When AI trusts authority: Who will question the evidence?

Human perception, institutional failures, and the danger of turning official narratives into artificial intelligence’s version of reality

How much of what we believe is actually true?

Have you ever wondered how easily we can be wrong about another person?

We meet someone, exchange a few words, observe their behavior, and form an impression. Sometimes we trust them immediately, and sometimes something about them makes us uncomfortable.

We might tell a friend, “I don’t think that person is trustworthy.” Our friend agrees. Another person shares a similar impression, and eventually, what began as an opinion becomes something everyone seems to know.

But what if everyone is wrong?

What if the person being judged has done nothing wrong? What if the agreement between those people is not evidence, but simply the repetition of an interpretation?

Now imagine that the person expressing the opinion is a manager, a doctor, or someone with institutional authority, and their judgement enters an official report. Personal impressions become professional assessments, words are carefully chosen, documents are signed, and conclusions begin to influence real decisions.

If the report appears objective, who verified the original interpretation? And what happens to the person affected when everyone involved believes the report?

This is not simply a question about human behaviour. It is a question about how societies decide what is true.

We rely on institutions, professionals, and official records because no individual can independently investigate everything. Yet those institutions are operated by people, and people can misunderstand, overlook, misinterpret, or misrepresent reality.

Now we are developing artificial intelligence systems capable of processing enormous amounts of human-produced information, including scientific publications, medical records, institutional assessments, government reports, and professional opinions. These documents may contain more than facts; they may also contain human interpretations of facts.

If human perception is sometimes wrong, what happens when artificial intelligence learns to reproduce it?

And what happens when the interpretation comes from an authority that everyone has been taught to trust?

The doctor who challenged medical authority

In the nineteenth century, Hungarian physician Ignaz Semmelweis investigated why so many women were dying from puerperal fever after childbirth.

Working at Vienna General Hospital, he observed that maternal mortality was substantially higher in a maternity clinic attended by doctors and medical students than in another attended by midwives.

In 1847, he introduced handwashing with a chlorinated lime solution before patient examinations, and mortality fell dramatically.

Today, the importance of hand hygiene in preventing infections is well established. Semmelweis’s findings were rejected by much of the medical establishment. Despite the dramatic reduction in maternal deaths, he was criticised and dismissed by influential colleagues who questioned his conclusions. Amid professional conflicts, he lost his position in Vienna, and his discoveries remained largely unaccepted during his lifetime.

In 1865, he was admitted to a psychiatric institution, where he died shortly afterwards at the age of 47. His contribution gained broader recognition after his death.

Now imagine an artificial intelligence system evaluating his findings in 1847. Would it have given greater weight to his observed results or to the prevailing medical consensus?

When evidence and authority disagree, what should determine an AI system’s judgement?

Romania’s Hexi Pharma scandal: What would AI have believed?

In 2016, Romania was confronted with revelations about diluted disinfectants supplied by Hexi Pharma to healthcare institutions.

Investigative reporting exposed serious problems involving product concentrations, effectiveness, and the systems responsible for verifying the quality of disinfectants used in hospitals.

Questions also emerged about earlier inspections, regulatory oversight, and whether warning signs had been adequately addressed.

These revelations came after the Colectiv nightclub fire of 30 October 2015. The fire killed 27 people that night, and another 37 subsequently died.

Many survivors suffered severe burns and required intensive hospital treatment. In the aftermath, reports concerning hospital-acquired infections, antibiotic-resistant microorganisms, and the treatment of patients transferred abroad raised serious concerns about conditions within the healthcare system. The Hexi Pharma revelations made the question of hospital infection prevention even more urgent.

This essay is not an attempt to reconstruct every medical outcome or determine individual responsibility for particular deaths.

It asks a question: What would an advanced AI system have concluded about Romanian hospital safety before the scandal became public?

Imagine returning to 2015. An AI system is asked to assess the safety of disinfectants and infection-prevention procedures in Romanian hospitals. It receives official documentation, inspection findings, product information, and available institutional reports. Some documents may contain reassuring conclusions, while others may contain warnings, inconsistencies, or incomplete information.

Would it trust official assurances?

Would it identify contradictions between reported compliance and available findings?

Would it investigate whether apparently independent reports relied on the same information?

Would it recognise missing evidence?

Or would it generate a polished assessment concluding that existing safety procedures were adequate?

Imagine the consequences if such a report were then used to reassure patients, healthcare professionals, or decision-makers. The report might be convincing, it might cite multiple official sources, and it might appear more objective because it was produced by artificial intelligence.

Could AI become an additional layer of authority over information that was never adequately verified?

Final thoughts

Artificial intelligence could help humanity identify institutional failures, challenge unsupported assumptions, and uncover problems that have remained unnoticed for years. But it could also reproduce human mistakes, giving unreliable information greater credibility and spreading those errors on an unprecedented scale.

The challenge is not to build AI systems that blindly trust or automatically reject authority, but systems capable of evaluating evidence, recognizing uncertainty, and questioning conclusions when necessary.

The cases discussed above illustrate how human decisions and interpretations can be mistaken, misleading, deliberately falsified, or, in some circumstances, entirely justified. They remind us that neither professional authority nor institutional recognition guarantees the accuracy of a conclusion.

Existing institutional failures cannot simply become the responsibility of artificial intelligence systems. At the same time, these failures should not prevent technological progress or discourage us from exploring how AI could contribute to more reliable assessments, greater transparency, and improved decision-making.

Yet this raises a far more difficult question: When will human decision-making become sufficiently accountable to prevent avoidable harm? If we fail to address the weaknesses in how existing institutions investigate errors, challenge authority, and resolve systemic problems, what will happen when those same failures can be reproduced, legitimised, or concealed on a much larger scale through artificial intelligence?

We need stronger regulations, but not only for AI. We also need effective mechanisms of accountability, independent oversight, and meaningful consequences for human and institutional failures. Otherwise, we risk building increasingly sophisticated technologies on foundations that remain fundamentally unchanged.

Written by the author with AI assistance for language editing and refinement.



Leave a Reply