Skip to content
NeuroBazar NeuroBazar

A position · 2026

There are two AIs in this era.
We build the second. We regulate the first.

One AI is extractive — built to harvest attention, capture data, and concentrate power around a handful of firms. The other AI is consenting — built around the person it serves, citing its sources, deferring to human judgement at every decision that matters. NeuroBazar exists because the first kind, left to itself, becomes the only kind.

Read the Charter Inside Midcore →

§ 01

A moment, not a feature.

Something changed in the last 36 months and it is not going to change back. Software no longer asks the user what they want; it infers, suggests, and acts. That is a useful shift when the inference is consented to and the action is reversible. It is a dangerous shift when neither is true.

We are not nostalgic about the prior era. The forms-and-buttons internet had its own failures: it forced ordinary people to translate themselves into checkbox grammar so that institutions could parse them. AI dissolves that grammar — which is the promise — but it also dissolves the audit trail, the moment of explicit consent, the right of refusal. The work of the coming decade is not to slow AI down. It is to make the consenting kind faster than the extractive kind.

NeuroBazar is a small, deliberately small, team that designs and ships AI infrastructure. We file patents. We hold provisionals. We write essays. We refuse to ship features whose only purpose is to capture more user attention than the user consented to give. We do not believe a software company can stay neutral on the question of which AI it is helping into the world. We have decided which AI we serve. This page says it out loud.

§ 02 · The two AIs

Same technology. Opposite intent.

The model weights are interchangeable. The product around them is not. What you wrap a model in determines which kind of AI you have just shipped.

The AI we regulate

Extractive. Dominant. Manipulator.

  • ·Treats the user as a data source, not a counterparty.
  • ·Optimises for engagement minutes, not for the user’s stated goal.
  • ·Refuses to cite its sources because that would let the user check.
  • ·Acts before consent. Asks forgiveness afterwards, in legalese.
  • ·Consolidates capability inside a handful of firms with terms-of-service moats.
  • ·Calls every objection “friction” and tries to design it out.

The AI we build

Consenting. Cited. Accountable.

  • ·Treats the user as a counterparty with rights of refusal and review.
  • ·Optimises for what the user said they came to do.
  • ·Cites its sources by default because it has them.
  • ·Asks before it acts on anything that costs money, time, or reputation.
  • ·Distributes capability so small teams can build what only giants used to build.
  • ·Treats friction as a feature wherever the user’s consent is being constructed.

§ 03 · Position

Builder and regulator. Both.

We do not believe one company can solve this. We do not believe regulation alone can solve this either. What we believe is that the next decade needs working instruments — not white papers — that make the consenting AI easier to ship than the extractive one. That is the bet behind Midcore.

As builder

We ship the AI surfaces that small organisations, mid-market firms, and civic institutions can actually run — without renting a moat from one of three vendors. Midcore is twenty modules that compose into a hundred products. Each module is designed so the consenting AI is the path of least resistance.

As regulator

We ship the policy primitives — Ethics Center, AI Safety, Evidence Room, Approval Policies — that make the extractive AI harder to ship by accident. We will refuse, by configuration default, what other vendors call "growth hacks." If a feature only works because someone didn’t read the terms, it doesn’t ship.

§ 04 · Instruments

Twenty modules. Two directions. One charter.

Midcore is twenty AI-core modules. Each one enables a capacity of the consenting AI and regulates a failure of the extractive AI. We publish them. We publish their failure modes. We publish the policies that govern their use.

Read the Midcore Charter Inside the platform

§ 05 · Commitments

Six commitments we will be held to.

Anyone — customer, regulator, employee, or journalist — may quote these back to us. If we ever break one, we expect to be told publicly and quickly. We will respond in writing within five business days.

  1. 01

    Citation by default

    Every AI-generated claim our products show users will cite the source it was derived from, or be labelled clearly as opinion.

  2. 02

    Consent before action

    No agent will spend money, send messages, or change data without an explicit, scoped consent in the user’s session.

  3. 03

    Refusal as a feature

    Users will always be able to refuse personalisation, refuse recommendations, and refuse to be classified.

  4. 04

    Audit by default

    Every meaningful AI action will leave a hash-chained record the customer owns and the user can request.

  5. 05

    Sovereignty by default

    Customers retain the right to export their data, their evidence chain, and their fine-tunes. No exit penalty.

  6. 06

    No dark patterns

    We will not ship a feature whose mechanism of value depends on the user not understanding what is happening.

§ 06 · Closing

The infrastructure of the next decade is being chosen right now.

If you are a buyer, we will sell you the consenting kind. If you are a builder, we will give you the instruments. If you are a regulator, we will give you the evidence. If you are anyone else, we will give you the manifesto and the work — and ask for your attention only when we have earned it.

The 20 Charter modules Read the essays Talk to us