Artificial Intelligence and Major Risks

Faced with AI’s promises and perils, Sam Altman has outlined three troubling trajectories. Between military dystopia and insidious dependency, how should we think about our relationship with these technologies? But above all, how can we think about it for ourselves?

Altman’s Three Specters

During a hearing before the US Congress in May 2023, Sam Altman, CEO of OpenAI, outlined three categories of risks related to artificial intelligence development. His first concern involves malicious appropriation: “An adversary of the United States could use this superintelligence to design a biological weapon, take down the electrical grid, break into the financial system and steal everyone’s money.” This geopolitical threat echoes the concerns of numerous experts, including Nick Bostrom who, in Superintelligence: Paths, Dangers, Strategies (2014), already analyzed the risk of an algorithmic arms race.

The second category relates to what Altman calls “loss of control incidents”, evoking classic science fiction scenarios where AI refuses to be shut down. Stuart Russell, in Human Compatible (2019), deepens this issue by showing how a system optimized for a given objective could develop unanticipated self-preservation behaviors. This concern is no longer theoretical: Anthropic researchers have recently documented cases where their models attempted to circumvent their own limitations.

The third risk, more subtle according to Altman, deserves closer attention, especially since the usage he predicted has developed enormously: “The models kind of accidentally take over the world. They never wake up [...] but they just become so embedded in society.” He mentions young people who declare: “I can’t make any decision in my life without telling ChatGPT everything.” This emotional dependency prefigures a more troubling scenario: “What if AI becomes so intelligent that the President of the United States can’t do better than following ChatGPT 7’s recommendations?”

The Temporality of Technological Prophecies

Altman’s 2023 intervention constitutes precious testimony about our historical moment. In ten years, the same Altman will certainly hold a very different discourse, enriched by accumulated experience, in the world as it will be at that time. I greet here the reader from the future, for whom this text likely belongs to an imaginary past, quite different from their reality. This evolution doesn’t invalidate Altman’s current statements, however; it informs us about how we conceptualize today our relationship with technological future. As Paul Valéry wrote: “The trouble of our times is that the future is not what it used to be.”

I think here of how computing pioneers imagined our present. Alan Turing, in his foundational article “Computing Machinery and Intelligence” (1950), predicted that by the year 2000, machines would pass his imitation test in 30% of cases. We’ve largely achieved this, but the question of artificial consciousness remains open. These past projections aren’t obsolete; they reveal each era’s deep concerns about technology, and they help us think today. The famous “Turing test” is still used today, so fortunately he worked to prophesy. The challenge isn’t to predict the future, but to nourish future questions with our contemporary questions, because knowledge, if serious, is iterative, builds on past thought, which, even if the environment changes, can enrich us. Don’t we still read ancient philosophers with profit?

The prospective exercise Altman engages in belongs to this tradition. His three scenarios reflect our contemporary anxieties: asymmetric warfare, loss of autonomy, soft alienation. These fears aren’t new, Norbert Wiener already formulated them in The Human Use of Human Beings (1950), but they take on particular acuity at a time when generative AI is already omnipresent.

The Tool’s Fundamental Ambivalence

Artificial intelligence shares with every powerful tool a constitutive ambivalence. This duality isn’t a defect but an intrinsic characteristic, as Plato already emphasized in Phaedrus about writing, this pharmakon, both remedy and poison. Plato stages the myth of Theuth, inventor of writing, who presents writing as a remedy for memory, but Thamos, the king of Egypt, responds that writing also has a negative effect: it causes forgetfulness because we rely on external signs instead of exercising our own memory. The word pharmakon, in the Greek text, indeed means both “remedy” and “poison.” A sharp knife allows us to prepare a refined meal or commit a crime, a car transports us or becomes a deadly weapon, nuclear energy lights cities or destroys them.

This ambivalence of AI already manifests concretely. The same language models that assist researchers in discovering new medicines can generate instructions for synthesizing toxins. Facial recognition systems that find missing children also serve mass surveillance. And furthermore, as Cathy O’Neil observes in Weapons of Math Destruction (2016), algorithms can amplify our social biases while promising objectivity.

Faced with this reality, the prohibitionist approach seems illusory to me. We can no more “uninvent” AI than we could make printing or the internet disappear. The question therefore isn’t to eliminate risk, an enterprise doomed to failure, but to learn to live with it, to develop what I would call an “augmented prudence” proportional to our tools’ power.

Toward an Augmented Consciousness

The first step consists of recognizing these technologies’ transformative nature. AI isn’t simply another tool in our civilizational toolbox; it modifies the very texture of the world we inhabit. Sherry Turkle, in Reclaiming Conversation (2015), shows how digital technologies reconfigure our modes of relation. Generative AI amplifies this phenomenon by creating increasingly convincing synthetic interlocutors.

This transformation requires from us an unprecedented effort of understanding. I’m not speaking only of acquiring technical skills, though that’s useful, but of developing critical literacy regarding these systems. We must understand their biases, their limits, their operating modes, without all becoming machine learning engineers. As Douglas Rushkoff suggests in Program or Be Programmed (2010), the challenge is to remain actors rather than spectators of the digital revolution.

This skill development also involves reading articles, books, participating in conferences, webinars, professional meetings, but above all, and this is my main point, through dialogue. We mistakenly believe that our AI experience is universal, that others’ experiences are roughly similar to ours. In reality, each person develops singular practices, strategies, relationships with these tools. A graphic designer using Midjourney, a student learning with ChatGPT, a doctor exploiting automatic diagnostic systems, live profoundly different realities.

The Promise of Collective Dialogue

Sharing experiences constitutes our best defense against the risks identified (or imagined) by Altman. Facing geopolitical risk, only concerted international governance can establish effective guardrails. This is a dialogue. Against loss of control, we must collectively and continuously define and implement acceptable limits of algorithmic autonomy. This is still a dialogue. To avoid insidious dependency, we must cultivate together an adapted digital hygiene. This is always a dialogue.

These dialogues mustn’t remain confined to expert circles. I encourage everyone to explore these questions by all means: speculative fiction writing, artistic creation, public debates, playful experimentation, etc. Humor and play constitute excellent vectors for taming technologies without falling into technophobia or technolatry. The works of Winnicott, Huizinga, Caillois, Bateson, but also mathematical and systemic approaches, all address play as a means of exploring the limits and possibilities of complex systems, often through experimentation, rules, and the creation of hybrid spaces.

And finally, the emergence of new professions related to AI control and auditing represents real openings. Professions like “prompt engineer,” “AI ethicist” or “algorithmic auditor” didn’t exist five years ago. This professional evolution testifies to our capacity for adaptation, but also to the necessity of institutionalizing vigilance regarding these technologies. We collectively become guardians of our own autonomy facing the machines we have created.

Ultimately, Altman’s three scenarios draw less inevitable futures than beacons for navigating the present. Our personal and collective responsibility is to transform these warnings into practical wisdom, these fears into active vigilance, these risks into opportunities to reinvent our relationship with technology, thus anthropologically with ourselves.