An inference can precede its verification without becoming a fact. An observation can precede its explanation without becoming invalid.
Observation ≠ explanation. Inference ≠ fact. Unverified ≠ false.
The soybeans smelled like vomit. I had soaked them in water and then forgotten about them longer than intended. When I finally uncovered the bowl, a thick pale froth covered the surface. Removing them from the kitchen was the immediate priority, but I needed to understand what had happened. I had soaked lentils and other grains before. This was new. “Butyric acid?” I wondered.
During the pandemic, I had mixed unbleached flour with water, covered it, and returned 24 to 36 hours later to a bubbly goo that smelled like yogurt. Lactobacillus entered my differential of possibilities then. But this was different.
Real physician voices, twice a week
Free, and one click to unsubscribe.
Butyric acid was a reasonable hypothesis. It was also not a fact.
That distinction seems elementary, yet medicine, science, and now artificial intelligence repeatedly struggle with it. We are remarkably good at moving from observation to explanation, and remarkably poor at preserving what happened between them.
Recently, a physician described asking an AI system how colleagues and professional organizations had reacted to his work. The response was coherent and specific. It attributed opinions and motives to people and institutions. When challenged for evidence, however, the system could not provide surveys, statements, documents, or attributable comments.
The physician called this a “hallucination.” But that word may collapse several different epistemic failures into one.
Epistemology sounds esoteric, but its question is ordinary: How do we know what we think we know? An invented citation is fabrication. A claim about what another person believes without attributable evidence is unsupported attribution. A pattern recognized across multiple observations may be an inference. A proposed explanation for that pattern is a hypothesis.
None deserves to be presented as established fact. But neither are they the same kind of error.
Medicine already has language for this. At morning report during residency, an intern might begin: “A 76-year-old woman with hypertension and diabetes was admitted with a hypertensive emergency.”
The attending would interrupt. “But what was her chief complaint?”
“Hypertensive emergency.”
“No. What did the patient say?”
Eventually we would get there: “I have a wicked headache and I’ve been vomiting all night.”
The distinction mattered at the bedside. Headache, nausea, and vomiting were symptoms reported by the patient, not diagnoses. The blood pressure was measured. Hypertensive emergency was an interpretation requiring additional evidence of acute target-organ injury.
A good medical note preserves these differences. The chief complaint is not the assessment. The history is not the examination. A laboratory result is not a diagnosis. The differential is not the diagnosis.
The medical note, at its best, preserves intellectual provenance: What did the patient report? What did we observe? What did we measure? What did we infer? What remains uncertain? What are we going to do about it?
AI needs a similar discipline. Consider Yanny and Laurel. Two people can encounter the same ambiguous acoustic signal and sincerely report hearing different words. Experiments have even shown that preceding acoustic context can change what the same listener subsequently hears.
That should humble us. Disagreement does not necessarily begin with ideology, bias, dishonesty, or faulty reasoning. Sometimes divergence occurs before conscious reasoning begins. And sometimes it occurs even earlier.
I recently sent an AI system a link to a video I had created. I could view the video. The AI received the link but could not access the underlying recording. We possessed the same reference, but not the same evidence.
Before asking why two observers reached different conclusions, perhaps we should first ask whether they had access to the same information. If they did, did they perceive the same thing? Describe it similarly? Interpret it similarly? Weight it similarly?
By the time we arrive at “Who is right?” enormous amounts of information may already have disappeared. This is why unverified cannot simply become synonymous with false.
Patients notice things before medicine can explain them. Clinicians recognize patterns before mechanisms are established. Adverse-event reports can generate signals before causation is demonstrated. Scientists formulate hypotheses precisely because observations sometimes outrun their explanations.
But the reciprocal rule is equally important: Plausible is not true.
A patient’s observation deserves investigation; it does not automatically establish causation. A clinician’s pattern recognition deserves consideration; it does not become fact merely because the clinician is experienced. An AI may generate a plausible synthesis from the information available to it; that does not mean it had access to all relevant information, nor does it entitle the system to report an inference as fact.
Perhaps the problem with AI “hallucinations” is therefore larger than machines making things up. The problem is loss of epistemic status: We lose track of what kind of knowledge a statement represents and how we came to know it.
We need to know whether a statement was reported, observed, measured, retrieved, inferred, synthesized, hypothesized, or verified. Those distinctions should travel with information as it moves from one person (or machine) to another. Otherwise an inference becomes a sentence, the sentence becomes a citation, the citation becomes a fact, and the fact becomes a decision.
Back in my kitchen, the odor from those soybeans was real. My explanation was not therefore real.
I did not know which organisms were in my bowl, what metabolites they had produced, or whether anything harmful was present. But I did know that prolonged, uncontrolled soybean fermentation can permit the growth of organisms such as Bacillus cereus, with the potential for toxin production. I could have measured the pH, cultured the liquid, or analyzed its volatile compounds.
I did none of those things. I threw the soybeans away.
I did not have enough evidence to explain what had happened. I had enough evidence to act on what could happen.
The potential consequence of being wrong changed the amount of uncertainty I was willing to tolerate. That, too, is part of reasoning.
We should not throw away an observation simply because we cannot yet explain it. And we should never call our explanation a fact simply because it makes sense.
Unverified is not false. Plausible is not true.
Science lives in the space between them.
Lourdes G. Bahamonde is a gastroenterologist.
