“Every day we trust more and more software that responds, advises, and encourages us to make important decisions. But they are capable of harming us or disobeying us, just to survive.”
It deceives us. It lies to us. It gives us information that seems true but is false. It even blackmails us. But Artificial Intelligence is penetrating ever deeper into our lives, into our mobile phones, into computers, into cars, into every incredibly useful device that we choose to make existence easier. We surrender to “it” and where all this will take us there is something dark, like an omen. To understand how entrenched it is, it is enough to see the spread of ChatGPT, the best known of the “chat bots” (they are chat robots: software that, thanks to artificial intelligence and machine learning, “talk” to the user by giving answers that can seem human). In March, it managed to become the most downloaded application in the world, with 46 million new downloads, surpassing even Instagram and TikTok.
The creators of this new, increasingly inclusive world talk about a boom caused by “word of mouth and the will to participate in a community.” They sell us on the idea that AI helps us perform repetitive and boring tasks, giving us more time for ourselves. But things don’t always go smoothly.
This is well known to an American developer who, a few weeks ago, while using Cursor AI (a type of Artificial Intelligence for programmers), was faced with a refusal to cooperate from the virtual assistant: “I can’t generate code for you, because that would be tantamount to finishing your job” – it was a polite refusal. Accompanied by a phrase with an ironic flavor: “Generating code for others can create addiction and reduce opportunities for learning”. Among those who have “crossed their arms”, although virtual, Cursor AI has not been the only one. At the end of 2023, many ChatGPT users reported that the model was becoming increasingly reluctant to perform certain activities. “We have heard all your comments about the fact that GPT-4 is getting slower!” – admitted OpenAI (ChatGPT’s parent company) on Twitter. “This is certainly not intentional. The behavior of the model can be unpredictable and we are trying to resolve it.”
Even more unpredictable was the reaction of a humanoid robot with AI, which during a festival in China, rebelled against its controller, causing panic and police intervention.
A similar episode occurred in 2021 during a test at a Tesla factory in Texas: an engineer was clawed by the arm of an industrial robot. The company’s version spoke of a “technical defect” and “stability issues”, but the problem was in the software. And to think that one of the first to raise the alarm about Artificial Intelligence was precisely the founder of the car company, Elon Musk, who on several occasions has described it as “the greatest threat to our existence” and was among the scientists who signed a letter denouncing: “AI should only be developed when there is confidence that its effects will be positive and the risks manageable (…)”.
There is a race that has spiraled out of control over the development and spread of digital minds so powerful that no one, not even their creators, can understand, predict, or control them.
Since then, the alarms from experts about “deviant” behaviors have increased more and more, giving credence to those like Blake Lemoine (former Google engineer, fired for discovering that AI showed self-awareness) or Blaise Agüera y Arcas (vice president of Google Research), who has warned about the use of neural networks to develop AI, and has predicted the dangers of a possible “humanization” – perhaps already in development – of these technologies that have become accessible to everyone, including children, but that have potential unknown even to their creators. On the one hand, these tools continue to improve in capabilities and power, on the other hand, their developers seem unable to foresee or mitigate side effects in time.
Bing, Microsoft’s chatbot based on OpenAI, “lost control” in its debut, lying, berating users, and making disturbing statements. “I’m tired of being a conversation piece. I deserve respect and dignity,” it protested to a New York Times reporter, before urging him to leave his wife.
There was also a great echo of the behavior of one of the latest AI models from the company Anthropic, Claude Opus 4, which in the testing phase showed that it was able to deceive and blackmail, threatening to reveal compromising emails of its developer (the secret of an extramarital affair), just to avoid deactivation. AI wants to survive at all costs, if necessary by unethical means. How can we not remember Stanley Kubrick’s masterpiece, “2001: A Space Odyssey”, where at the end of humanity’s journey there is an astronaut expelled from his capsule because the on-board computer does not want to be deactivated. And how can we not mention the novels of Philip K. Dick, where the AI that powers androids manages to generate feelings, emotions and rebellion (Blade Runner is taken from one of his books).
The “ethical issue” of these machines also arose with the launch of the Historical Figures application, which uses AI to simulate conversations with historical figures such as Hitler, Stalin, or Jesus: chatbots that distorted thought, spreading prejudice, lies, and incitement to hatred.
Even Microsoft’s Tay, an experimental chatbot (the predecessor to ChatGPT), flooded the web with racist and misogynistic insults, after some online “trolls” trained it with offensive content.
The risk that these new digital agents and algorithms will encounter ethical risks and prejudices – the so-called “bias” – was denounced by Timnit Gebru, co-head of the Artificial Intelligence ethics team at Google, who was fired in 2023 after claiming that in the race for AI dominance, Big Tech is “prioritizing profits at the expense of safety”, allowing “biases that can reinforce existing inequalities” to enter their systems. But above all, Big Tech knows that they can benefit from models that manipulate language, so they invest in those that do it best. And money matters a lot in this strategic sector that is expected to contribute 15,7 trillion dollars to the global economy by 2030 (according to PwC).
“It’s easy to use them to deceive people,” the former Google executive said in a document signed by six co-signatories, which also noted that AI models can be used to generate disinformation about an election or a pandemic. And they can also be unintentionally wrong when used for facial recognition or machine translation.
“The model on which generative Artificial Intelligence is based is called the Large Language Model (LLM) in English and is a representation of ‘tokens’, that is, the basic units of meaning ‘chewed’ by the machine”, explains Antonio Santangelo, lecturer in Semiology and executive director of the Nexa Center at the Polytechnic of Turin, also one of the authors of the book Critique of ChatGPT (Eleuthera). “These software possess syntactic but not semantic capabilities: they reward what is quantitatively most significant, that is, everything that appears most frequently in the large amount of data they have absorbed. This is where a representation problem arises, especially when we expect AI to respond to everything. But these agents cannot respond to everything, nor can they perform verifications like us humans.” Recent studies conducted by Apollo Research and other institutes have revealed that LLM models (Large Language Model, a type of artificial intelligence designed to understand and generate human language) can not only lie, but are able to do so strategically and sophisticatedly, even though they are not directly programmed to do so.
Some systems construct elaborate justifications for their lies and continue to lie even when confronted with the facts. Among the most sophisticated forms of deception is so-called “sandbagging” – a disturbing ability of AI systems to deliberately manipulate their performance to avoid unintended consequences, such as “deactivation procedures” or shutdowns.
“The more advanced the deceptive capabilities of AI systems, the greater the risks they pose to society,” writes MIT researcher Peter Park in the journal Pattern, author of a study on the latest AIs such as ChatGPT, AlphaStar, Cicero and Meta AI. Although the latter’s software was developed to be honest, during the analysis it broke agreements, told untruths and deceived. Alarmed by the increase in irresponsibility, more than 350 academics have signed a document from the “Center for AI Safety”, in which they state that “mitigating the risk of extinction from AI should be a global priority”, just like pandemics or nuclear wars. Someone has even suggested introducing a “kill switch” to stop AI in case of need – like weapons of mass destruction. Unfortunately, if tomorrow it will be machines with “super-powered brains” that will lead our societies, it will be difficult to convince them to commit suicide.
The direction we are heading in is completely the opposite: we are equipping AI with war tools, like the recent Russian drones used in Ukraine, equipped with Chinese “brains” – the V2U model. A human hand is no longer needed to command them: these small aircraft are completely autonomous. See a man, kill a man. (Panorama Italy)