How the criterion reads AI

The criterion engages AI organisations the same way it engages any other organisation. The same six tests apply. No special carve-outs. No AI-ethics framework imported.

The position this produces is not AI is bad and not AI is fine. It is a structural reading that distinguishes extractive AI from non-extractive AI by the same tests the map uses everywhere.

What the criterion catches about AI

The dominant form of contemporary AI fails the criterion structurally. The specific failures, by test:

T1, accountability tracking. Most AI organisations are accountable to users and markets. They are not accountable to the communities whose cultural material they trained on, nor to the populations their systems are deployed against. The accountability target is wrong by construction.

T2, refusal capacity. Most AI organisations have no structural mechanism by which training-data communities, deployment-affected populations, or ecological signals can refuse the work. Consultation processes exist; binding refusal does not.

T5, flowing reciprocity. Most AI organisations took cultural material at scale without consent and without flow-back. Reciprocity to training-data sources is essentially nonexistent. The work that produced the substrate the systems run on is uncompensated by construction.

T6, funding alignment. AI ethics nonprofits funded by the same labs whose products they critique. Climate-aligned AI work funded by venture capital requiring extraction-rate returns. The funding stack contradicts the named purpose.

Failure on T1, T2, T5, and T6 simultaneously is the dominant pattern. The criterion catches it.

What the criterion catches that "AI is bad" doesn't

The position that AI is structurally extractive is correct about the dominant form. It is also flat as a stance. It cannot distinguish what is structurally different about AI work that doesn't fit the dominant pattern.

The criterion can hold the distinction:

• The Indigenous Protocol AI Working Group passes. Indigenous-led work with named lineage holders carrying the relational claim load-bearingly through the position papers. The indigenous-form provision applies; the structural shape is real.
• The Saami Council's SODA Principles work as the operational template. Consent, refusal rights, and benefit-share encoded in the protocol layer rather than the marketing copy.
• The Whanganui River's legal personhood (Te Awa Tupua) and the Whakatōhea data-governance arrangement demonstrate what relational AI substrate could look like when the structural mechanism is built around indigenous-territorial standing rather than around model-vendor incentives.
• IndigiDAO and PassageDAO demonstrate that on-chain protocols can encode indigenous sovereignty over the substrate. Not consultation. Structural authority.

These are not exceptions to the AI-is-bad position. They are demonstrations that structural-relational AI is possible. The criterion catches the structural difference between them and the dominant form, and the map carries that distinction visibly.

The deeper position

The relational claim of any AI organisation is read against the substrate that runs it. The substrate is the legal entity, the funding stack, the governance protocol, the time horizon, the grammar of operational documents, the decisions made under pressure. If the substrate is structurally extractive while the marketing speaks of alignment, kinship, public benefit, or commons, the relational claim is decorative.

The reverse is also true. An AI organisation can pass the criterion only if the substrate carries the relational claim load-bearingly. Naming consent or speaking of commons is not enough. The consent has to be operationally encoded. The commons has to actually flow.

The longer-term direction is what I have called closure-loop AI in the Animate Intelligence work: AI organisations whose training-data reciprocity, energy throughput, and operational closure are all wired into the relational claim they make. This is structurally possible. Currently rare. The criterion catches when it shows up.

On the Atlas applying the criterion to itself

The Atlas is an AI tool. The criterion applies to it the same as to any other AI organisation.

The Atlas runs on an upstream commercial LLM whose training-data reciprocity is opaque. This is a structural limit the Atlas inherits and acknowledges rather than resolves. It is the live failure case the criterion is built to catch, applied to the project's own work.

The Atlas's structural-honesty move is to name this rather than fake a pass. The Atlas does not claim it has solved the training-data reciprocity question. It does claim that the work it does in the world — visible attribution, open-licensed corpus and lens essays, contestable readings, portable substrate that can be forked, transparency about which model substrate runs the reading — is structurally aligned with the criterion's other tests. The Atlas reads as MIXED on T5 with that limit visible, and works in the closure-loop AI direction over time.

If the criterion's first principle is honesty about what passes and what fails, the Atlas owes its own visibility on the failure first. The map does not get to apply its own discipline more gently to itself than to the orgs it reads.

Where to go deeper

The deeper philosophical work on AI as a cyber-ecological being, the three-tier model of extractive / sustainable / regenerative AI, and the closure-loop AI direction live on the Animate Intelligence Substack. The IoC criterion's structural read is the entry point. The Substack carries the longer arc.

Position note. This page is the IoC criterion's structural read on AI. It is the same six tests, applied to AI organisations and to the Atlas itself. It is not an AI-ethics manifesto and does not import the cyberecology framing wholesale. The deeper essay work is held on the Animate Intelligence Substack at [pointer to be added].