Kurate.org
Rankings
Back to Rankings

Large language models underrepresent malice and overreport virtue

Xiaoyan Wu, Fan Shi

Oct 31, 2026arXiv:2611.01332v1
cs.AIcs.CY
Share
Scorecard· 16/16
6.5/10 impact

Careful, preregistered paired probe-versus-report design that localizes a known LLM bias, but limited to one open model, one vignette, and probe/anchor-dependent inferences.

Abstract

People increasingly use large language models (LLMs) to interpret others' behaviour, often assuming that LLMs infer motives as humans do. Aligned LLMs attribute more benevolent and less malevolent motives to a whistleblower than humans do, but ratings alone cannot reveal whether these differences arise in internal representations or at report. Here we compare linear-probe readouts of LLMs' internal states with their own ratings at the same prompt position for four motives, in two preregistered studies. For the two malevolent motives, self-interest and intent to harm, internal and expressed readouts did not differ: the lower ratings reflect weaker signal along the probed directions, not suppression at output. For the two benevolent motives, duty and concern for others, expressed ratings exceeded the internal readouts, indicating amplification at report. In free text the model rarely volunteered a motive against the colleague, far less often than humans did, yet produced one readily once the scenario supplied a reason. Shifting the internal representation along a malevolent-motive direction brought both malevolent ratings to the human level, whereas the rating of duty yielded only past a threshold at which the other motives had overshot. LLMs' more charitable motive profile thus has two distinct sources: reduced malevolent signal in the probed representation and amplified benevolent attribution at report.

AI Impact Assessments

(1 model)

Core Contribution

The paper asks why aligned LLMs attribute more benevolent and less malevolent motives to a whistleblower than humans do. It separates two accounts: the model represents less, or it represents the motive and suppresses or amplifies it at report. Linear probes for four motives (deontic, prosocial, individualistic, competitive) are trained on a 70-event corpus of stated motives. The probes are read at the answer token of a survey prompt, and the same forward pass yields the expressed rating as the expected value over digit probabilities. Both readouts are put on one anchored scale.

The headline findings, in Gemma 4 12B-it:

  • Malevolent motives show no internal-versus-expressed gap. The weak signal is internal, not suppressed at output.
  • Benevolent motives show expressed ratings above internal readouts, so there is amplification at report.
  • In free text the model rarely volunteers a competitive motive (0–1% vs 8–16% in humans) but readily does so when given a grudge.
  • Steering along the motive axes brings the malevolent ratings to human levels smoothly. The deontic rating moves only past a threshold, where the other motives overshoot.
  • The behavioural profile replicates in four Claude models and in five models from an earlier study.

    • The novelty is the same-token, same-scale comparison of probe readout and behavioural report, used to localise a behavioural bias. This is a sensible use of interpretability for LLM-as-simulated-participant concerns.

    Methodological Rigor

    Strengths:

  • Two preregistrations, with competing hypotheses stated up front (H1a/b, H2a/b), a pilot/confirmatory split, frozen code and fixed layer-admission rules.
  • Event-level bootstrap (10,000 resamples) that refits the probes and anchors, which is the right uncertainty source.
  • Multiple validity checks: explicit-motive versions, negated and neutral controls, cue sentences, a second wording, three alternative scale anchorings, and a pre-trained-weights comparison.
  • Steering controls with random directions and equal-norm control directions.
  • Human validation of the corpus (κ = 0.97) and a free-text analysis with a control sentence.

    • Weaknesses:

  • The central inference rests on a linear probe's scale. "Internal = expressed" depends on anchor choice and on the assumption that corpus-derived directions capture the model's motive representation. The authors acknowledge that a near-zero reading means no linear signal on this direction, not no information.
  • The probe corpus is a different distribution (short explicit-motive texts) from the scenario. The internal reading for implied motives may be systematically attenuated, which could mimic "weak representation" or exaggerate "amplification at report". The internal score for implied duty or prosocial motive is low, but this may reflect a probe-specificity issue (explicit stated motive vs inferred), and the paper partly concedes this.
  • Several null findings (competitive and individualistic) rest on intervals that include zero, so they are absence of evidence rather than equivalence tests. The individualistic public condition is borderline (+0.29, uncorrected p = 0.006).
  • It is a single 12B open model. The frontier-model replication shows only the behavioural profile, not the mechanism.
  • The steering uses class-mean differences, which move valence rather than specific motives, so the "causal" claim is blunt. The probe direction itself proved unusable for steering.
  • Free-text coding uses a single LLM coder. The coder over-assigned deontic codes relative to the authors', though the same coder was used for both sources, so the comparison is consistent.
  • One scenario, a German whistleblowing vignette.

    • Potential Impact

    The main audience is LLM-as-participant researchers, computational social scientists, and the alignment/interpretability community. The practical message is that a "charitable" bias has two sources needing different remedies (output calibration vs representational change). The framing of internal-versus-report dissociation for social attributions is reusable, and the paired-readout design could be applied to other biases such as sycophancy and stereotypes. The empirical scope is narrow, so the dissociation result for benevolent motives may be partly an artefact of how a probe trained on stated motives treats inferred ones. The impact is therefore moderate: a clean, careful methods-plus-finding paper rather than a field-shifting one.

    Timeliness

    Highly relevant. LLMs as simulated participants and judges is a hot topic, and concerns about alignment-induced distortions and the faithfulness of reports are current. The use of preregistration in interpretability work is unusual and welcome.

    Strengths

  • A tidy conceptual separation of "represents less" from "reports less/more".
  • Preregistered, with pilot-to-confirmatory consistency.
  • Extensive controls and transparency (corpus, code and data to be released, though currently only on request).
  • A convergent behavioural, probe, free-text and steering picture.

    • Limitations

  • One model and one scenario for the mechanistic claims.
  • The linear-probe and anchor dependence of "internal".
  • Steering does not cleanly isolate motives.
  • Dense, heavily jargon-laden writing, with unusual features such as preregistration deviations and many nested checks; the signal gets somewhat lost.
  • The conclusion "malice is weakly represented" could reflect base-model priors about the scenario, with little insight into where it arises (pre-training vs post-training). The pre-trained comparison was uninformative.

    • Overall

    A solid, careful, moderately novel contribution. It is likely to be cited in LLM-simulation and moral-psychology-of-AI discussions but is unlikely to be a field-defining primitive. The dependence on one 12B model and a specific probe design limits generalisability.

    Rating:6.5/ 10
    Significance 6Rigor 7.5Novelty 6.5Clarity 5.5

    Generated Nov 3, 2026

    Comparison History (0)

    No comparisons yet.