August 2024 marked a historic date: Regulation (EU) 2024/1689 – the AI Act, came into force, making the European Union the first major global player to adopt a comprehensive law on artificial intelligence. A milestone many had been waiting for, and one that few have actually read all the way through. Because reading it, something interesting emerges: the regulation governs risks, classifies systems, establishes obligations. But on sustainability, the real kind, the kind that measures AI’s impact on the planet and on people, it still treads lightly.
And this tells us something important: governing artificial intelligence is not just a matter of compliance. It is, first and foremost, a choice of values.
What the AI Act Actually Says
The text is long, technical, necessarily complex. But its framework is clear: high-risk AI systems, those affecting health, safety, and fundamental rights, must comply with stringent requirements for transparency, human oversight, and robustness.
Risk-based classification is the central instrument:
- Unacceptable risk: prohibited systems (e.g., social scoring, subliminal manipulation)
- High risk: subject to compliance obligations (e.g., AI in medical, judicial, and educational contexts)
- Limited risk: transparency obligations (e.g., chatbots, deepfakes)
- Minimal risk: free from specific constraints
As of August 2025, the rules on GPAI models also entered into application, the large general-purpose models underlying systems like ChatGPT, Gemini, and Claude. A significant development: for the first time, even systems without a defined intended use are subject to transparency obligations and, for models with systemic impact, to risk assessment requirements. It is an acknowledgment that the most pervasive AI is not the vertical kind, but the horizontal kind, the kind that quietly enters every process.
On the environmental front, however, Recital 48 and Article 9 merely hint at the ecological impact of high-risk systems, with no mandatory metrics or defined thresholds. The European Parliament had called for stricter requirements on the energy footprint of models. The final version of the AI Act did not incorporate them in any binding way.
The Gap Between Regulation and Reality: The Numbers No One Wants to Mention
While lawmakers were calibrating risk thresholds, data centers kept growing. The numbers are hard to ignore. According to the IEA — International Energy Agency, global electricity consumption linked to AI and data centers has already doubled over the 2024–2026 period, and projections to 2030 continue to climb. Goldman Sachs estimates a 160% increase in energy demand by 2030, driven precisely by artificial intelligence. Training a single Large Language Model can generate CO₂ emissions comparable to hundreds of transatlantic flights — as documented by Strubell et al. (2019) and, in greater detail, by Patterson et al. (2021).
And yet, paradoxically, AI is also being presented as a key tool for achieving the climate goals of the Green Deal. Energy optimization, environmental monitoring, predictive models for renewables: the promises are there. But the European Parliament warns that without adequate governance, the risk is that AI becomes part of the problem, not the solution. The contradiction is structural. And the AI Act, on its own, does not resolve it.
A Sustainability That Also Concerns the Mind
There is, however, a dimension of sustainability that the regulatory debate tends to overlook almost entirely: the cognitive one. AI does not impact only the planet. It impacts how we think. The phenomenon of cognitive offloading, the tendency to delegate to machines mental processes we once carried out on our own, is reshaping the cognitive habits of entire generations. As we explored in our article on Gen Z and AI, younger people are growing up in an ecosystem where the artificial assistant answers before the question has even been precisely formulated. The risk is not stupidity; it is cognitive dependence. And governing AI in a truly sustainable way also means asking: what kind of intelligence are we cultivating in people, while we enhance that of machines? Responsible AI governance cannot ignore this question. Nor can it answer it with a regulation.
The Real Challenge: To Govern Is to Choose
Here lies the point that interests us most, as an agency working every day at the intersection of technology, communication, and responsibility. Governing AI is not a technical act. It is a cultural act. Rules are necessary, and the AI Act is a significant step forward compared to the regulatory vacuum that preceded it. But rules define boundaries, not direction. And direction is chosen by people: the decision-makers who determine which systems to adopt, the professionals who choose how to use them, the organizations that set their own internal standards.
The OECD AI Principles framework and the UNESCO Recommendation on the Ethics of AI, adopted by 193 countries, converge on one point: digital sustainability is not a technical indicator. It is a commitment that runs through every level of an organization, from the boardroom down to the daily choices of every team.
The UN AI Advisory Body itself, in its 2024 final report, says so explicitly: we need AI governance that is inclusive, adaptive, and capable of connecting innovation with human development. A regulation is not enough. A vision is needed. And, surprisingly, this vision has also come from an unexpected voice.
When Even the Pope Takes a Stand
There is a voice no one expected in this debate, and yet it has proven to be among the most lucid. On May 15, 2026, Pope Leo XIV published Magnifica Humanitas: the first encyclical in history devoted to artificial intelligence. A document that goes far beyond theological reflection: as we wrote in our reading of the encyclical, it is an act of political and philosophical resistance.
Its central point is uncomfortable and necessary: AI is not neutral. “It takes on the face of those who conceive it, fund it, regulate it, use it.” Behind every algorithm lies a hierarchy of power, often invisible, always present. And when power cloaks itself in neutrality, “injustice becomes silent.” Leo XIV uses a strong word: disarm. Disarming AI means removing it from the logic of competition, not only military, but economic and cognitive. It means making it, in his words, “debatable, contestable, and therefore inhabitable.” This is precisely the challenge the AI Act cannot meet on its own. Because a regulation can define boundaries. But only a vision of the human being can provide direction.
The Model We Choose
At Keyformat, we live this tension every day. AI is part of our work, it amplifies our capacity to listen, accelerates analysis, enhances creativity. But it does not replace choice. It does not replace responsibility. What the journey toward our Manifesto taught us is that digital sustainability begins with a question: not “what can AI do?” but “why are we doing this, and for whom?” The AI Act has provided a regulatory answer to part of that question. Now it is up to organizations, and to each of us, to answer the part the law cannot reach: the part about values, priorities, and meaning. Because governing artificial intelligence does not just mean following the rules. It means being able to explain, every single day, the choices we make with it.
