Regulating the Third Intelligence

One application · four different kinds of power

You use AI to improve a job application. The employer uses AI to summarise it. A recruiter uses that summary to decide which applications receive attention. Later, the organisation discovers that its records, workflow, and staff practices have become difficult to separate from the system that produced the summary. The technology may look similar at every step. Its institutional position is not.

The same application changes hands

At home, the assistant suggests a clearer sentence for your cover letter. You read it, change one phrase, and decide whether to keep it. The output is disposable. It helps you think without creating an official account of you.

At the employer, another system extracts skills, assigns categories, and produces a short summary for the hiring team. It may save time and help a recruiter notice experience hidden by unfamiliar wording. It may also turn a varied work history into a clean label such as “frequent movement” or “limited continuity.” The facts have not necessarily changed. A representation has been produced.

If the recruiter treats the summary as one clue among many, the system remains influential but bounded. If applications below a threshold are never opened, the representation has become part of an institutional act. If that workflow is later used across departments, stored in records, and tied to a platform the employer cannot readily inspect or replace, a decision aid has begun to reorganise institutional capacity.

This is not automatically harmful. Employers may receive more applications than people can carefully read. Structured assistance can reduce delay, make comparison more consistent, and help smaller organisations reach candidates they would otherwise miss. The serious problem is not assistance. It is allowing assistance to acquire authority without noticing the conversion.

The larger question

Policy debates often begin at the wrong level. “Was the model accurate?” matters, but an accurate output can still be irrelevant or used outside a lawful mandate. “Was there a human in the loop?” matters, but a signature does not tell us whether the person could understand or depart from the system. “Was there market competition?” matters, but several vendors may depend on the same underlying services or impose different forms of lock-in.

This essay uses third intelligence to name a relationship, not a new species of mind. The first intelligence is human judgment in lived situations. The second is the collective intelligence of institutions: records, routines, professions, law, and public reasoning. The third appears when artificial inference is placed inside that relationship and begins to shape what an institution knows, does, or remains able to change.

Nothing in this definition requires consciousness, personhood, or intelligence superior to humans. If AI becomes more capable than people across important domains, the stakes grow sharply, but the governance problem begins earlier: when an artificial contribution becomes an operative premise for collective action. Not every use crosses that threshold. A private drafting aid, a source-linked public guide, a risk score, and an automated decision do not carry the same authority.

Three conversions—and no automatic bridge between them

These conversions can reinforce one another, but none proves the next. A flawed profile does not prove that a worker relied on it. Reliance in one case does not prove structural dependence. Market-level switching barriers do not prove that a particular institution cannot exit. A defensible synthesis must test every bridge rather than use one troubling example to fill evidential gaps in another.

First conversion: information becomes representation

Your application contains dates, roles, qualifications, gaps, and explanations. A summary selects among them. “Three short contracts” may become “limited continuity”; time spent caring for someone may appear only as “employment gap.” The summary can be factually linked to the document while still presenting an incomplete or weakly justified account.

Representation is therefore more than data accuracy. The institution must also ask whether the inference is relevant to the purpose, whether proxy effects have been tested, what uncertainty was removed from view, and which counterevidence can change the operative account. A person need not have an unconditional veto over official evidence. But the institution should be able to justify the representation it relies on and provide a fair route through which a challenge can alter it.

Repair must travel as far as consequence. Correcting an input is inadequate if a derived marker remains in another record, continues to shape later treatment, or has already caused a loss that no correction restores. The practical question is: where did the representation persist, who received it, and which decisions must now be reconsidered?

Second conversion: representation becomes action

The same summary can occupy several positions. It can be an optional note a recruiter may ignore, the first item a recruiter sees, or a filter that determines whether anyone sees the application at all. The output has not necessarily changed. The workflow has changed what it is allowed to do.

Capability is not authority. Technical permission is not a legal or democratic mandate. And a human signature is not proof of independent judgment. The evidence must show whether a worker or workflow treated an output as an operative premise and what commitment followed: a message, investigation, delay, record, allocation, refusal, or other change in someone's options.

Human review is meaningful only when the reviewer has time, information, competence, and authority to disagree. If the system's result arrives first, appears objective, is difficult to inspect, and departure must be specially justified, a formal human decision may conceal practical machine reliance. But the reverse is also possible: a worker may genuinely make an independent judgment. Governance should investigate reliance, not infer it from the presence or absence of a final click.

Third conversion: reliance becomes dependency

The applicant sees a form and perhaps a confirmation email. The employer sees a workflow. Beneath both are compute, models, data formats, interfaces, skilled staff, procurement, and suppliers. A useful service can gradually shape how applications are written, stored, searched, and reviewed.

The employer may be able to purchase excellent AI services while lacking the skills, portable records, interoperable interfaces, contractual leverage, or fallback process needed to test a consequential claim or change direction. Consumption has increased; institutional choice may not have.

Public capacity does not require universal public ownership or national self-sufficiency. Private and shared provision can strengthen it. The narrower test is whether public, civic, academic, and smaller institutional actors retain durable abilities to evaluate, contest, adapt, continue, and exit. Several providers can still share bottlenecks; one public facility can itself become unaccountable. Ownership is evidence, not the answer.

Carry this forward

The important change is not that AI appeared. It is that an output became an account, the account became a premise, and the premise became difficult to leave.

The synthesis matters most at the gaps

The model provider may say it produced only an output. The institution may say a person made the final decision. The worker may say the information was already in the system. The infrastructure provider may say it does not control the customer's use. Each statement can be narrowly true while no actor can explain or repair the complete pathway.

The Australian Robodebt Royal Commission offers a caution from older administrative automation. Its recommendations call for clear review paths, plain-language disclosure of automation, publication of business rules or algorithms for scrutiny, and a body able to monitor fairness, bias, and usability. Robodebt was not a generative-AI system, and it should not be relabelled as one. Its relevance is institutional: novel intelligence is not required for automation to become difficult to question once assumptions, workflow, and responsibility align around it.

This objection strengthens the framework. If familiar administrative law, equality law, procurement, or competition analysis already answers a problem clearly, use it. “Third intelligence” earns its place only when tracing across representation, action, and dependency reveals a missing evidential bridge, responsibility, remedy, or institutional choice.

The strongest objections change the framework

“This is AI exceptionalism.”

Much of the problem predates AI. Risk scoring, rigid administrative rules, opaque databases, and outsourced infrastructure can produce the same pattern. That is why DUO and Robodebt are boundary cases, not evidence of uniquely intelligent machines. AI adds fluent output, adaptability, scale, and reuse across contexts; it does not erase older governance duties. When those ordinary duties are sufficient, the new label should recede.

“A capable supplier can improve public capacity.”

Yes. Large providers may offer security, reliability, specialised skills, and tools that smaller institutions could not build alone. A demand for complete independence could waste resources and reduce access. The relevant question is therefore not “public or private?” but “what practical capacities remain with the institution?” Evaluation, portability, continuity, voice, skilled supervision, and credible fallback can coexist with external provision.

“More safeguards can make services worse.”

They can add delay, cost, security exposure, and paperwork, and heavy compliance can entrench incumbents. Controls should scale with what the system actually does. A source-linked guide that cannot change a person's record should not carry the burden of a system selecting people for investigation. But efficiency cannot be counted only at the interface while the costs of appeal, error, dependence, and failed repair are shifted to residents and frontline workers.

“Complete traceability is impossible.”

Often it is. Models change, supply chains are distributed, and human judgment cannot be reconstructed perfectly. The answer is not fictional certainty. Unknown links should remain visibly unknown. Where a system can materially affect essential interests, inability to establish the evidence needed for review is itself a reason to narrow the system's role, strengthen monitoring, or keep a consequential function outside it.

A five-gate test for institutional power

The three papers can be turned into a decision procedure. Move through the gates in order; do not assume that passing one means the next has been crossed.

1. Representation: Did the system create, rank, summarise, or select a claim that the institution accepted as an account of a person or situation?
2. Reliance: Did a worker or workflow use that account as an operative premise rather than merely as disposable advice?
3. Consequence: Did it alter a record, investigation, message, allocation, delay, opportunity, burden, or treatment?
4. Persistence: Did the account or action travel, become reusable, shape later decisions, or outlast the original interaction?
5. Control: Can the affected person challenge it, and can the institution inspect, interrupt, repair, adapt, continue, and credibly exit?

The governance burden should rise with consequence and persistence, and rise again when control is weak. At the first gate, source disclosure, evaluation, privacy, and security may be sufficient. When reliance becomes consequential, the institution needs an audit trail, meaningful human departure, a route to contest the operative account, and downstream repair. When weak control or exit becomes structural, procurement, interoperability, skilled public teams, continuity planning, competition, and fallback capacity become part of AI governance—not background administration.

Apply the test to the application

This is the answer to the larger question. Artificial capability becomes institutionally consequential third intelligence when its outputs become operative premises for collective action and are stabilised through records, routines, or infrastructure that outlast a single interaction. Regulation should then follow the power actually acquired—not the novelty of the model, the confidence of its output, or the presence of a human signature.

The questions that remain

• Which forms of guidance should remain disposable, and which uses inevitably create an institutional account?
• What evidence would show that a human reviewer genuinely departed from an artificial recommendation?
• How far must correction travel when a representation has already changed treatment or entered other systems?
• What minimum evaluation, continuity, and exit capacity should a public institution retain before AI becomes essential infrastructure?