Governance, regulation and possible futures

To make these families concrete, it helps to sketch their extreme versions, bearing in mind that these are scenarios, not predictions.

The bright side. In the most favorable hypothesis, often defended by optimists (Dario Amodei's essay Machines of Loving Grace, chapter 7, is one illustration), a powerful and well-controlled AI would act as an accelerator of knowledge. One can imagine decades of medical progress compressed into a few years: diseases that are incurable today understood and treated, early diagnosis made widespread, accelerated drug design (chapter 14). Add to that a scientific golden age (materials, energy, climate), a personalized tutor for every student (chapter 15), and, if productivity gains were widely shared, a form of material abundance freeing up time. This is the scenario in which AI delivers on its promises without its risks coming to pass.

The dark side. Conversely, several worrying trajectories do not even presuppose a hostile superintelligence. The most-discussed of the painful transitions is that of poorly anticipated mass unemployment: if automation (cognitive, then physical, chapter 17) outpaced job creation and the gains were concentrated, inequality would explode. A second risk is political: highly capable AI tools can serve generalized surveillance and an unprecedented concentration of power, among a few states or a few companies, up to a possible durable lock-in of an authoritarian order (chapter 21). A third lies in deliberate misuse: large-scale cyberattacks, industrialized disinformation, even a lowering of the barrier toward biological weapons (chapters 20, 22 and 24). A fourth, more insidious, was formalized in 2025 under the name gradual disempowerment: without any abrupt takeover, humanity could gradually lose its grip on its economy, its culture and its institutions, simply by delegating ever more decisions to systems faster and cheaper than itself, to the point where its interests cease to be at the center. At the extreme, finally, stands the scenario of a rapid loss of control in the face of a highly advanced and misaligned AI (the AI 2027 narrative, section 24.3). Less spectacular but often cited, a risk of meaning completes the picture: what would become of work, effort and the pride of creating in a world where the machine did almost everything better and faster?

None of these images is certain, and reality will probably mix the best and the worst depending on domains and regions. It is precisely because the range is so wide that the trajectory depends on variables we can, in part, bend.

The decisive variables are four in number, and each points back to a chapter of this course: the success (or not) of alignment and safety (chapter 24); the quality of governance and coordination (this chapter); the distribution of benefits, between concentration and democratization (chapters 9 and 17); and the evolution of geopolitical rivalry (chapters 8 and 22). As for the arrival of a possible artificial general intelligence (AGI), experts' estimates range from a few years to several decades, or even never: which is to say, the uncertainty is vast.

---

Key takeaways (chapter 25)

• Three regulatory approaches: Europe (rights and risk, extraterritorial effect), United States (market and export controls), China (state-led steering).
• The European AI Act classifies uses into a pyramid of risks (unacceptable, high, limited, minimal); general application on 2 August 2026 (high risk, transparency), penalties up to 35 million euros or 7% of turnover, against the backdrop of the protection-versus-competitiveness debate (Digital Omnibus).
• Governing frontier models raises the questions of mandatory audits (the aviation analogy), safety institutes, and state control, all hampered by the difficulty of international coordination in a context of a race.
• The futures range from a bright side (abundance, scientific golden age, transformed medicine and education) to a dark side (mass unemployment, surveillance and concentration of power, misuse, gradual disempowerment, loss of control), by way of a "muddling through" path; four variables will decide: alignment, governance, distribution, geopolitics. The arrival of an AGI remains highly uncertain.
• Throughline: AI is a general-purpose technology; the right stance is informed clear-sightedness, for its future will depend on our collective choices.

---