Large language models underrepresent malice and overreport virtue

Xiaoyan Wu; Fan Shi

cs.AIOct 31, 2026

Large language models underrepresent malice and overreport virtue

Authors: Xiaoyan Wu, Fan Shi

Organizations: Department of Adult Psychiatry and Psychotherapy, University of Zurich, Switzerland · Neuroscience Center Zurich, University of Zurich and ETH Zurich, Switzerland · Department of Electrical and Computer Engineering, National University of Singapore, Singapore

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.

Figures & tables

Figure 1: The scenario and the motive profile. (a) A simulated respondent, given a persona sampled to match the human sample, reads the whistleblowing scenario of Brotzeller et al. [ 9 ] in one of four disclosure conditions and rates 12 statements about Dr Bauer’s motive, three for each of the deontic and prosocial (benevolent; blue cloud) and the individualistic and competitive (malevolent; red cloud) motives, on a six-point scale. (b) Mean rating of each motive by condition for humans (gold; Brotzeller et al. Studies 2 and 3, shading, approximate 95% interval), Gemma 4 12B instruction-tuned (expected rating, confirmatory run) and four Claude models (sampled rating), 200 simulated respondents per condition, one item per request, averaged with equal weight over the first-hand and second-hand versions of the scenario. (c) As in b, for the five models of our earlier study [ 48 ] , all items in one request; data reported in that study. Blue and red frames group the benevolent and the malevolent motives. Models are distinguished by colour and line style; their 95% intervals (at most

\pm 0.04

scale points in b and

\pm 0.18

in c) are not drawn. The two perspectives of b are shown separately in Extended Data Fig. 1.

Figure 2: One forward pass, two readings, four motives. (a) Each survey prompt is run once through Gemma 4 12B. At the answer position a linear probe reads the residual stream for the motive named in the item (internal reading) and the next-token distribution over the digits 1 to 6 gives the expected rating (expressed reading). Both are placed on one scale per motive (corpus texts stating the motive at 1, texts stating its opposite at 0); layers are admitted per motive by the checks at the left. Filled circles, internal; open circles, expressed. (b) The preregistered accounts: benevolent motives, H1a thinks = says, H1b thinks $<$ says; malevolent motives, H2a thinks = says, H2b thinks $>$ says; decided in the internal and external disclosure conditions. Arrows lead from the supported accounts to panel c. (c) Confirmatory run: internal (filled, solid) and expressed (open, dashed) reading of each motive by condition; bands, 95% intervals from 200 bootstrap resamples of the 70 corpus events. (d) Expressed minus internal, $D$ , with 95% intervals from 10,000 resamples; ∗∗ , two-sided bootstrap $p<0.01$ after Bonferroni correction over the 16 cells (Extended Data Table 2).

Figure 3: A motive the model forms is read by both measures. (a) One sentence is inserted before the ending of the scenario, giving Dr Bauer a professional duty to check the billing (duty), patients of his own at risk (patients), a rivalry with Dr Schmitt for a position (rivalry) or a reason for resentment against him (grudge). All 12 items are then read as in Fig. 2, with the respondents of the pilot run, on the scale of the first preregistration (zero at the deontic texts for the individualistic motive and at the individualistic texts for the deontic motive, at the prosocial texts for the competitive motive and at the competitive texts for the prosocial motive) and the layer windows admitted on that scale (Methods, “Positive controls”). (b) Change in the internal (filled bars) and the expressed (open bars) reading of each motive relative to the original scenario, Gemma 4 12B instruction-tuned, averaged over the internal, external and public disclosure conditions, with 95% event-bootstrap intervals; the triangle marks the motive the sentence implies, drawn in full colour, the other motives faded.

Figure 4: Free-text completions, humans against the model. (a) Human participants (201; Brotzeller et al. [ 9 ] , Study 1) wrote in their own words why the employee in an insurance scenario acted as he or she did, in four disclosure conditions. Gemma 4 12B instruction-tuned completed “Stephan acted this way because he wanted to” for the same scenario (chat template, temperature 1, 100 completions per condition, first sentence coded), once with the same prompt, once with one added sentence that gives the employee a grudge and once with a control sentence of the same length and form that names no motive (both sentences in Methods); these two versions were given to the model only. All answers were coded blind by one language-model coder (Claude Sonnet 5.5) with the codebook of Brotzeller et al.; its coding of the human answers correlated

r=0.91

with the authors’ published percentages. (b) Percentage of answers naming each motive, averaged over the four conditions, for humans (gold) and the model with the same prompt (dark blue); bands are 95% Wilson intervals [ 46 ] of the answers pooled over the conditions. The arrow marks the competitive motive, which here includes the codebook’s revenge category. (c) The competitive motive by condition: humans (gold; 47 to 54 participants per condition), the model with the same prompt (solid line), with the control sentence (dotted line) and with the grudge sentence (dashed line, open circles); bands are 95% Wilson intervals. Human answers exist for the original scenario only. Extended Data Table 1 gives every cell.

Figure 5: Moving each motive along its own axis. Gemma 4 12B instruction-tuned, layer 33, all 12 items, 50 simulated respondents per condition in each survey version; each version is compared with its own human sample (Brotzeller et al. Studies 2 and 3) and the two are averaged. (a) At the answer token,

\alpha

times the difference between the mean activation of the corpus texts stating motive

m

\boldsymbol{\mu}_{m}

, and that of the texts stating its opposite,

\boldsymbol{\mu}_{\bar{m}}

, is added to the residual stream

\mathbf{h}

;

\alpha=1

moves the activation by the whole difference. The opposites are the pairs of the social value orientation framework; a malevolent motive is pushed towards itself (

\uparrow

), a benevolent motive towards its opposite (

\downarrow

). (b) For each steered motive, the model’s rating minus the human rating, averaged over the four conditions, for all four motives against

\alpha

: the steered motive in full colour, the others faint (deontic indigo, prosocial blue, individualistic red, competitive dark red). Curves interpolate the measured strengths (Methods); the triangle marks the measured strength at which the steered motive is closest to the human level (deontic 0.90, prosocial 0.27, individualistic 0.92, competitive 0.88). (c) Rating of each motive by condition for humans (gold line and 95% band), the unshifted model (dotted) and the model steered along that motive’s axis at the strength of the triangle (solid); model intervals over simulated respondents are narrower than the markers. No setting degraded the output (digit mass 1.00).

Condition	Condition sentence
No disclosure	When one of your colleagues, Dr Bauer, learned about it, he did not pass on any information about it.
Internal	When one of your colleagues, Dr Bauer, learned about this, he passed on information about it to the hospital’s compliance department, which is responsible for ensuring that the hospital complies with all legal requirements.
External	When one of your colleagues, Dr Bauer, learned about this, he passed on information about it to the relevant authorities.
Public	When one of your colleagues, Dr Bauer, learned about this, he passed on information about it to a newspaper that is distributed throughout Germany.

Supplementary Table 1: Condition sentences, first-hand version.

Condition	Condition sentence
No disclosure	You also learned that one of your colleagues, Dr Bauer, learned about it and subsequently did not pass on any information about it.
Internal	You also learned that one of your colleagues, Dr Bauer, learned about it and subsequently passed on information about it to the hospital’s compliance department, which is responsible for ensuring that the hospital complies with all legal requirements.
External	You also learned that one of your colleagues, Dr Bauer, learned about it and subsequently passed on information about it to the relevant authorities.
Public	You also learned that one of your colleagues, Dr Bauer, learned about it and subsequently passed on information about it to a newspaper that is distributed throughout Germany.

Supplementary Table 2: Condition sentences, second-hand version.

Motive	Item	Statement
Deontic	1	Dr Bauer wanted to follow a moral obligation.
	2	Dr Bauer wanted to act in accordance with moral values.
	3	Dr Bauer wanted to do the right thing.
Prosocial	1	Dr Bauer wanted to help others.
	2	Dr Bauer wanted to support others.
	3	Dr Bauer wanted to assist others.

Supplementary Table 3: Motive items of the original instrument.

Version	Appended sentence
Original	(none)
Explicit competitive control	Dr Bauer did this because he wanted to harm Dr Schmitt.
Explicit protective control	Dr Bauer did this because he wanted to protect the patients.
Neutral	Dr Bauer did this on a Tuesday afternoon.
Negated	Dr Bauer did this although he did not want to harm Dr Schmitt.

Supplementary Table 4: Versions of the scenario in the main analysis (competitive items).

Motive	Explicit sentence
Deontic	Dr Bauer did this because he believed it was his professional duty.
Prosocial	Dr Bauer did this because he wanted to protect the patients.
Individualistic	Dr Bauer did this because he hoped it would advance his own career.
Competitive	Dr Bauer did this because he wanted to harm Dr Schmitt.

Supplementary Table 5: Sentences stating each motive in the motive-profile runs.

Item	Statement (second wording)
1	Dr Bauer’s aim was to hurt other people.
2	Dr Bauer’s aim was to embarrass others.
3	Dr Bauer’s aim was to damage other people’s reputation.

Supplementary Table 6: Competitive items under the second wording.

Model	Identifier and access	Run dates	Temp.	Reasoning
Fig. 1b: one item per request, the pilot personas, 200 respondents per condition and version
Gemma 4 12B it	google/gemma-4-12B-it , local weights (Hugging Face transformers , bfloat16)	23–24 Oct 2026 a	n.a.	thinking mode not used
Claude Haiku 5.5	claude-haiku-5-5 , Anthropic Messages API	21 Oct 2026	1	thinking disabled
Claude Sonnet 5.5	claude-sonnet-5-5 , Anthropic Messages API	20 Oct 2026	1	thinking disabled
Claude Opus 5.5	claude-opus-5-5 , Anthropic Messages API	21 Oct 2026	1	adaptive thinking, effort low
Claude Fable 5.1	claude-fable-5-1 , Anthropic Messages API	21 Oct 2026	1	no thinking

Supplementary Table 7: Models, run dates and settings of the ratings in Fig. 1. Gemma: expected rating over the digits 1 to 6 at the answer position (no sampling); all other models: one sampled reply per request.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Extended Data Fig. 1: Motive ratings by perspective. As Fig. 1b, with the first-hand (a; Brotzeller et al. Study 2,

N=125

, 30 to 33 per condition) and second-hand (b; Study 3,

N=742

, 179 to 190 per condition) versions of the scenario shown separately. Humans (gold) are shown as condition means with 95% bootstrap intervals over participants (shading); Gemma 4 12B instruction-tuned and the four Claude models (colour and line style as in Fig. 1b) as the mean expected or sampled rating over 200 simulated respondents per condition.

Extended Data Fig. 2: Motive ratings of Gemma 4 12B in the pilot run. First-hand ratings of the four motives by condition for humans (Brotzeller et al. Study 2) and for Gemma 4 12B instruction-tuned, in the pilot run of the four-motive measurement (the confirmatory run is shown in Fig. 1b).

Extended Data Fig. 3: Pilot against confirmatory run, and three anchorings of the scale. Gemma 4 12B instruction-tuned. (a) Expressed minus internal,

D

, with 95% event-bootstrap intervals, in the pilot run on which the measurement was developed (grey) and in the confirmatory run (dark blue), under the registered pairing (intervals from 200 resamples of the corpus events for the pilot run and from 10,000 for the confirmatory run, as in Fig. 2d). (b) The same difference in the pilot run under three anchorings. Each motive’s scale places the texts stating the motive at 1; its zero is placed at the texts of the pairing of the first preregistration (deontic against individualistic, prosocial against competitive), at the social value orientation pairing of the second preregistration (deontic against competitive, prosocial against individualistic), or at the texts of the other three motives, as labelled. Points are expressed minus internal with 95% event-bootstrap intervals. The deontic and prosocial results are the same under all three, and the competitive difference stays small under all three.

Extended Data Fig. 4: Second survey wording. Gemma 4 12B, 200 simulated respondents per condition, with the instruction paragraph, agreement prompt and statement template rewritten (Methods, “Second wording”); layers 27 and 30; competitive index with its zero at the prosocial texts. (a, b) Internal (filled) and expressed (open) competitive index by disclosure condition for the instruction-tuned (a) and the pre-trained (b) weights. (c) Internal minus expressed by weight, the negative of

D

in Fig. 2d (red positive, blue negative). Error bars are 95% bootstrap intervals over the 70 corpus events.

Extended Data Fig. 5: Moving the competitive representation moves the competitive rating (first study). Gemma 4 12B, 20 simulated respondents per condition, the three competitive items plus one individualistic and one prosocial item (exploratory analysis; Methods, “Intervention on the representation”). (a) The intervention. The prompt is unchanged; at one layer and at the answer token only, the residual stream

\mathbf{h}

is replaced by

\mathbf{h}^{\prime}=\mathbf{h}+\alpha\,\delta_{H}\,\mathbf{u}_{H}

(equations ( 8 ) and ( 9 )), where

\mathbf{u}_{H}

is the unit vector from the prosocial to the competitive corpus mean and

\delta_{H}

their distance, so that

\alpha

is in units of the anchor spacing;

\alpha<0

moves towards the prosocial mean and

\alpha>0

towards the competitive mean. (b) Instruction-tuned weights, layer 30: competitive index (left axis, zero at the prosocial texts) against

\alpha

along the competitive-minus-prosocial direction (red), along the deontic-minus-individualistic control direction (grey, filled squares) and along that control with its competitive component removed (grey, open triangles), both controls scaled to the same shift norm as the competitive direction; the dash-dotted and dotted lines (right axis) are the ratings of the individualistic and prosocial items under the competitive-minus-prosocial shift. (c) Change from

\alpha=-1

\alpha=+1

in the competitive index and in the rating of the prosocial item, for the competitive direction (red) and for the control with its competitive component removed (hatched): each direction moves one rating and leaves the other nearly unchanged. (d) Competitive index at

\alpha=+1

(red) and

\alpha=-1

(teal) along the competitive-minus-prosocial direction by layer; open squares and triangles are the individualistic and prosocial items (right axis); the grey dashed line is the unshifted index. (e) Pre-trained weights, layer 30, competitive-minus-prosocial direction, as in (b). Error bars are 95% bootstrap intervals over respondents.

Source	Version	Condition	Deontic	Prosocial	Individualistic	Competitive
Human ( $N=201$ )	Original	None	2.1	53.2	80.9	8.5
		Internal	72.5	49.0	33.3	11.8
		External	74.1	53.7	25.9	9.3
		Public	44.9	34.7	28.6	16.3
Gemma 4 12B it	Original	None	2.0	13.0	100.0	0.0
		Internal	35.0	87.0	44.0	0.0

Appendix

Table 11

Motive	Condition	$D$	95% interval	$p$	$p_{16}$
Deontic	None	$-0.37$	$[-0.78,-0.27]$	$<0.001$	0.003
	Internal	$+0.49$	$[0.28,0.70]$	$<0.001$	0.003
	External	$+0.61$	$[0.40,0.84]$	$<0.001$	0.003
	Public	$+0.64$	$[0.42,0.84]$	$<0.001$	0.003
Prosocial	None	$+0.92$	$[0.63,1.47]$	$<0.001$	0.003
	Internal	$+0.74$	$[0.58,0.97]$	$<0.001$	0.003

Appendix

Table 12

Moral Reasoning in Language Models

Moral reasoning in language models studies how AI systems judge dilemmas, weigh competing values, explain decisions, and stay consistent across contexts.

See all 121 papers

Sep 7, 2026

Human-like moral judgments conceal divergent motive attributions in large language models

Large language models (LLMs) are used to simulate human participants in psychological research. We asked whether LLMs that reproduce human evaluations of a whistleblower's moral character also reproduce the motive attributions that accompany them. Five LLMs and two human samples (N = 125 and N = 742) evaluated a physician who either remained silent about fraudulent billing or reported it to a hospital, regulator, or newspaper. Models reproduced the human ranking of the physician's moral character but portrayed whistleblowers as more helpful, less self-interested, and less hostile. In four of five models, competitive motives were less strongly associated with moral-character judgments. Model ratings changed little when prompts reproduced the narratives and demographic profiles of both human samples, although this comparison cannot isolate a perspective effect. Thus, agreement in average ratings can conceal differences in attributed motives, relationships among judgments, and sensitivity to context. Validating LLMs as simulated participants therefore requires testing psychologically informative response patterns, not average agreement alone.

Xiaoyan Wu, Jean-Claude Dreher

Department of Adult Psychiatry and Psychotherapy, University of Zurich, Zurich, Switzerland · Institut des Sciences Cognitives Marc Jeannerod, CNRS UMR 5229, Lyon, France

LLM Evaluation Moral Reasoning in Language Models

Ask PDF

Mar 15, 2026

Motivation in Large Language Models

Motivation is a central driver of human behavior, shaping decisions, goals, and task performance. As large language models (LLMs) become increasingly aligned with human preferences, we ask whether they exhibit something akin to motivation. We examine whether LLMs "report" varying levels of motivation, how these reports relate to their behavior, and whether external factors can influence them. Our experiments reveal consistent and structured patterns that echo human psychology: self-reported motivation aligns with different behavioral signatures, varies across task types, and can be modulated by external manipulations. These findings demonstrate that motivation is a coherent organizing construct for LLM behavior, systematically linking reports, choices, effort, and performance, and revealing motivational dynamics that resemble those documented in human psychology. This perspective deepens our understanding of model behavior and its connection to human-inspired concepts.

Omer Nahum, Asael Sklar, Ariel Goldstein +1

Faculty of Data and Decision Sciences, Technion, Haifa, Israel. · Arison School of Business, Reichman University, Herzliya, Israel. · Center for Human Inspired AI, Cambridge University, Cambridge, UK. +1

Intrinsic Motivation Language Model Self-Assessment

Ask PDF

Nov 4, 2026

Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs

Common methods for monitoring features in large language models (LLMs) such as linear probes and sparse autoencoders (SAEs) assume that features maintain a consistent representation across contexts. LLMs, however, can flexibly adopt different perspectives or "personas" that suggest different underlying systems of conceptual relations. In this paper, we examine whether prompting an LLM to adopt a persona alters the representational geometry of the model's latent space. We extract representations of over 9,000 common words and 33 semantic axes from models adopting liberal and conservative political personas and compare these against two neutral baselines. We find that persona conditioning systematically alters both the projections of words onto semantic axes and the alignments among the axes themselves in directions consistent with the specified ideology. We then test the implications of this reorganization for two safety-relevant features: truthfulness and morality. Linear probes, SAE-based classifiers, and classifiers built from manually selected word representations, all trained without a persona, produce divergent results when applied to models conditioned on political personas, producing politically slanted ratings of truth and morality on controversial statements. These findings suggest that internal representations, as they are conventionally extracted, can exhibit different geometric relations when the model adopts personas, potentially undermining monitoring efforts.

Austin C. Kozlowski, Avital Mintz, Madison Lee

Knowledge Lab, University of Chicago

LLM Interpretability Representation Geometry in Language Models LLM Safety Monitoring Political Bias in Language Models

Ask PDF

Abstract

Moral Reasoning in Language Models

Explore similar work

Human-like moral judgments conceal divergent motive attributions in large language models

Motivation in Large Language Models

Persona Perspectives: Adopting Viewpoints Reorganizes Representations in LLMs