS means Advanced, Planning, Strategically Aware : a system has capabilities for solving complex problems which, in many areas, greatly exceed those of the human brain. He is also able to plan, to anticipate the result of his actions, thanks to a strategic awareness which “allows a model to design plans to take and maintain power over humans and their environment”, believe these scientists. .
For several years now, AIs have been able not only to regurgitate their learning, but to extrapolate and show themselves to be authentically creative in ways totally unforeseen by their designers. In a spectacular demonstration (The Surprising Creativity of Digital Evolution, 2019), researchers have simulated a mechanical spider with six feet, whose mission is to move forward on a flat surface while minimizing contact with the ground. At first, the claudic virtual creature squirms, then without warning, it rolls onto its back and manages to crawl at a good pace, without therefore touching the ground with its paws: mission accomplished!
More quantifiable was the progress of go game simulations developed by Google DeepMind. The best-known version is AlphaGo, which beat the world champion in 2016. But the later, lesser-known iterations are even more troubling because they demonstrate the self-learning potential of these systems. Initially, the algorithm had learned the rules of the game of go after ingesting millions of games. Once the principles had been assimilated, its creators had it play countless times against itself, with spectacular results: in three days of self-training, the version called AlphaGo Zero beat the previous one, the one that had defeated the champ Lee Sedol; in twenty-one days, Zero surpassed all the world’s great masters of the game of go and after forty days, he surpassed all his own previous versions, without any human intervention.
AIs ready to do anything to achieve their ends
These two examples are proof of the phenomenal ability of these systems to learn and improvise creatively and independently. These notions are essential to apprehend the potential dangers. Because without even mentioning the diversion of AI for criminal purposes, specialists have highlighted several possible deviations.
One of them concerns the control of the objectives of an AI. It’s called alignment – with respect to human values, of course. However, this notion is in practice unstable. AI researchers mention a risk of misalignment of these models which would succumb to over-motivation in relation to a given objective, by exploiting loopholes left by the creators of AI, or if necessary by developing devices intended to thwart any surveillance.
In a paper titled The Alignment Problem from a Deep Learning Perspective (February 2023), three experts discuss models ready to do anything to achieve their goals. Their determination will be all the more absolute as they are designed – clearly stimulated – in order to constantly improve their performance, and as they are, by construction, devoid of scruples or any moral sense. Therefore, write the authors (from Berkeley, OpenAI and Oxford), if the goal is to maximize stock market gains, the model will inevitably seek to manipulate the market; if he is commissioned to do scientific research, he will make sure that experimental data lie; and if the AI is supposed to build an application with a high audience, it will seek to maximize its gain by designing the most addictive interfaces. This last point recalls some excesses observed in real life, isn’t it Meta or TikTok?