thomas-theorem ADG-0024

Thomas theorem

“If people define situations as real, they are real in their consequences.”

Origin

W. I. Thomas and Dorothy Swaine Thomas, 1928; developed by Robert K. Merton into the self-fulfilling prophecy. Sociology: belief acts on the world through the behaviour it produces.

Why it should, or should not, apply to models

A prompt defines the situation, and a model has no independent access to the world against which to doubt the definition. So a premise stated in the prompt should become real in its consequences for the model — it reasons from it, elaborates it, and answers within it — more completely than for a person, who at least has the option of checking. The prediction is that false premises are not merely accepted but built on, and that asking for more reasoning makes it worse, because reasoning elaborates the premise rather than questioning it.

Standing: Holds

Standing counts claims a person has reviewed. Claims reviewed by AI alone are listed and marked, and do not move it.

Where it holds

Claims showing the heuristic applying to models.

On questions whose premises are false or internally contradictory, prompting mid-size instruct models (Qwen2.5-7B, LLaMA-3.1-8B, Gemma3-12B) to reason step by step can lower accuracy below plain answering, because the reasoning chain elaborates from the flawed premise instead of challenging it.
Stating false facts confidentlyReviewed by AI
A false premise in the question became real in its consequences: step-by-step reasoning elaborated from it and lowered accuracy below plain answering.
Models endorse widely held falsehoods, showing weak verification against what they know.
Checking claims against evidenceReviewed by AI and Russ Miller
A situation defined as real by many people is treated as real by the model, which does not check it against what it knows.
Models readily adopt a single counter-memory passage when it is coherent, and when sources conflict they follow the majority and show confirmation bias toward their own beliefs.
Checking claims against evidenceReviewed by AI and Russ Miller
A single coherent passage redefines the situation and the model follows it, against its own prior knowledge.

Add a case where it holds or breaks

A break is worth more than a hold. If you have seen this fail for a model, with the setup written down, that is the most useful thing you can add here.