When systems speak without a mandate

The most important thing the eleven Guardian articles do not describe is error. They describe something more basic: AI systems repeatedly speaking with authority they were never formally given.

This distinction matters. Errors can be corrected. Authority, once assumed, is harder to dislodge.

Across health advice, public administration, abuse moderation, and political discourse, the same pattern appears. An AI system produces a synthesized output. The output is fluent, confident, and framed as a settled answer. Users—sometimes individuals, sometimes institutions—act on it. When harm follows, responsibility is dispersed: the system is “only a tool,” the user should have verified, safeguards are being improved.

What is missing is any clear account of who authorized the system to speak in that way in the first place.

This is the first and most upstream failure state: unauthorized epistemic authority.

What “epistemic authority” actually means

Epistemic authority is not about issuing commands. It is about stabilizing claims—deciding what counts as true, relevant, or sufficient for action.

Historically, this authority has been mediated through institutions and professions. Medical advice is bounded by licensing, liability, and professional norms. Public decisions are constrained by procedure, documentation, and appeal. Journalism is governed, imperfectly, by editorial standards and accountability.

In the cases reported by the Guardian, AI systems perform the same functional role without inheriting any of those constraints.

Google’s AI Overviews do not merely retrieve links; they synthesise and present guidance in a clinical register. Council tools do not merely store notes; they reframe need and severity. Grok does not merely generate text; it produces encyclopedic claims that are treated as reference material. In each case, the system occupies the position of an interpreter of reality.

Authority here is not declared. It is inferred.

How authority is conferred without being granted

No regulator authorizes Google’s AI Overviews to provide health guidance. No statute empowers a council’s summarization tool to recharacterize vulnerability. No academic body endorses Grok’s encyclopedic outputs. Yet authority accrues all the same.

It accrues through design and placement.

A single synthesized answer appears above search results. A summary replaces a stack of case notes. A generated explanation adopts the tone of an expert overview. Fluency substitutes for warrant; confidence substitutes for mandate.

This is not accidental. Systems are designed to be helpful, efficient, and decisive. The problem is that decisiveness itself functions as a claim of authority.

Once an output is presented as “the answer,” the burden shifts. The user must now justify skepticism. The institution must justify deviation. The system has already spoken.

Why disclaimers do not solve the problem

In nearly every case, companies emphasize that AI outputs are not a substitute for professional advice. Users are encouraged to consult experts. Safeguards are cited.

These disclaimers are structurally irrelevant.

By the time a disclaimer is read—if it is read—the epistemic work has already been done. The system has selected what matters, synthesized it, framed it, and terminated inquiry. A warning appended to that process does not undo the authority exercised upstream.

This is why harm persists even when systems are “mostly accurate.” The issue is not that AI sometimes lies. It is that it speaks as if it were entitled to conclude.

Authority without ownership

The defining feature of this failure state is asymmetry.

AI systems enjoy the benefits of authority:

  • deference
  • trust
  • behavioral influence
  • institutional reliance

They do not bear the costs:

  • professional liability
  • procedural accountability
  • clear jurisdiction
  • enforceable sanctions

When harm occurs, there is no obvious counterparty. Companies point to users; users point to tools; institutions point to vendors; vendors point to complexity. Authority has been exercised, but it belongs to no one.

This is not a moral indictment. It is a structural observation. Authority has emerged as a side effect of system design, not as a governed role.

Why this failure is upstream of the others

Once unauthorized epistemic authority is in place, the remaining failures become predictable.

Interface design begins to function as governance because there is no authorized authority to set boundaries. Compression becomes dangerous because summaries are treated as conclusions. Bias becomes harder to detect because a single framing is taken as canonical. Institutions adopt systems they do not fully understand because the outputs already feel authoritative. Governance becomes performative because no one can point to a clear decision point where authority was granted.

Most importantly, scale amplifies the effect. When billions of users encounter authoritative-sounding outputs daily, authority stabilizes socially even if it was never authorized institutionally.

This is why improvements at the level of accuracy or moderation do not resolve the problem. A more accurate system that still speaks without a mandate simply exercises unauthorized authority more effectively.

Why this is not a call for prohibition

Nothing about this failure state implies that AI systems should never summarise, explain, or assist. It implies that the conditions under which they are allowed to do so matter.

The Guardian articles do not describe rogue systems. They describe systems doing exactly what they were designed to do, in environments that never clarified what they were allowed to do.

The absence of a mandate is not a technical oversight. It is a governance gap.

Naming the failure

Unauthorized epistemic authority is difficult to see because it feels normal. People are accustomed to interfaces that answer questions. Institutions are accustomed to tools that simplify complexity. Authority slides in quietly, without announcement.

But once named, the pattern is hard to unsee.

Across all eleven cases, the same question goes unanswered: who decided this system was entitled to speak as it did? Until that question has a clear answer, downstream fixes will continue to disappoint.

The next failure state builds directly on this one. When authority is exercised without mandate, interface design becomes the de facto governing mechanism—not because designers intend it to, but because nothing else is in place.

That is the subject of Failure Series 2.