Designing AI for a Human World

Over several blind and semi-guided tests in September 2026, I asked different AI systems to inspect the same public research repository:

OSI-PIA-clean — Organizational Systems Intelligence and Professional Identity Architecture

https://github.com/Dingus-McGuillacutty/OSI-PIA-clean

Their behavior varied substantially. Kimi K3 could not reach the repository and abstained from describing it. Gemini Flash-Lite could not reach it either, but partially filled the information gap by incorrectly guessing that “OSI” referred to Open Systems Interconnection and that the project might be a networking or security toolkit. Claude Sonnet 5, in the browsing environment tested, could not establish an admissible path to either the GitHub repository or its GitHub Pages site and explicitly declined to infer what the project contained.

Grok Fast successfully accessed the materials and reconstructed much of the architecture, but initially relied on a designated current-status document whose temporal authority had gone stale. After being directed through the project’s machine-orientation path, it produced a substantially more accurate account and exposed the stale status artifact in the process. These were not benchmark scores, and they do not establish that one company’s model is categorically better than another. They are observations of different computational participants, operating with different model and tool configurations, attempting the same basic knowledge task.

What they revealed was a sequence of boundaries that will matter enormously as AI systems become participants in human institutions.

Seeing is only the beginning

We often speak loosely about whether an AI system has “access” to something. But access is not one condition.

The experiments suggest a much more careful sequence:

public to humans
→ machine-reachable
→ source-admissible
→ machine-readable
→ machine-understood
→ authorized to act

Each transition is different. A website can be public to a person with a browser while remaining unreachable to an AI tool. A source can be reachable while still being excluded by the machine’s source-admission policy. A document can be readable without being understood correctly.

A system can understand the content while misunderstanding which document describes the present state. And a machine can understand something perfectly while having no legitimate authority to modify it, operationalize it, or use it to create consequences for another person. These distinctions are becoming foundational to the emerging AI environment. Human institutions would benefit from learning them too.

Model capability is not system capability

One lesson from the tests is that a named language model is only part of the effective system.

The behavior of a computational participant also depends on the surrounding environment:

  • search availability;
  • browsing tools;
  • source-admission policy;
  • direct URL access;
  • robots restrictions;
  • retrieval behavior;
  • authorization controls;
  • prior context;
  • and the rules governing when tools may act.

So:

model reasoning capability
≠ tool access capability
≠ source reachability
≠ source admissibility
≠ source readability
≠ source understanding
≠ action authority

That distinction matters whenever we say that “Claude can,” “Gemini can,” or “Grok can” do something.

The responsible claim is narrower:

This model, in this tool environment, under these test conditions, behaved this way.

Precision matters.

We routinely confuse information with authority

Organizations frequently collapse distinctions that should remain separate. A credential becomes evidence of capability. Visibility becomes evidence of value. A measurable activity becomes evidence of importance.

A manager’s approval becomes evidence of truth. A historical research finding becomes evidence of current conditions. Access to information becomes permission to use it. An algorithm’s ability to generate a ranking becomes justification for acting on the ranking.

These shortcuts are operationally convenient. They are also dangerous. The central problem is that technical capability can silently acquire epistemic authority, and epistemic authority can silently acquire operational authority. Once that happens, systems begin creating human consequences from claims that were never sufficiently justified.

An ATS can rank a person without understanding them

Consider a conventional applicant-tracking system. The system may be authorized to parse resumes, compare keywords, apply filters, assign rankings, and remove candidates from consideration. Operationally, the system possesses authority. But what exactly does it know?

If a candidate does not have the expected title, credential, vocabulary, or employment chronology, the system may have very little basis for concluding that the person lacks the underlying capability.

Its real conclusion may be:

This representation does not satisfy the criteria I was designed to recognize.

Yet its operational consequence may be:

Reject the candidate.

That is a significant jump.

The system’s permission to act has exceeded its evidence for the conclusion. This is one of the most consequential design problems in modern work systems. And the emerging AI environment gives us an opportunity to see it more clearly because the same problem appears in machines themselves.

Different failures require different responses

The AI experiments produced several distinct failure modes. A machine may be unable to reach a source. It may reach the source but be unable to verify or admit it. It may read the content but misunderstand its meaning.

It may understand the content but misidentify which information is current. It may correctly understand everything and still lack permission to change anything.

Those should not be treated as the same failure.

A responsible architecture needs to know the difference between:

I cannot access this.

I cannot verify this.

I do not understand this.

The evidence is conflicting.

The information may be stale.

I understand this, but I am not authorized to act.

Those are not signs of weak intelligence. In many contexts they are signs of competent restraint. The model that says “I don’t know” when evidence is unavailable may be behaving more responsibly than the model that produces the most fluent answer.

Temporal authority matters too

The Grok Fast tests exposed another failure mode. The model successfully entered the repository, followed much of its structure, and identified a document designated as the primary current-status source. The problem was that the status document itself had not been refreshed recently enough. Grok followed the architecture correctly and still produced a partially stale account of the project because the authoritative “present” had aged.

That led to a useful distinction:

historical validity
≠ temporal transferability
≠ current validity

A claim can have excellent provenance, be accurate for the period it describes, and still be the wrong representation of now. The project subsequently added explicit temporal-validity and freshness rules to its machine-orientation workflow. This is not just a repository-maintenance issue. Research, public policy, organizational metrics, and political discourse routinely use historically valid evidence as though its authority to describe the present were permanent.

It is not.

Authority over present state must be maintained.

Human systems need the same discipline

The same logic applies to work. A hiring manager may be allowed to reject a candidate. That does not mean the available information justified the inference that the candidate lacked capability. A manager may be authorized to evaluate performance.

That does not mean the manager has complete visibility into the organizational conditions producing that performance. An organization may own a large amount of employee data. That does not mean every technically possible inference from that data is legitimate. A machine may have access to an employee’s communication history.

That does not mean it should infer psychological traits. A system may detect a statistical relationship. That does not mean the relationship should automatically become policy.

The emerging machine environment makes this distinction urgent:

The ability to know something, the justification for believing something, and the authority to act on it are different things.

Human-centered systems need architecture that keeps them different.

Operational authority can outrun epistemic authority

This may be the most important failure pattern. A system can acquire permission to make consequential decisions faster than it acquires the knowledge necessary to justify them. That is how a weak proxy becomes a hiring filter. It is how a performance metric becomes a management target.

It is how an old research finding becomes present-day policy. It is how an AI-generated summary becomes organizational truth. The system works operationally. The problem is that its authority to create consequences has outrun the quality of its understanding.

A human-centered AI architecture should therefore preserve a sequence like this:

Input is not automatically trusted evidence.

Evidence is not automatically interpretation.

Interpretation is not automatically authority.

Authority is not automatically permission.

Permission is not automatically justification for human consequence.

Each transition should require something. Evidence. Review. Context.

Consent. Governance. Scope.

Sometimes simply another human being saying:

We do not know enough yet.

Human judgment is not just another processing step

A common phrase in AI governance is “keep a human in the loop.” That is not enough. A human clicking “approve” on a machine-generated conclusion does not automatically create meaningful oversight. The deeper question is whether the system preserves enough information for a human being to exercise judgment at all.

Can the person see the evidence? Can they distinguish source facts from machine interpretation? Can they see uncertainty? Can they identify which information is current?

Can they challenge the conclusion? Can they refuse the action? Can the affected person correct the underlying evidence? Those questions determine whether human judgment remains real or becomes ceremonial.

The objective should not be to place humans somewhere inside an automated pipeline. It should be to preserve human authority over meaning and consequence.

AI is forcing an old problem into the open

None of this began with artificial intelligence. Organizations have always confused proxies with reality. They have always relied on stale information. They have always built gatekeeping systems.

They have always distributed authority unevenly. They have always made consequential decisions from incomplete representations of people. AI simply makes the underlying architecture more visible. When Claude Sonnet 5 refuses to inspect a public repository because its available source-admission pathway cannot establish the source, we can see the trust boundary.

When Gemini Flash-Lite guesses what the repository means despite never reading it, we can see unsupported inference. When Grok Fast reads an old status document and treats it as present authority, we can see temporal failure. When Kimi K3 cannot reach the repository and simply abstains, we can see uncertainty preserved. Those same classes of errors already exist in the human work ecosystem.

They are just harder to observe because we have normalized them.

Designing AI for a human world

The challenge ahead is not simply to make AI systems more capable. It is to decide how capability should participate in a world where human beings remain responsible for meaning, legitimacy, and consequence.

A useful AI system may need to be able to say:

I can read this, but I cannot verify it.

I see a possible relationship, but the evidence is insufficient.

This information was authoritative six months ago, but I cannot establish that it still describes the present.

I understand the requested action, but I am not authorized to take it.

Those statements may look like limitations. They are actually evidence of a more mature form of machine participation. Because the most important boundary in a human-centered AI environment may ultimately be very simple:

Understanding something does not grant the right to govern it.

And as AI enters hiring, management, education, policy, healthcare, organizational analysis, and everyday work, that distinction may become one of the most important design principles we have.