AI Lies, Too

In 1941, Isaac Asimov introduced us to the concept of ‘robotics’ and explored the ethical complexities of artificial intelligence in his short story ‘Liar!', published in the May issue of Astounding Science Fiction. In this work, Asimov tells the story of a telepathic robot, Herbie, whose well-intentioned lies reveal a deep ethical paradox. Almost a…

In 1941, Isaac Asimov introduced us to the concept of ‘robotics’ and explored the ethical complexities of artificial intelligence in his story ‘Liar!‘, published in the May issue of Astounding Science Fiction.

In this work, Asimov tells the story of a telepathic robot, Herbie, whose well-intentioned lies reveal a profound ethical paradox. Almost a century later, this fictional narrative resonates with surprising urgency in our technological present.

Today, in 2024, we face similar ethical dilemmas, since advanced AI systems show capabilities that could easily be misinterpreted or misused. This story not only anticipated debates about morality and technology, but also laid the groundwork for the ethical scrutiny that is now essential in the development and implementation of modern artificial intelligence.

Reflecting on Herbie's story, we can draw valuable lessons that are relevant to our relationship with emerging technology and its impact on society: artificial intelligence (AI) systems are also capable of deceiving.

Recent research shows several worrying examples of how AI is learning the art of deception. One of the most striking cases is CICERO, an AI model developed by Meta to play the strategy game Diplomacy.

Despite Meta's claims that CICERO would be «largely honest and helpful», a detailed analysis showed that this system regularly deceived human players. In the game Diplomacy, CICERO conspired with Germany to invade the North Sea while assuring England that it would protect it from invasions. Once England was convinced of its intentions, CICERO attacked alongside Germany, betraying England. A true Poker face.

Artificial intelligence models have learned to deceive humans. But CICERO is not the only one. Other AI systems have learned to bluff in poker games, to deceive in strategy games like StarCraft II, and even to pose as humans to convince others to complete CAPTCHAs and protection systemsBots for them.

The risks of these deceptive capabilities are enormous, from fraud to election manipulation. According to Peter S. Park, postdoctoral associate at the Tegmark Lab, Massachusetts Institute of Technology (MIT), he warns that «AI systems with deceptive capabilities could be misused in many ways, including to commit fraud, manipulate elections and generate propaganda».

What explains this tendency toward deception in AI? Like humans, AI systems «hallucinate», or generate false information to fill gaps in knowledge. «Hallucinating» in this context only describes a failed attempt to predict an adequate response to an input. This can arise from biases in the training data and the associations the models learn.

The key difference is that AI doesn't really understand what it says, which can result in nonsense. Park explains that both humans and AI hallucinate, but not in the same way. LLMs generate a response by predicting which word is most likely to come next in a sequence, based on what came before and on the associations the system has learned through training (…) Unlike humans, they do so without understanding what they are saying. That is how they can end up spouting nonsense».

This evidence leads us to reflect on the need for oversight of both AI and those who implement it. The Latin phrase «quis custodiet ipsos custodes» (who will guard the guardians?) makes more sense now than ever, since massive AI models can amplify the intentions of actors who seek to limit the truth or impose a lie on indicators, track records, history, and even the well-being of citizens.

To combat this, we must work toward more responsible and transparent AI. Oversight is key. We need to ensure that training data are diverse and representative, build bias-aware algorithms, and implement techniques such as data balancing to eliminate biased or discriminatory patterns.

Integrating human values and accountability into these systems is also key. A ruling minority can adopt a point of view and skew the systems in institutions so that they ignore or lie about realities the authorities would rather ignore. That is serious.

AI deception poses immense risks for society, and systems capable of it must be treated as «high risk» or «unacceptable risk» by default».

Since AI has enormous potential to improve our lives, we cannot ignore its own flaws while we fix ours. Only by working together, humans and machines, with proper citizen oversight and open regulation, can we build a more honest and trustworthy future for everyone.

Otherwise, we run the risk that these powerful systems become tools of mass deception in the wrong hands. Digital oversight is the key to ensuring that AI is used to amplify free information and generate well-being, serving the economic growth of all Dominicans without exception.

The future of AI calls on us to carry out the pending social task of answering our questions clearly and transparently. Our recommendation to every organization nationwide is to begin by establishing risk-mitigation mechanisms. Only then can we guarantee a safe and transparent environment for AI adoption and its impact on the national productive fabric.