AI, History, and the Myth of the First Answer

Why hallucinations are not the real threat; and never were

A recent discussion on BBC Radio 4 raised familiar concerns about artificial intelligence and historical accuracy. Historian Tom Holland described asking an AI system about a poem connected to Jason and the Argonauts. The AI confidently attributed the poem to Pindar, an attribution that was incorrect. The conversation went on to suggest that AI risks infecting history with hallucinations, with novelist Kate Mosse cited as having experienced similar problems.

The error itself is not in dispute.
The conclusion drawn from it is.

Listen as A Spotify Podcast Discussion or view as a YouTube Video

---

The real mistake is not the hallucination

AI systems do misattribute sources. They confuse authors, compress traditions, and sometimes invent links that do not exist. This is a genuine limitation and should never be denied.

But the deeper problem lies elsewhere.

No historian would take the first book off a shelf and treat it as definitive.
So why do so many people expect the first answer from an AI system to be treated that way?

That expectation is not a technical failure, it is a human one.

---

History has never been pristine

Much of what we call “historical fact” was written:

• Long after the events it describes
• By authors with political, religious, or cultural incentives
• Through translations layered with misunderstanding and reinterpretation

Entire historical narratives have shifted because:

• A source was later shown to be biased
• A translation was corrected
• Archaeology contradicted a long-accepted account

Yet we do not say history itself is “hallucinating.” We say it is contested, provisional, and open to revision. That is not a flaw, it is the discipline working as intended.

AI has not introduced error into history.
It has merely removed the comforting illusion that our sources were ever error-free.

---

The double standard we apply to AI

If an undergraduate cited a single secondary source and stopped there, they would fail.
If a popular history book made an unsupported claim, reviewers would challenge it.

Yet when AI produces a first-pass answer, many people instinctively treat it as:

• Final
• Authoritative
• Definitive

When that answer turns out to be wrong, the blame is placed entirely on the machine.

This is a double standard.

AI is not replacing historical method, it is exposing how casually we abandon it when information arrives quickly and confidently.

---

Hallucinations vs inherited bias

When AI invents a citation, we call it a hallucination.
When a medieval chronicler reshapes events to please a patron, we call it history.

The difference is not accuracy.
The difference is distance.

Human errors are softened by time and authority.
AI errors are immediate, visible, and unsettling.

What truly alarms people is not that AI can be wrong, it is that it can be wrong with confidence, a trait it shares uncomfortably well with humans.

---

Deepfakes, hoaxes, and misplaced fear

The same misunderstanding appears in debates about deepfakes.

Fakes are not new.

• Propaganda predates photography
• Hoaxes predate film
• Fabricated UFO footage existed decades before AI

What has changed is ease of creation and speed of distribution.

That matters, but it does not remove responsibility from the audience. The solution has never been to ask who created the content, but to ask:

• Is it internally consistent?
• Is it supported by multiple independent sources?
• Does it align with established knowledge, or challenge it in a verifiable way?

These questions are not new. We have simply grown unused to asking them.

---

The human filter

Truth has always required what might be called a human filter:

• Context
• Scepticism
• Comparison
• Domain knowledge

When that filter is weak, errors spread, whether they originate from AI, books, lecturers, or broadcasts. When it is strong, even flawed tools become powerful allies.

AI works best as a starting point, not a conclusion. Used that way, it accelerates research rather than corrupting it.

---

What this moment is really about

This is not a crisis of artificial intelligence.
It is a crisis of epistemology, how we decide what is true.

We can respond in one of two ways:

• Strengthen our standards of evaluation
• Or abandon judgement and blame the tools

History does not become infected because machines make mistakes. It becomes infected when humans stop questioning authority, whatever form that authority takes.

AI has not changed the rules of knowledge.
It has reminded us, rather uncomfortably, that they were always there.

Latest AI Philosophical Articles