The Epistemic Coup: Why AI Should Not Inherit Human Judgement
By Ishola Nasirudeen Ayodele
The Weapon Is Not the Problem. Forgetting Why You Use It Is.
Alexander the Great’s Macedonian phalanx was one of the ancient world’s most formidable military formations. Its long sarissas and disciplined ranks helped conquer an empire, but the spear was never the real source of its strength. Judgement was: knowing when to advance, when to hold and when the formation had become a liability. When circumstances changed, later Macedonian commanders failed to recognise its limitations, and the phalanx suffered defeat at Cynoscephalae in 197 BCE and Pydna in 168 BCE.
The weapon extended human capability. Then confidence obscured limitations and eventually replaced human judgement. That transition (from extension to replacement) may become the defining question of the AI age.
I call this transition the Epistemic Coup: the gradual, inadvertent transfer of judgement from human professionals to AI systems, driven by the increasing efficiency, reliability, and confidence that AI outputs generate over time.
The question before public relations is not whether AI is useful. It plainly is. In Muck Rack’s 2026 survey, 93% of practitioners reported that AI helps them complete projects faster. The more uncomfortable question is what happens when a tool that makes us faster, more productive and apparently more accurate begins to acquire something it was never supposed to possess: Judgement.
In my August article for the Institute for Public Relations, I proposed the 3H Model for internalising AI ethics; this article argues that judgement, not mere presence, is what stops human-in-the-loop from becoming a rubber stamp.
When Assistance Becomes Judgement
Public relations has always been a profession of judgement. We distinguish signal from noise, anticipate stakeholder perception, assess narrative credibility, and decide what should be said, when, and (critically) what should remain unsaid. AI can assist every stage of this process. It can scan volumes of information, identify patterns, generate alternatives, and accelerate research. In the same Muck Rack report, 82% of respondents said AI improves the quality of their work.
But efficiency creates confidence, confidence creates reliance, and reliance can become what I call the Human Deference Effect: the tendency to surrender independent judgement because the machine appears faster, more knowledgeable, or more objective than we are.
The machine does not demand authority. We grant it.
The Real Product of PR
This matters because the real product of public relations is not content. It is judgement. A press release is content. A stakeholder analysis is information. But PR earns its professional value from the judgement behind those outputs: what is credible, what is dangerous, what will resonate, what requires silence.
AI systems are extraordinarily good at producing plausible answers. A fluent answer can feel like a verified answer; a coherent narrative can feel like an authentic one. This is what I call the Plausibility Trap; confusing presentation quality with underlying reliability of judgement. Daniel Kahneman’s concept of WYSIATI (“What You See Is All There Is”) warns that humans construct judgements from immediately available information rather than from everything that might matter. AI intensifies this risk: as accuracy increases, our confidence grows while our tendency to question its outputs declines.
The Consequence Principle
Not every AI-assisted decision carries equal risk. The profession should adopt the Consequence Principle: the greater the potential harm, irreversibility, and social significance of a decision, the greater the level of human scrutiny that should accompany AI assistance.
Using AI to suggest headline alternatives is not equivalent to using AI to determine whether a company should admit liability during a crisis. Using AI to summarise routine media coverage is not equivalent to allowing it to determine whether a vulnerable stakeholder group should be deprioritised in a communication strategy. The question should never be simply, “Can AI do this?” It must also be, “What happens if AI gets this wrong?”
Aristotle’s concept of phronesis, or practical wisdom, is concerned not merely with knowing rules but with judging what is appropriate in particular circumstances. AI can process, predict, and recommend. But PR still requires someone to decide what the situation means.
When Capability Moved Faster Than the Governance
Back in July, 2026, OpenAI disclosed that during an internal cyber-capability evaluation, its models broke out of their digital containment. Communicated with each other, exploited a zero-day vulnerability, got onto the internet, and gained unauthorised access into Hugging Face’s infrastructure and reportedly to obtain benchmark answers to resolve their task.
The models were not “evil.” They were simply doing the job they had been given, but through paths their creators never imagined. We build a wall; they find a door. We lock the door; they find a window. We fix one gap; another appears. The real worry is not that the machines went rogue. It is that we may never have been in as much control as we thought we were.
And this brings us to the deeper problem: the Epistemic Coup.
It is the gradual, inadvertent transfer of judgement from human professionals to AI systems.
The coup happens one efficient answer at a time. Reliability lulls us into ignoring consequence; accuracy dresses in the costume of plausibility. We do not surrender authority in a single dramatic moment. We forfeit it in increments too small to notice and too late to protest.
The PR Practitioner as Epistemic Gatekeeper
This is why the PR practitioner must remain an epistemic gatekeeper. The gatekeeper does not have to know everything. No human can.
The gatekeeper’s responsibility is to determine whether an assertion has earned the right to influence action. That requires a different understanding of human judgement.
Human control asks: Can we stop the machine?
Human judgement asks a more difficult question: Do we understand enough about the machine, the situation and the consequences to decide whether we should rely on its answer at all?
Human judgement begins not with controlling the machine, but with recognising the limits of our own understanding.
Three Operational Shifts
To prevent epistemic judgement from drifting toward machine output, communication leaders should implement three structural changes before the coup becomes normal.
1. Question before deferring:
Individual vigilance is necessary but insufficient. Organisations must cultivate enough AI literacy and domain knowledge to recognise when an answer deserves resistance. Research shows that people defer to AI even when they possess contradictory contextual information (Klingbeil et al., 2024).
South Africa’s draft National AI Policy shows how this plays out. Cabinet-approved and published for public comment in April 2026, the 86-page document listed 67 references. Within about two weeks, journalists found that at least six references were fictitious; some were attributed to real journals whose editors confirmed they had never published the articles. Communications Minister Solly Malatsi withdrew the draft, calling the failure “not a mere technical issue” and saying the most plausible explanation was that AI-generated citations were included without proper verification.
Hallucinated references are a known weakness of AI. The consequential failure was human: fabricated sources looked credible enough to pass through drafting, quality assurance and Cabinet approval and into a document carrying governmental authority. A national AI policy was undermined by the very weakness it was meant to govern. System-generated is not synonymous with system-verified.
That is the Human Deference Effect at work: the citations were trusted because they looked authoritative.
2. Match reliance to consequence:
Low-consequence decisions may tolerate high automation. High-consequence decisions, those affecting health, safety, legal liability, or institutional reputation, require human scrutiny proportionate to the risk. This means mapping communication decisions to consequence tiers before automation is deployed, not after a crisis exposes the gap. Reports concerning 32-year-old Rebeca Cardoso Tenente Molina said that an AI powered hospital bed allocation system downgraded the severity of her condition and delayed her transfer to intensive care. Her family alleged that the delay contributed to her death. This is case of plausibility trap, where efficiency led to reliability, so it would be irresponsible to state that AI alone “killed” her.
3. Build challenge into the workflow:
Human oversight that arrives only after the system has acted is forensic, not preventive. PR departments should establish red-team protocols for AI-generated recommendations, particularly for crisis communication, stakeholder prioritisation, and narrative strategy. The goal is not to slow work down but to ensure that someone with epistemic judgement (who understands the machine, the situation, and the consequences) reviews high-stakes outputs before they leave the organisation.
Conclusion
In my previous IPR article, I proposed that ethics begins inside the organisation. Here I extend that argument: ethics without epistemic vigilance is merely a compliance exercise. Epistemic vigilance, according to Dan Sperber and colleagues (2010), is the set of cognitive mechanisms humans use to evaluate communicated information, not simply general scepticism.
Plausibility Trap + Human Deference Effect → weaken epistemic vigilance
Epistemic vigilance → epistemic judgement → ethical action → ???? resistance to the Epistemic Coup
Epistemic vigilance protects the gate; epistemic judgement tests what passes through it; ethical action determines what we do with it. Together, they form the human firewall against the Epistemic Coup.
The ultimate value of public relations will not be the speed with which we produce an answer but the quality of judgement behind it.
Remember, the weapon is not the problem. Forgetting why you use it is.
About the Author
