AI is a pyramid: built for free or at low cost by everyone, owned by a few, opaque even to its own architects. What if reliability — and security — became a luxury?
We admire the pyramids for their point. We forgot our hands. The scaffolding that allowed the blocks to be erected has disappeared — and with it, the memory of the work that made the monument possible.
I believe we are erecting a comparable scaffolding today. The artificial intelligences that we use every day are not the finished building: they are the transitional phase, the construction site. And the metaphor holds to the end, on three levels that must be looked at together.
Built by the multitude. With every interaction, we work. Our exchanges, our corrections, our preferences serve by default to improve these systems — the suppliers say so themselves, and cutting off this sharing is an option to activate, not an initial state. Not their raw knowledge, which comes from texts written by others and is currently being judged in court, but their fine tuning, their alignment. Real, unpaid, and above all unrecognized work: presented as simple consumption of a service. “But it’s voluntary,” one will say. Less and less: knowing how to use these tools has become a professional expectation, and refusal is starting to cost like a delay. Free in law, constrained in fact.
Owned by a few. Fluidity is not reliability. The model produces the probable, not the verified — it can make you faster and wronger at the same time, with the same confidence whether it is right or wrong. Reliability, true reliability — sourcing, verifying, tracing each statement — is not in the model: it is in what we build around it. And what is built around it sells. The coming divide does not concern the model, it concerns access to the data which anchors it to reality: the citizen will have the public web; the lawyer will have his legal bases, the doctor his clinical bases, the hospital which can pay will have audit and traceability. Everything happens as if reliability was not given to the greatest number of people because elsewhere, it is sold. I cannot prove intention; I am describing an effect.
And sealed rooms. This is the third plan, the most disturbing. The pyramids guard interior spaces whose function we still do not understand. The models too: their internal spaces still escape the very people who designed them — it’s the whole unfinished business of interpretability. We are building a monument whose builders do not know how to read the plans.
The first week of June 2026 gave this structure a dated reading. On June 1, Anthropic confidentially filed its IPO prospectus, a week after a fundraising valuing the company at around $965 billion. On June 4, the same laboratory published a call to provide the industry with a coordinated pause capacity. On June 9, it launched its most powerful model – and its version without guardrails, reserved for verified partners, in collaboration with the American government. IPO, call for caution, launch: the triad takes nine days. I do not read into it any conspiracy or innocent coincidence, and I do not question anyone’s sincerity: I see a structure where even the call for caution is now formulated from a position of interest. When the guardian of the brake and the owner of the accelerator are the same hand, the question is no longer the good faith of the hand — it is the architecture which entrusts it with the two pedals.
And this architecture has just crossed a threshold. The model reserved for verified partners knows, according to the technical report of its creator, how to identify and exploit new vulnerabilities in major systems. The risk of free distribution is real — let’s admit it without reservation. But the effect of the response is also: the best cyber defense capacity in the world becomes a reserved access asset, allocated by a private company linked to a State. The included take a defensive lead; those excluded – less resourced administrations, public hospitals, infrastructures in the South – remain more exposed. The question is no longer just who can attack with these models. It’s also who has the right to defend themselves with it.
What to do? On an individual level, treat these tools for what they are: not oracles, but the draft of a brilliant and fallible collaborator. Regenerate, compare, require sources, cross-check outside. It’s the right gesture — but it won’t be enough. We cannot ask everyone to reconstruct their access to the facts alone.
The real answer is collective. There are public libraries, public universities, public hospitals. There must be AI of general interest — not to replace private innovation, which is useful and legitimate, but to ensure that access to the verified, and now the ability to protect oneself, depends neither on the ability to pay nor on a certificate of trust issued by a company. The target is not “the best model in the world”: it is a public infrastructure that is sufficiently efficient, auditable, multilingual and sovereign for essential uses — school, health, justice, administration, research. It is not a utopia: UNESCO has been defending it since 2021, the United Nations since the Global Digital Compact of 2024, and projects like OpenEuroLLM are already building it for European languages. All that remains is to expand it — towards languages and cultures that the market will never deem profitable.
The pyramids still stand; their scaffolding has disappeared, and with them the hands which assembled the blocks. These were constrained by force; ours are through productivity. But the fate of the monument could be the same: we remove the scaffolding, and we forget the work that no one paid for what it was worth. Unless, this time, we decide to remember it — and build a monument that is for everyone.
The issue is not to slow down the AI. It is to prevent reliability — and security — from becoming a luxury.
