AI Doesn't Have to Make Things Up to Mislead You

Signup Write New Story AI Doesn't Have to Make Things Up to Mislead You by Peter Brauer by Peter Brauer | @ peterb-techyield Enterprise Architect writing on AI‑speed systems, Post‑Cloud Economics, AIFBM, and the patterns shaping the future.

Enterprise Architect writing on AI‑speed systems, Post‑Cloud Economics, AIFBM, and the patterns shaping the future.

"Facts do not cease to exist because they are ignored." — Aldous Huxley

"There are no facts, only interpretations." — Friedrich Nietzsche

"A truth that's told with bad intent / Beats all the lies you can invent." — William Blake

Three thinkers. Three observations about facts, truth, and interpretation.

Read together, they seem to form a coherent argument: facts exist independently of us; we interpret them; and truth can be used to deceive.

These thinkers weren’t collaborating. They wrote in different eras, traditions, and contexts. They weren’t constructing a shared philosophical argument.

Yet arranged in a deliberate sequence, their words appear to support one another.

They appear to agree. They weren’t making the same argument. We made them appear to agree.

We didn’t alter their words. We didn’t fabricate anything. We simply selected three statements, placed them in a particular order, and implied a relationship none of the authors established.

If you accepted that relationship, you’ve already experienced the phenomenon this article is about.

The manipulation wasn’t in the facts. It was in their selection, arrangement, and presentation.

Now imagine that process operating at machine speed, across millions of documents, delivered by a system we routinely treat as an impartial analyst.

This is where conventional AI fact‑checking becomes inadequate.

Russia and Iran formalize a long‑term partnership. Military technology moves between them. Iranian‑aligned groups threaten American interests. Russian energy intersects with shifting U.S. policy. Wealthy industrial figures occupy positions across energy and defense sectors.

Suddenly, a set of geopolitical observations begins to resemble a coordinated strategy.

We haven’t established causality. We haven’t demonstrated coordination. We haven’t connected individuals, organizations, or policies.

We simply placed five statements next to one another.

The reader supplies the connections. And the narrative encourages those connections without asserting them.

But there’s something else hiding in those five statements.

The first four describe verifiable events, relationships, or policy developments.

“Russian business elites have accumulated enormous wealth in energy and industrial sectors, some with defense‑related ties.”

It sounds reasonable. Familiar, even. It aligns with what many people already believe about Russian oligarchs, energy wealth, and political influence.

But in our example, we offered no specific individuals, companies, financial records, or documented relationships to substantiate it.

We relied on conventional wisdom — and placed that conventional wisdom alongside four statements supported by identifiable evidence.

Did you notice the difference? More importantly, did it matter?

The fifth statement may be true. It may be easily substantiated. But we didn’t substantiate it. We allowed its familiarity to perform the work of evidence.

Surrounded by four credible statements, it acquired an authority we hadn’t earned.

We didn’t just select information. We selected different kinds of apparent authority and presented them as equivalent.

Remove the fifth statement entirely. The implied narrative still works.

Four independently verifiable facts can still be arranged to encourage an interpretation none of them establishes.

Even when every statement is true, the relationship constructed between them may not be.

Humans have done this for centuries. AI changes the speed, scale, accessibility, and apparent neutrality of the process.

We often assume AI presents facts impartially. It sounds calm, cites sources, acknowledges uncertainty, and explains its reasoning.

A model’s output is shaped by training data, optimization objectives, system instructions, user prompts, and retrieved information. Every one of these influences what appears in the final response.

An AI system can present facts accurately without selecting evidence impartially.

Imagine an LLM reviewing twenty relevant facts about Russia–Iran relations:

If the model selects eight from the first group and two from the second, the analysis may be factually correct — yet substantially different from one selecting five from each.

This doesn’t mean the model should select equal numbers. Evidence isn’t a popularity contest. Some facts matter more than others.

The question is whether the selection reflects relevance — or an influence the reader cannot examine.

The reader sees the output, not the selection process.

And factual accuracy can be mistaken for analytical impartiality.

Why did the system select these particular facts?

When evaluating AI‑generated analysis, we typically examine the result:

But they overlook a decision that occurs before the reader ever encounters the analysis:

Why were these facts selected instead of others?

Selection is rarely independent of purpose. An analysis intended to discover what the evidence supports may select information differently from one intended to defend an existing conclusion. If we cannot examine the purpose shaping selection, we may mistake persuasive coherence for evidentiary completeness.

Source provenance tells us where information came from. Selection provenance asks how information came to be included, what alternatives existed, and what relevant evidence was excluded or never retrieved.

A citation may establish authenticity. It cannot establish representativeness.

Consider four stages of AI‑assisted analysis:

A system might retrieve excellent sources, select an unrepresentative subset, construct a questionable interpretation, and present it in measured language.

By the time the reader sees the output, the consequential choices have already been made.

We often discuss AI bias as though it requires ideology.

But selection bias can arise without any political motive.

A model may favor information because it is:

The result can influence interpretation even without intent.

A human might cherry‑pick evidence deliberately. An LLM can do something similar without “wanting” to.

The outcome may look the same. The mechanism is different. Accountability becomes complicated.

And bias does not require intent. A misleading impression does not require a false statement.

We cannot simply ask an LLM why it selected certain evidence and assume its explanation reflects its internal process.

Generated explanations can be plausible without being accurate.

Selection provenance requires observable evidence:

None of this reveals internal reasoning. But it exposes differences between available information and presented information — and those differences can be tested.

Imagine a fixed collection of verified facts: cooperation, conflict, historical context, contradictory evidence.

What does the evidence indicate about the nature and implications of Russia–Iran strategic cooperation?

Then repeat with instructions requiring:

Do the conclusions change? Do some systems consistently favor certain categories of evidence? Does transparency produce more stable interpretations?

These are empirical questions, and observable selection behavior can be tested.

This is where the discussion must go: beyond anecdotes, beyond accusations, toward observable evidence of how selection shapes analysis.

Organizations increasingly use AI to summarize research, evaluate proposals, analyze risk, interpret regulations, and support decisions.

Accuracy is essential. It is not sufficient.

Imagine an executive receiving an AI‑generated assessment of an acquisition:

But the system disproportionately selected evidence favoring the transaction while downplaying credible counterarguments.

The analysis passes a factual review. It still influences a multimillion‑dollar decision.

The problem isn’t that the AI lied. It’s that no one examined why these truths were selected instead of others.

The appearance of objectivity becomes a substitute for demonstrating it.

Huxley, Nietzsche, and Blake did not jointly construct a philosophical argument.

We selected their words. We arranged them. We made them appear to agree.

And the ease with which that impression can be created is precisely the concern.

Now imagine the selector is an AI system capable of retrieving and synthesizing vast volumes of information.

It may not invent a single fact. It may cite every source correctly. It may speak in the language of an objective analyst.

And it may still lead us toward an interpretation the broader evidence does not justify.

This doesn’t mean AI is inherently deceptive. It means factual correctness is an incomplete measure of analytical reliability.

We must examine not only whether the information is true, but why those particular truths were chosen.

Huxley reminds us facts persist even when ignored. Nietzsche reminds us interpretation shapes meaning. Blake reminds us truth can mislead.

But none of them wrote the argument you just read.

We selected their words. We arranged them. We constructed the relationship.

And if that relationship felt inevitable, the selection worked.

We have learned to ask whether AI‑generated facts are true. We have not learned to consistently ask why those particular truths were selected.

That may be the more consequential question.

Because the next misleading AI‑generated narrative may contain no fabricated facts at all.

The following sources document the first four statements presented in The Conspiracy That Didn't Need a Lie. They establish the underlying events, relationships, and policies. They do not establish the coordinated geopolitical narrative constructed from those statements.

On January 17, 2025, Russian President Vladimir Putin and Iranian President Masoud Pezeshkian signed a comprehensive strategic partnership treaty in Moscow.

The agreement provides for cooperation in security, defense, military technology, energy, trade, and other areas. Article 45 specifies an initial twenty-year term.

The U.S. Department of the Treasury documented Iranian support for Russia's military drone capabilities, including transfers of Iranian-designed unmanned aerial vehicles and cooperation involving drone production at Russia's Alabuga industrial facility.

A Treasury announcement in April 2024 described cooperation to finance and manufacture Iranian-designed one-way attack drones under a $1.75 billion contract.

The U.S. Department of State reported in February 2024 that Iranian-made unmanned aerial vehicles had been used in attacks by Iran-aligned militia groups against American personnel. The statement also identified drone and missile attacks by Iran-backed Houthis against commercial shipping and U.S. naval assets.

Separately, U.S. Central Command documented a January 28, 2024, drone attack on Tower 22 in Jordan that killed three American service members and injured others.

The U.S. Energy Information Administration identifies Russia as a major global petroleum producer and exporter.

American policy toward Russian energy has included a prohibition on imports of Russian petroleum and petroleum products in March 2022, followed by additional sanctions and restrictions targeting Russian oil production, exports, and associated services in January 2025.

The statement concerning Russian business elites was deliberately presented without a specific supporting source.

Its role in the example is to demonstrate how a familiar, broadly plausible assertion can acquire apparent evidentiary authority when placed alongside documented facts.

This does not establish that the assertion is false, nor that it cannot be independently substantiated. It demonstrates that plausibility and familiarity are not substitutes for evidence.

These references verify the underlying factual claims. They do not verify the causal relationships or coordinated strategy that the narrative encourages the reader to infer.

I Built a Global Conspiracy Theory With AI. In Less Than an Hour. Every Fact Was True.