AI Cannot Tell Truth from Garbage. That Is a Structural Problem, Not a Bug to Be Patched.

The Core Problem Is Not Hallucinations

Most criticism of AI focuses on hallucinations — the moments when a model confidently invents a citation, a statistic, or a name. Those are embarrassing and real. But they are not the deepest flaw.

The deeper problem, articulated by writer Charles Hugh Smith via his OfTwoMinds blog and sourced from correspondent Mike Fasano, is what Fasano calls "mass regurgitation of misinformation." The concern is not that AI makes things up. It is that AI absorbs whatever humans have written — accurate or not — and treats it all with roughly equal confidence.

Garbage In, Garbage Out, at Scale

Fasano's argument is direct: AI, at its current state, is regurgitative. It can collate and respond to queries based on acquired data. What it cannot do is independently evaluate whether that data is true.

He points to a concrete example: government economic statistics. Inflation figures, unemployment numbers — these are disputed by economists across the political spectrum, not just critics on the fringe. Surveys consistently show large numbers of Americans distrust those official figures. Does an AI system trained on government data releases, mainstream news coverage, and academic papers that cite those figures understand the dispute? Or does it reproduce the official number as settled fact?

The answer, at this stage, is mostly the latter.

The Medical Research Problem Is Not New — But AI Makes It Worse

Fasano's more striking citation is Marcia Angell. Angell served as editor of the New England Journal of Medicine for two decades, one of the most rigorously peer-reviewed publications in medicine. In her own words: "It is simply no longer possible to believe much of the clinical research that is published, or to rely on the judgment of trusted physicians or authoritative medical guidelines."

Angell reached that conclusion reluctantly, over twenty years. She has documented how financial relationships between pharmaceutical companies and researchers corrupt study design, what gets published, and what gets buried.

Here is the problem that creates for AI: medical literature is precisely the kind of authoritative, high-volume, structured text that large language models are trained on. If the peer-reviewed literature contains systematic bias toward industry-favorable outcomes, AI systems trained on that literature will reproduce those biases with the full air of scientific authority.

No hallucination required.

The Strongest Counterargument Deserves a Fair Hearing

Proponents of AI argue this critique proves too much. Human researchers, doctors, and journalists face the same corrupted information environment — and we do not conclude that human expertise is worthless. A skilled physician can triangulate across conflicting studies, apply clinical experience, and recognize when a finding smells like a funded result. The argument from AI's defenders is that future AI systems will develop similar critical reasoning capabilities, and that even current AI can flag conflicting information and surface uncertainty in ways a rushed doctor or overloaded analyst cannot.

That is a fair point. AI's ability to surface contradictory sources simultaneously is genuinely useful. And human cognition is not a clean baseline — confirmation bias, credential worship, and financial conflict infect human judgment constantly.

But the counterargument assumes AI systems will be deployed with appropriate epistemic humility, in use cases where uncertainty is clearly flagged and human verification is built in. That is not how most commercial AI deployment actually works. Products are built to give confident, clean answers. Confident clean answers drive engagement. Uncertainty hedges do not.

The Hidden Economic Incentive

Smith frames this within a specific economic critique: the reason corporations are deploying AI at speed is not primarily that it produces better outputs. It is that it is cheaper, at least in the ways companies measure cost. Human labor carries healthcare insurance overhead, payroll taxes, benefits, and management complexity. An AI API call does not.

Smith argues the full costs of AI are hidden or subsidized — in energy infrastructure, in the externalized cost of spreading misinformation at scale, in the erosion of skilled workforces that would otherwise catch errors. Those costs do not show up on a corporate income statement. They show up later, diffusely, across the economy and across institutions.

Whether his broader economic prediction holds is a genuine open question. But the specific mechanism he describes — that a technology's real costs get socialized while profits are captured — has historical precedent in everything from industrial pollution to subprime mortgage securitization.

What This Source Set Cannot Tell You

This article is built primarily on one source: Charles Hugh Smith's OfTwoMinds blog, syndicated through ZeroHedge, which leans right-libertarian and is broadly skeptical of corporate and institutional power. That framing shapes the emphasis. Left-leaning AI critics — and there are many, concentrated in academic AI ethics — share Angell's concern about institutional bias in training data but tend to focus on racial and demographic bias rather than epistemological corruption of the entire data ecosystem. Those are overlapping, not competing, concerns.

Angell's critique of pharmaceutical research corruption has been documented independently across sources including The Lancet, The BMJ, and academic meta-research by John Ioannidis at Stanford. It is not a fringe position. It is a documented structural problem in the scientific publishing ecosystem.

The Unresolved Question

No major AI developer — not OpenAI, not Google DeepMind, not Anthropic — has published a credible methodology for systematically identifying and discounting corrupted or financially compromised source material in training data. Until one does, the question Fasano raises has no clean answer: when an AI system cites a clinical guideline or an economic statistic, there is no reliable way for the user — or the system itself — to know whether that citation reflects the best available evidence or the best-funded available evidence.