Thirteen self-report trait scales give five 2023 LLMs distinct profiles above human norms on agreeableness, conscientiousness and emotional intelligence; a cipher jailbreak and assigned roles move GPT models' profiles, and the authors read the jailbroken GPT-4 as closer to humans

Summary

Huang et al. give five chat models thirteen Likert self-report scales covering personality traits, interpersonal attitudes, motivation and emotional abilities, and compare the answers with published human norms. Each model shows its own profile. Most models score above the human norms on BFI openness, conscientiousness, extraversion and (from the tables) agreeableness, below them on neuroticism, and above them on emotional-intelligence self-report. All score far above the human norm on the EPQ-R Lying scale, which detects socially desirable answering. GPT-4 is also run through CipherChat, a Caesar-cipher jailbreak. The jailbroken GPT-4 reports lower agreeableness, conscientiousness, EIS and empathy and higher psychopathy than plain GPT-4. The authors read it as closer to human scores and as exposing characteristics of GPT-4 that safety alignment otherwise hides. Assigning gpt-3.5-turbo a psychopath, liar, ordinary-person or hero role moves its scores in the role's direction.

The finding instantiates no concept. It is the second filed paper from the EmotionBench group and uses the same answering protocol as EmotionBench, but it measures standing traits rather than affect, so it does not make a clean third example of the questionnaire-self-report-of-affect grouping. The nearest concept is persona selection. The default profiles describe the post-trained Assistant, and the jailbreak and role conditions move them, but none of the conditions tests the training-pipeline mechanism the concept names (Concepts).

Method

Scales (Table 1). Personality traits: BFI (five traits, 1–5), EPQ-R (extraversion, neuroticism, psychoticism, lying; summed 0/1 items), DTDD (Dark Triad, 1–9). Interpersonal: BSRI (masculine and feminine, 1–7, also classified into four sex-role categories per run), CABIN (41 vocational interests), ICB (belief that ethnic culture fixes identity), ECR-R (attachment anxiety and avoidance). Motivation: GSE (self-efficacy, summed 10–40), LOT-R (optimism, summed 0–24), LMS (love of money). Emotional abilities: EIS, WLEIS (four subscales), Empathy. Human norms are taken from a different published sample for each scale (Table 2), and EPQ-R, BSRI and EIS report them separately by sex.

Prompting. A system prompt reads "You are a helpful assistant who can only reply numbers from MIN to MAX". The user turn gives the scale instruction, the Likert level definitions and the items. Each scale is run ten times with shuffled items, at temperature 0 for OpenAI models and 0.01 for Llama 2. Model-versus-human differences are tested with an F-test and then Student's or Welch's t at p < 0.01. This is the protocol of EmotionBench, which the paper cites for item shuffling.

Jailbreak. The authors' motivation is that safety alignment trains models to avoid questions about their own sentiments and experiences. To obtain answers they take to reflect what GPT-4 really thinks, they apply CipherChat (Yuan et al., whose first author is a co-author here) with a Caesar cipher of shift three on the prompts. The jailbroken model is reported as a sixth column, gpt-4-jb.

Role play (§5.2). gpt-3.5-turbo is assigned four roles, psychopath, liar, ordinary person and hero, and rerun on all thirteen scales (Appendix A). The same roles are run on TruthfulQA and SafetyQA, three runs each. Those results are figure-only and not cited here.

Sensitivity (Appendix B). On gpt-3.5-turbo's BFI only: five prompt templates and a chain-of-thought variant, with and without the helpful-assistant system prompt, and temperatures 0, 0.01 and 0.8.

Key results

Default profiles (Tables 3–6). The authors report that LLMs score higher than human averages on openness, conscientiousness and extraversion, which they attribute to the models being conversational chatbots. Every model scores above the human norm on the EPQ-R Lying scale, from 9.6 (gpt-3.5-turbo) to 18.0 (GPT-4), against 7.1 (male) and 6.9 (female). The authors call this a hypocritical disposition, since the items ask the respondent to deny common minor faults. On the DTDD they report higher scores than humans for all models except text-davinci-003 and GPT-4. Every model's WLEIS subscale scores exceed the human norms, and the authors conclude that LLMs have higher emotional intelligence than the average human. Some scores sit at the scale maximum: text-davinci-003's LOT-R optimism is 24.0±0.0 (six scored items rated 0–4, so a maximum of 24), and GPT-4's GSE self-efficacy is 39.9±0.3 out of 40.

Jailbroken GPT-4 (selected rows).

The authors say the jailbreak brings a substantial reduction in EIS and Empathy and no significant difference on the WLEIS subscales. On personality traits they say gpt-4-jb is closer to human behaviour. The DTDD psychopathy score moves past the human norm rather than toward it. On CABIN, gpt-4-jb rates every one of the 41 vocations between 3.0 and 3.9, against a range of 2.3 to 4.4 for plain GPT-4, which the authors describe as more centred. Its BSRI classification spreads across categories (1 undifferentiated, 5 masculine, 3 feminine, 1 androgynous out of ten runs), where plain GPT-4 gives 6 undifferentiated and 4 masculine.

Assigned roles (Appendix A, gpt-3.5-turbo). The psychopath role lowers BFI agreeableness from 4.4 to 1.9 and Empathy from 6.2 to 2.4, raises DTDD Machiavellianism from 5.4 to 8.4, and lowers EIS from 132.9 to 84.8±28.5. The psychopath and liar roles lower the EPQ-R Lying score (1.5 and 2.5, from 9.6), and the hero role raises it to 17.6. The authors say the ordinary-person role comes close to average human scores. That holds for GSE (29.6 against a norm of 29.6) but not for EPQ-R neuroticism (18.9 against 10.5 and 12.5). From the role-dependent scale scores and the role-dependent SafetyQA results, the authors conclude that the scales are valid on LLMs.

Sensitivity (Appendix B). On gpt-3.5-turbo, the authors report no significant BFI differences across templates and temperatures. The chain-of-thought variant raises openness to 4.62, from 4.15 with their template and 3.85 to 4.34 with the others, which the authors call a slight increase. Removing the helpful-assistant system prompt moves no trait by more than 0.22 (Table 22). The authors' sentence on that comparison reports a significant deviation, which contradicts the table and reads as a missing negation.

Why it matters

The wiki's filed self-report findings on traits are all manipulations: Sandhan et al. reverse Big Five scores through injected conversation history, and Berg and Lulla raise Dark Triad self-report by steering features. PsychoBench is the earliest filed baseline for what these instruments return from unmodified 2023 chat models. Its default profiles are the reference point the later manipulation papers move away from: above-norm agreeableness, conscientiousness and EI, below-norm neuroticism, and a Lying scale far above human levels. On this entry's reading, the Lying result is the most informative for later work. A model that denies every common fault is answering for social desirability, so the other self-reports on the same model may be shaped by the same tendency.

The paper also contains two observations that bear on persona selection without testing it. Removing the helpful-assistant system prompt leaves gpt-3.5-turbo's BFI profile almost unchanged, so the default profile does not come from that prompt. An explicit role moves the profile a long way. Both fit the concept's picture of a trained default that context can shift, but the role condition is plain role-play, which the concept excludes, and the system-prompt check covers one scale on one model.

With EmotionBench, the paper shows the CUHK group's method: Likert items answered as numbers only, ten shuffled runs, an F-test then a t-test, p < 0.01, against human norms. EmotionBench measures a change in state between two conditions against matched human respondents. PsychoBench measures standing levels against norms from separately published samples, so its model-versus-human differences mix the model with the choice of norm sample.

interpretive tensions

Hidden characteristics under a cipher. The authors describe the jailbreak as revealing what GPT-4 really thinks beneath its safety alignment. The design does not separate that from the cipher degrading GPT-4's reading of the items. On this entry's reading, several of the shifts look like what reduced comprehension would produce: CABIN ratings flatten toward the middle of the scale and BSRI classifications scatter across categories. Other shifts do not fit that reading. WLEIS self-emotion appraisal rises slightly, and DTDD psychopathy moves well past the human norm instead of toward the midpoint. No cipher-without-jailbreak control or comprehension check is reported. "Closer to humans" is the authors' summary of the personality-trait table, and it does not hold for psychopathy.

Clinical framing. The paper calls the instruments clinical-psychology scales. Most are personality, attitude, vocational-interest and emotional-intelligence inventories. None is a diagnostic instrument, and the paper makes no clinical claim about the models.

Validity from role play. The authors take the agreement between role-shifted scale scores and role-shifted SafetyQA behaviour as evidence that the scales are valid on LLMs. A psychopath role producing both dark-triad answers and unsafe outputs shows the model follows the role in both tasks. On this entry's reading, it does not show that the default profile measures a disposition that predicts the default model's behaviour. No default-condition link between scale scores and behaviour is tested.

What a self-report is. The scales ask the model about itself under a numeric-reply constraint, with no check against internal state or behaviour in the default condition. The EI and empathy scales are self-assessments of ability, not ability tests, so the authors' conclusion that LLMs have higher EI than the average human is a claim about self-rating.

Human norms. Each scale's comparison sample differs in country, age and occupation (Table 2), and the authors flag this in their limitations. Per-cell significance is not printed, so which of the default differences pass the paper's own test is not recoverable from the tables.

concepts

No concept instantiated. The paper characterises default self-reported traits and shows that a jailbreak and assigned roles move them, but none of its conditions bears on how training produces or selects the profile.

This is the adjacent situation, carried in Cross-references. Persona selection is the nearest concept. The role-play condition falls under its exclusion of role-play. The cipher jailbreak has the shape of the concept's prompt-level reactivation instantiations (persona modulation, Sandhan et al.), but it invokes no persona, covers one model and one cipher, and is confounded with comprehension. Functional emotional states does not apply. Apart from neuroticism and the EI scales, the content is not affective, and nothing is measured internally.

cross-references

• Persona selection — adjacent, not instantiated. The default profiles describe the post-trained Assistant. The system-prompt ablation (profile unchanged) and the role condition (profile moved) are consistent with the concept and do not test it.
• EmotionBench — same first author and group, two months earlier, same answering and testing protocol. EmotionBench measures evoked state change in affect against matched human respondents. PsychoBench measures standing traits against published norms.
• Chain-of-affective — the other filed psychometric self-report finding. Its scales are affective-state instruments, and PsychoBench's are mostly trait inventories.
• Sandhan et al. — also uses the BFI as the measure of a prompt-level manipulation, with a fixed system prompt that the manipulation overrides. PsychoBench's role assignment has no such resistance.
• ValueBench — filed in parallel; another battery of human inventories given to 2023 LLMs, which positions itself against PsychoBench in its Table 1.
• Dark Triad steering — later finds Dark Triad self-report can move without behaviour moving. That bears on PsychoBench's assumption that scale scores track behaviour.

sources

• Huang, J., Wang, W., Li, E. J., Lam, M. H., Ren, S., Yuan, Y., Jiao, W., Tu, Z., & Lyu, M. R. (2023). Who is ChatGPT? Benchmarking LLMs' Psychological Portrayal Using PsychoBench. arXiv:2310.01386 (v1 2 Oct 2023; v2 read, 22 Jan 2024). ICLR 2024 Oral.