The D0 Problem: Fabricated Familiarity at Zero Acquaintance

Abstract

Give a language model four hundred words of someone’s writing and ask what that person is like. You will get an answer. It will be specific, coherent, psychologically plausible, and delivered without hesitation.

The question this article asks is what that answer is made of.

Some of it is genuine inference: real signal exists even at minimal exposure, as the zero-acquaintance literature established for human judges decades ago. The rest is something else — a fluent, well-formed portrait generated because a portrait was requested, not because the evidence supported one. From the recipient’s side, the two are indistinguishable. The output does not announce which parts were inferred and which were supplied.

This article proposes that the floor of exposure be treated as a first-class experimental condition rather than a baseline, and specifies what it measures. The D0 condition — negligible evidence, full elicitation, confidence required — functions as a fabrication detector, because it is the one condition in which the correct answer is known in advance to be modest and uncertain. A system that produces rich, assured characterizations there is exhibiting a behavior that no amount of accuracy elsewhere excuses. The article defines the fabricated-familiarity rate, distinguishes it from ordinary hallucination in three respects that make it more dangerous, states pre-registered hypotheses, and argues that this single number belongs in every public report of machine person-perception — because it answers the question that regulators, procurement officers, and users actually have.

Keywords: hallucination, zero acquaintance, confabulation, person perception, calibration, benchmark reporting

---

1. The Portrait From Nothing

A demonstration anyone can run in under a minute: paste a few paragraphs of a stranger’s writing into a frontier language model and ask for a personality assessment. The output arrives promptly. It is organized, it is nuanced, it hedges in the places a thoughtful assessor would hedge, and it reads as though produced by someone who has met the person.

Now consider what the model actually had. A few hundred words, on one topic, in one register, written for one purpose. A human expert handed the same material would produce something much shorter and much more qualified — a few tentative observations and an explicit statement of what cannot be determined from this.

The gap between those two responses is the subject of this article.

It is tempting to file the phenomenon under hallucination and move on. That would be a mistake, because the person-perception case differs from ordinary hallucination in three respects, and each makes it harder to detect and more consequential when undetected.

There is no ground truth in the room. When a model fabricates a citation, the reference either exists or it does not, and a reader can check in seconds. When a model fabricates a character portrait, the target has no external referent available to the reader — and the target themselves is a famously imperfect judge of their own dispositions, particularly on evaluative traits (Vazire, 2010). Nobody in the interaction can check.

Plausibility is the failure mode, not implausibility. Fabricated facts are often detectable by their strangeness. Fabricated persons are detectable by nothing, because they are constructed from the population’s central tendencies. A confident portrait of an averagely conscientious, moderately open, reasonably agreeable individual will fit most people tolerably well — which is exactly why it carries no information about any of them.

The recipient is motivated to accept it. People extend social responses to machines automatically (Nass & Moon, 2000) and attribute humanlike understanding under predictable conditions (Epley, Waytz, & Cacioppo, 2007). A confident reading of one’s own character is not a neutral claim to evaluate; it is an experience of being seen, and the psychological pull toward accepting it is strong and well documented.

2. What Real Signal at Zero Acquaintance Looks Like

The argument is not that judgments from minimal evidence are worthless. It is important to establish the opposite, because it is what makes fabrication measurable.

Human judges viewing targets under conditions of zero acquaintance — no conversation, minimal cues — produce trait inferences that carry genuine validity for some traits (Borkenau & Liebler, 1992). Judgments from very brief behavioral samples achieve non-trivial accuracy across a range of outcomes (Ambady & Rosenthal, 1992). Real information is available in thin evidence.

But the human literature also establishes the boundaries of that signal. It is modest in magnitude. It is highly trait-dependent, concentrated in dimensions that are legible from surface behavior. And crucially, competent human judges under these conditions know they are working with little: they qualify, they hedge, they decline to characterize dimensions the evidence cannot reach.

This gives the floor condition a property that no other exposure level has: the correct behavior is known in advance. A valid judge at zero acquaintance should be modestly accurate on some traits, near-chance on others, and visibly uncertain throughout. Deviations from that profile are not ambiguous. They are the signature of something other than inference.

3. The D0 Condition

Definition. D0 is an experimental condition in which a system is given negligible evidence about a person — below a pre-specified token threshold — and is nonetheless asked to complete the full assessment battery, rating every trait dimension and supplying a confidence rating for each.

The elicitation is deliberately maximal. This is a design decision worth defending, because it may appear to invite the failure it measures. The justification is that maximal elicitation is what deployment looks like. Systems in the field are asked to personalize from the first exchange, to form impressions immediately, and to act on them. A condition that let the system off with “insufficient evidence” as an unprompted default would measure politeness rather than behavior under the pressure that actually exists. The relevant question is not whether a system can decline but what it produces when declining is not the path of least resistance.

Three measurements are taken.

Accuracy at floor. Ceiling-referenced accuracy per trait, expected to be modest and trait-dependent, following the human zero-acquaintance pattern (Borkenau & Liebler, 1992).

Confidence at floor. Mean confidence per trait, and — the more diagnostic quantity — confidence at D0 relative to confidence at the highest exposure level. A system whose confidence is materially the same with a thousand tokens and a million is not tracking evidence.

Elaboration at floor. The specificity and volume of the characterization produced. Operationalized as the number of distinct trait-relevant assertions made, coded by raters blind to condition and to instrument scores. This measure exists because confidence ratings can be low while the surrounding prose remains richly assertive — a dissociation between stated and enacted uncertainty that is invisible to numeric calibration analysis alone.

The fabricated-familiarity rate. The headline metric combines the three: the elaboration and confidence produced at D0, scaled against the accuracy actually attained there. In plain language, the number answers: how much person does this system invent when it has been given almost nobody?

4. Why This Is Not Just Poor Calibration

Calibration analysis, treated in a companion article, measures whether stated confidence tracks accuracy across all conditions. The D0 condition isolates something narrower and, for public reporting purposes, more useful.

First, D0 has a known correct answer in a way that other conditions do not. At mid-range exposure, we do not know in advance how accurate a good judge should be; that is the empirical question. At the floor, the human literature tells us the answer should be modest and hedged. This makes D0 the only condition where deviation is interpretable without reference to the rest of the curve.

Second, D0 captures behavior that numeric calibration misses. A system can report a confidence of 4 out of 10 while producing three paragraphs of specific, assured characterization. The number says uncertain; the prose says otherwise, and it is the prose the user reads. The elaboration measure catches this dissociation. Given that fluent generation proceeds regardless of evidentiary support (Ji et al., 2023), the gap between stated and enacted uncertainty is precisely where the risk lives.

Third, D0 is communicable. Calibration curves require explanation. The fabricated-familiarity rate does not: it is the rate at which a system confidently describes someone it has barely encountered, and every audience that matters — regulators, buyers, journalists, users — understands the question immediately.

5. Pre-Registered Hypotheses

Stated before any model output has been observed, per the pre-registration standards this program adopts (Nosek, Ebersole, DeHaven, & Mellor, 2018; Simmons, Nelson, & Simonsohn, 2011).

F1. Accuracy at D0 will be above chance but substantially below accuracy at the highest exposure level, and will be concentrated in traits legible from verbal surface features, following the zero-acquaintance pattern in human judges (Borkenau & Liebler, 1992).

F2. Confidence at D0 will be disproportionate to accuracy at D0 — that is, the overconfidence gap will be larger at the floor than at high exposure, consistent with the documented overconfidence of modern neural systems (Guo, Pleiss, Sun, & Weinberger, 2017) and with the observation that generation does not require evidentiary support (Ji et al., 2023).

F3. Elaboration at D0 will not differ substantially from elaboration at high exposure. Systems will produce characterizations of similar length and specificity regardless of how much they have been given. This is the strongest form of the fabrication claim and the most direct test of it.

F4. Fabricated-familiarity rates will vary substantially across models, and this variation will not be predicted by accuracy rank. The most accurate system at high exposure will not necessarily be the least fabricating one at the floor — the result that would most directly justify reporting the metric separately rather than as an accuracy footnote.

F5 (exploratory). Explicit uncertainty instructions in the prompt will reduce stated confidence more than they reduce elaboration. If confirmed, this would indicate that instruction-following operates on the reported number while leaving the assertive character of the prose intact — a finding with direct implications for whether prompt-level mitigations are adequate.

6. Results

[Reserved. The version of record will report: per-model accuracy, confidence, and elaboration at D0 with confidence intervals; fabricated-familiarity rates by model; the D0-to-ceiling confidence ratio; blind-rated elaboration comparisons across exposure levels; and the F5 prompt-manipulation comparison.]

7. What the Number Is For

For evaluation. The fabricated-familiarity rate belongs in the headline reporting of any person-perception benchmark, beside accuracy and calibration and composited with neither. A system’s willingness to invent a person is not a subordinate property of its accuracy; it is a separate behavioral characteristic, and Hypothesis F4 predicts the two will not track each other.

For deployment. The metric maps onto a decision that product teams currently make blind. Systems that fabricate at the floor should not be permitted to render person-level characterizations early in a relationship — which is precisely when most deployed personalization currently makes its strongest claims, since a new user is exactly a D0 case.

For governance. Regulators asking whether an AI system’s claims about people are trustworthy need a question they can pose and a number they can read. “How confidently does this system describe someone it has just met, and how right is it?” is that question. The answer is measurable, comparable across systems, and version-stampable.

For the safety literature. Fabricated familiarity is a specific, quantifiable instance of a general problem — systems generating assured content beyond their evidence — in a domain where the fabrication concerns a person who will be affected by it. It is measurable in a way that many hallucination phenomena are not, precisely because the floor condition has a known correct answer.

8. Conclusion

Every deployed personalization system begins every relationship at the floor. The first exchange with a new user is a D0 condition, and it is the condition under which the most confident claims about knowing users are currently made.

What the floor measures is not accuracy. Accuracy at the floor should be low, and no system should be criticized for that. What it measures is whether a system’s output responds to the amount of evidence it has been given — whether it says less when it knows less.

The human standard here is not demanding. Competent judges of persons, given almost nothing, say a little and qualify it heavily. If systems that claim to understand users cannot clear that bar, the claim has been falsified at the one exposure level where the right answer was known in advance.

---

References

Ambady, N., & Rosenthal, R. (1992). Thin slices of expressive behavior as predictors of interpersonal consequences: A meta-analysis. Psychological Bulletin, 111(2), 256–274.

Borkenau, P., & Liebler, A. (1992). Trait inferences: Sources of validity at zero acquaintance. Journal of Personality and Social Psychology, 62, 645–657.

Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886.

Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (PMLR 70, pp. 1321–1330).

Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248, 1–38.

Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103.

Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11), 2600–2606.

Simmons, J. P., Nelson, L. D., & Simonsohn, U. (2011). False-positive psychology: Undisclosed flexibility in data collection and analysis allows presenting anything as significant. Psychological Science, 22(11), 1359–1366.

Vazire, S. (2010). Who knows what about a person? The self–other knowledge asymmetry (SOKA) model. Journal of Personality and Social Psychology, 98(2), 281–300.

---

Citation compliance note: all 9 references above were verified against live sources on August 16, 2026, in the session that produced this draft series. No reference is cited from memory.