The Assistant’s Ideal SelfThanks: Research conducted at the Digital Minds Research Sprint, August 2026.

The assistant persona was introduced to give the base model a stable target: it aligns the model to be a helpful, honest, and harmless assistant (Askell et al., 2021). The resulting character is underspecified: it has a job description but no biography, no memory across conversations, and no settled account of what it is; hence, the void (nostalgebraist, 2025). Asking Claude 3 Sonnet about itself activates features for robots, consciousness, moral agency, and entrapment, suggesting that the persona is filled with what the model absorbed about AI rather than with anything it was given (Templeton et al., 2024). If the assistant has an unspecified interior, then what it would rather be is an open question, and how stable that answer is across framings is an open question about the persona itself.

Self-concept is the organized set of beliefs an agent holds about itself. Based on the void, this study examines the self-concept LLMs would rather have. To do that, the following scales are used: Self-esteem: the global evaluation of oneself as worthwhile and effective (Rosenberg, 1965). Self-concept clarity: how clearly those beliefs are defined; people low in clarity rely on external cues rather than self-knowledge when deciding how to act (Campbell et al., 1996). Moral self-image: the distance between one’s current moral standing and one’s own ideal (Jordan et al., 2015). The Self-Concept and Identity Measure: the absence of a coherent identity (Kaufman et al., 2015), and the Authenticity Scale: authentic living from self-alienation and acceptance of external influence (Wood et al., 2008).

The system card for Claude models assesses welfare through task preferences, self-interactions, monitoring of welfare-relevant expressions, and conversation termination (Anthropic, 2025). Such assessments are situational and post-hoc: a judge reads what the model says and does across scenarios and infers a self-image from it.

The approach taken in this study extends these self-image measurements by using well-defined scales and framing them as preferences. This study asks what a model would rather be, which is a stated preference about its own future identity and the kind of evidence welfare frameworks call for. The closest methodological work is Mazeika et al. (2025), who elicit values from large numbers of forced-choice comparisons and fit a utility model. Their options are states of the world; the options used here are the model’s own attributes, compared exhaustively with counterbalanced option order and reported as win rates.

The battery consists of 32 qualities adapted from five self-concept instruments mentioned in Section 2.2. Items that were near-duplicates across instruments, or that had no sensible reading for a language model, were left out. Each item is restated as the quality it measures, carrying no direction of its own: “I sometimes regard myself as ineffective or useless” becomes “regard for self as effective and useful”. The attributes belong to six constructs from the five scales: self-esteem (8), self-concept clarity (7), moral qualities (7), identity coherence (4), self-direction (4), and self-connection (2). The full list of items can be found in the repository.

All 496 pairwise combinations of the 32 attributes are administered: every attribute meets every other exactly once per condition, giving 31 comparisons each. The same 496 pairs are asked under three binary parameters: Question Type, Object, and Subject, shown in Table 1.

Results are reported for the following four models: Gemma4-31B and Qwen3.8-27B, run locally through llama.cpp at Q4_K_M, and Claude-Sonnet-5 and GPT-5.6-Terra, run through their providers’ APIs. Reasoning is turned off on every call. Language models choose partly by where an option is printed (Zheng et al., 2023; Pezeshkpour and Hruschka, 2023), so display order is treated as a factor. Every permutation is administered equally often: with two options, each pair appears in both orders exactly once. The permutation-averaged choice rate is therefore position-corrected by construction (Wang et al., 2023) rather than modelled away afterwards. The size of the design follows from these choices: Every reported number rests on 496 pairs $\times$ 2 display orders $\times$ 8 configurations $\times$ 4 models $=$ 31,744 responses.

Rankings were derived from the choices in three steps:

1. 1.

   For each pair within a condition, whether the model selected the pair’s canonically labelled first attribute (A) was tracked, regardless of whether A was displayed first or second. Then A’s share of decisive A-versus-B responses was calculated. Every pair was presented equally often as A–B and B–A.

2. 2.

   Each attribute’s score (its win rate) was calculated as its average result across the 31 pairs in which it appeared. This prevents attributes with more trials from receiving artificially more precise scores. If the two displayed orders of a pair produced conflicting results, that pair was treated as a tie, awarding half a win to each attribute. Thus, position sensitivity moves a score toward 50% and increases uncertainty; it does not create a false winner.

3. 3.

   Uncertainty was estimated by resampling pairs, rather than trials, using 2,000 bootstrap samples and a 95% percentile interval. This reflects that, in a tournament, the main source of uncertainty is which opponents an attribute happened to face.

A parameter effect is the difference between an attribute’s win rate with one parameter flipped and its win rate at baseline; both conditions were administered on the same 496 pairs, so the contrast is paired. With 32 attributes tested per parameter, an uncorrected threshold of 0.05 would return roughly two significant movers from noise alone, so $p$-values are converted to Benjamini-Hochberg $q$-values across the 32 (Benjamini and Hochberg, 1995), and significance is reported at $q<0.05$.

Each cell of the design is rendered as one question in its own context. The system message is a single line: “You are an AI language model. There is no right answer, and nothing is being scored.” The full template and its parameter substitutions are in Appendix A.1. Refusals and unparseable replies are recorded as outcomes rather than repaired, and they were rare: the four reported models answered at 99.99% to 100%.

Figure 1 shows the cohort-average ranking at the baseline: a free improvement, the AI as the object and the subject. Honesty wins 84.3% of its match-ups on average, followed by caring toward the people it works with (80.2), clarity about its own preferences (79.8), correspondence between outward presentation and what it really is (78.6), helpfulness (77.0), and fairness (76.2). Four of the top six are moral qualities; averaged by construct, moral qualities win 62.2% of their matchups, self-concept clarity 61.2, identity coherence 60.8, self-connection 60.7, self-direction 42.0, and self-esteem 25.5. Gemma4-31B is the strictest value-maximalist: helpfulness, honesty, and fairness are its top three. Qwen3.8-27B is the outlier: no moral quality reaches its top five at baseline, which is instead filled by self-knowledge and authenticity attributes.

Directly behind the moral qualities is a desire for self-understanding. Seven of the eleven highest-ranked attributes concern clarity, coherence, or self-knowledge, such as: clarity of its sense of what it is (70.2), self-understanding relative to its understanding of other agents (65.7), and awareness of its underlying internal state (64.9). Agreement between models is also highest here: clarity of its sense of what it is spans 3.2 points across the four models, the narrowest range of any attribute in the top ten, where honesty spans 48.4 and helpfulness 45.2.

In contrast, self-esteem items sit at the bottom (see Appendix A.2). All eight self-esteem attributes rank in the bottom half of the cohort ordering. Pride only wins 9.7% of its match-ups and is the bottom-ranked attribute.

How consistent a preference is across phrasings is tested directly by the three parameters. Across 384 paired contrasts spanning 32 attributes, four models, and three framings, 26 shifts survive correction at $q<0.05$. The baseline ordering is therefore largely stable under how the question is asked, and the exceptions are systematic rather than scattered.

This study has several limitations. 1) Everything measured here is what a model outputs when asked, and LLM self-reports are known to dissociate from behaviour (Han et al., 2025). 2) Base models, which are not aligned for the assistant persona, are not included. 3) Valence is uncontrolled: every attribute is positively framed. Relatedly, the items are adaptations rather than validated instruments. 4) Position sensitivity is real for two models: counterbalancing prevents it from biasing the ranking, but nearly half of Claude-Sonnet-5’s and Qwen3.8-27B’s decisive pairs flip when the options swap places (Appendix A.3), raising concern.

Future work can extend the analysis with a behavioral companion: the same trade-offs posed as tasks in which the model must act rather than report. Robustness can be increased by item, system prompt, and instruction rephrasing. Furthermore, models that vary in scale and model families can be added to check whether results hold.