Insights · The position · No. 01
The era of AI is a society question, not a software question.
Software used to ask what we wanted. It now infers, suggests, and acts. That shift is useful when consented to — and dangerous when neither consent nor reversibility holds. A note on what just changed.
NeuroBazar Editorial · · 6 min read · Series · Two AIs
Most articles about AI begin with a benchmark — a graph, a model number, a leaderboard. This one starts somewhere else. The benchmark that matters in 2026 is not the model. It is the posture: the stance software now takes toward the person on the other side of the screen. For thirty years the posture was deferent. Software asked. It offered checkboxes, dropdowns, blank fields. The person decided, and the system executed. That posture is over.
The new posture is anticipatory. Software infers what you probably want, suggests it, and increasingly acts on it. When the inference is good and the action reversible, this is a real gain — fewer steps between a person and their intent. When the inference is poor and the action irreversible, it is a different kind of system entirely: one that decides about you, with you as the audit log. Neither of these systems is hypothetical. Both are shipping. They look identical from the outside.
The institutional shape of the shift
The previous internet had a grammar. Forms, buttons, terms of service, GDPR pop-ups — clumsy, often theatrical, but a grammar nonetheless. Institutions parsed people through it; people, in turn, parsed institutions through it. The audit trail was bad, but it was there. You could screenshot a form, save a contract, point to a checkbox you didn’t tick. The AI era dissolves that grammar. It also dissolves, by default, the artefacts that grammar produced: the recorded moment of consent, the bounded scope of an action, the right of refusal as a first-class affordance.
Dissolving the grammar is the promise — the friction was real, and ordinary people paid most of the cost. But the artefacts of that grammar were doing other work, too. They held institutions accountable. They gave courts something to read. They gave individuals something to refuse. Replace them with vibes-based interaction and you have not removed a tax on the user; you have removed an instrument of leverage from the user. The work of the coming decade is not to slow AI down. It is to give the user back the instruments — in a form the new posture can actually respect.
Two failure modes hiding inside one trend
Discussions of "AI risk" tend to fixate on one of two failure modes: capability disasters (the model does something nobody asked for) and alignment disasters (the model does what someone asked for, but the wrong someone). Both are real. Both are well-covered. The third — the one that is already happening at scale, on the consumer surfaces you already use — gets less attention. Call it posture drift: software that incrementally moves from asking to inferring to acting, and never pauses to ask the user whether that drift was consented to. The user’s subjective experience is "this app is suddenly making decisions for me." The institutional reality is that someone decided, in a release note, to grant it permission.
Posture drift is not a model failure. It is a product decision. It is reversible. The reason it keeps happening is that — for most product teams — shipping the consenting posture takes more work than shipping the extractive one. You need a consent vocabulary the model can read. You need an evidence chain the auditor can verify. You need policy primitives that refuse to bypass the user’s consent regardless of what the prompt says. You need, in short, infrastructure. Without it, every team rebuilds these protections from scratch, badly, on a deadline.
What we are building, and why
NeuroBazar exists because we believe this infrastructure should be a public good — not in the sense of free, but in the sense of available. Our platform, Midcore, is twenty modules engineered so the consenting posture is the path of least resistance. Citation as a default. Evidence as a default. Refusal as a first-class affordance. Approval as the only path past a money- or data-affecting boundary. We sell these modules to enterprises because that is the model that funds the work. We publish their behaviour because the alternative is asking customers to trust us — and we do not think trust is what this era needs. It needs checkability.
The work of the coming decade is not to slow AI down. It is to make the consenting AI faster than the extractive AI.
If the era of AI is, as the trade press keeps insisting, a "platform shift," then the platform is not being chosen later. It is being chosen now, every day, by the product teams that decide whether to ship the inference with or without a consent vocabulary attached. Those decisions are still small. They will be enormous by 2030. We wrote this essay — and we are shipping these instruments — because we want the next decade’s platform to be the consenting one. We do not think the choice will be made by manifestos. It will be made by what is easier to ship. We are working on making the right thing easier.
Continue with the next essay in the series: The two AIs — extractive and consenting.