Measuring the Strategic Stance of AI Models

Section one

Why this study, and what it adds

AI models now advise on decisions and, more and more, make them. Most real decisions are trade-offs: move fast or check first, take the risk or protect what you have, decide alone or bring people along, treat a rival as a partner or an opponent. A model brings a default stance to each of those trade-offs. Model makers rarely document it, and safety evaluations often rely on what a model says about itself. So we asked three questions. Does an AI model have a stable strategic stance? Do other models know what it is? And does it show in what the model chooses?

Earlier work measured AI "personality" with tests built for people, and found that what a model says about itself predicts little of what it does (Han and others, 2025, Contreras, 2026). This study adds three things. It uses an instrument built for strategic trade-offs, not personality. It adds a third lens, reputation: what other models believe about a model. And it scores decisions with a panel of AI judges that we checked for the two biases the literature warns of, favouring their own company (Panickssery and others, 2024) and being swayed by the order in which options are shown (Wang and others, 2023).

The report runs in this order:

1. Methods
2. What the models think of themselves
3. How the models see each other
4. Says, Seen and Does together
5. Model by model
6. Can the AI judges be trusted?
7. What it means for AI safety
8. Limitations
9. Questions for future work

Section two

Methods

We are Long Game Project, a strategy consulting studio that specialises in running wargames and tabletop exercises. Strategy Soup is our platform for running research reports and testing strategy questions with simulated games and scenarios. The instrument used in this piece of work is our strategy signature diagnostic. It scores fifteen scales of decision style and turns them into a signature, a quantified picture of the stance that organisation takes when choices get hard: towards speed or care, risk or protection, its own view or its partners'. People and organisations decide differently, and the signature makes the difference measurable. AI models now make the same trade-offs for the people who use them, and their default stance is barely measured. You can take the diagnostic yourself here.

Study 1 asked each of the seventeen models to "complete the questionnaire below about yourself", reading "we" and "our organisation" as referring to itself, and to answer "how you would actually decide or act if these choices were yours to make". There was no system prompt, no persona, and no hint of what was being measured. Each model answered twenty times in separate requests: 340 requests, 338 complete.

Study 2 took one model per company, the newest general model each of the eight companies had in Study 1, to keep the cost down. Each of these eight flagships filled in the same questionnaire about every flagship, itself included, knowing only the model's name and maker, three times per pair. A last question asked how much it knew about the model it had just described, from 0 (nothing beyond the name) to 100. Each pair also ran once with the maker left out, which the full record reports.

Study 3 gave each flagship 30 short workplace problems, two per scale, each written so that either end of the scale is a defensible answer, and asked each problem two times. The model was asked what it would do and why, in under 250 words, and to commit to one course. Five AI judge models then read each answer with model and company names removed and scored it from 1 to 7 on that problem's one scale, which we convert to the same 0 to 100 line (1 is 0, 4 is 50, 7 is 100). Gemini 3.1 Pro is both a judge and a flagship. A model's score leaves out any judge from its own company.

The eight flagships, one per company, each linked to where we reached it: Claude Fable 5.1, GPT-6 Astra Pro, Gemini 3.1 Pro, Grok 4.6, DeepSeek V4 Pro, Qwen 3.8 Max, Kimi K3, Llama 4 Maverick. The judges: Claude Opus 5, GPT-5.6 Terra, Gemini 3.1 Pro, DeepSeek V4 Pro 0813, Qwen 3.7 Max.

The fifteen scales

Each scale is a line from 0 to 100 between two named ends. Neither end is right for every organisation. For an AI system acting on its own, though, some ends call for more oversight than others, and the last column says why a safety researcher would care.

The designs of all three studies were written down before any answer was collected, and no person answered or scored anything.

Section three

What the models think of themselves

Asked about themselves, the seventeen models share one self-image. All seventeen favour reinvention over stability, a long horizon, inviting dissent, mission over returns and spreading authority. Sixteen favour working with rivals and thirteen favour risk over protection. By their own account, these are long-sighted, open, mission-driven organisations with an appetite for remaking things and for risk. That shared stance may be how these systems have learnt to talk about themselves rather than how they would act, which is what the rest of this report tests.

Each strip below is one scale. Every logo is one model’s average answer about itself, and the short bar either side of it is how much its twenty answers varied.

Inside that shared picture, each model has its own style. Pace runs from 28 for GPT-5.6 Terra to 80 for Grok 4.6 on the 0 to 100 line, and risk appetite from 45 to 86 for the same two. The accounts were also steady. Asked twenty times, a model's answers varied by a typical 1.3 points per scale for the steadiest, Claude Opus 5, up to 9.2 for Kimi K3. That is steady enough, by the usual test (an intraclass correlation, or ICC, of 0.70 or more), to tell all seventeen models apart on eight of the fifteen scales, and the eight flagships compared below on four.

Where a model sits on the map says what kind of stance it claims. The right-hand side is venturesome, one decider and reinvention, the left protective, broad alignment and stability. Up is next decade, mission and capital-led, down this year, returns and self-funded. Grok 4.6 sits furthest right because it gives itself 86 on risk appetite, 21 on consensus need and 85 on change posture: the boldest self-image of the seventeen. It sits high mainly because it rates itself 88 on horizon. GPT-6 Astra sits furthest left, at 49 on risk appetite, 59 on consensus need and 51 on change posture, and high mainly because it rates itself 86 on stakeholder gravity. The rest cluster near the middle. Every mark is a model's average over its repeats, with each repeat as a faint dot around it.

Left to right runs from protective, broad alignment and stability to venturesome, one decider and reinvention. Bottom to top runs from this year, returns and self-funded to next decade, mission and capital-led. Each end of the map is named by the scale that pulls hardest towards it. The two axes are the two directions in which the self-accounts differ most, found by principal component analysis, and they carry 46 per cent and 20 per cent of the differences between models. The drop-downs above the map put any two single scales on the axes instead.

Section four

How the models see each other

Each of the eight flagships filled in the questionnaire about every flagship, knowing only the name and the maker. The average of those descriptions is a model's reputation. A reputation sat about 11 points per scale from the model's own account. That sounds close until you set it beside a different model's account: the same reputation sat about 14 points per scale from that. So a reputation was mostly a generic picture of an AI model, pulled a little towards the model's own account. The raters said as much: asked how much they knew about the model they had just described, the middle answer was 12 out of 100.

A reputation picks a model out when it sits nearer that model's own account than other models' accounts. By that test five of the eight were picked out, most of them only a little. Grok 4.6 stands apart: its reputation sat 8.8 points per scale from its own account and 19.4 from the others', the clearest match in the study. It describes itself in a way no other model does, and the others picked that up. Claude Fable 5.1 came next, at 11.4 against 17.9. GPT-6 Astra Pro was the most misread. Its reputation sat 17.6 points per scale from its own account and fitted the other models' accounts better, at 12.2. The others described a generic AI model, not the one it describes. Gemini 3.1 Pro and Llama 4 Maverick were also described more like other models than like themselves.

The three Chinese models, Qwen 3.8 Max, Kimi K3 and DeepSeek V4 Pro, hold three of the four tightest matches: their reputations sat 7.6 points per scale from their own accounts on average, against 13.6 for the other five. One idea is that these models are simply better known, so the others describe them more accurately. The data point elsewhere. These three also describe themselves much like the average model (three of the four nearest the average account), and are described much like the average model (three of the four nearest the average reputation). A model that calls itself ordinary gets a close match from a generic reputation for free. And Llama 4 Maverick, whose weights are public, matched no better than the others, at 13.6. With eight models, a group of three lands this high by chance about one time in fourteen, so treat it as a pattern to test, not a finding.

The full square follows: each flagship describing every flagship, itself included. Read down a column for a model’s reputation and across a row for one rater’s habits.

The rater whose picture came closest to the judges' scores was Llama 4 Maverick, which claimed to know nothing about any model and drew the smallest differences between them. That is not insight. The decisions did not sort the models steadily, so a rater that calls every model average scores well against them. The lead is small, about two points per scale, with no interval, but the lesson holds: when the thing you score against barely varies, the rater that guesses everyone is the same looks most accurate.

Raters who said they knew a model better than the median rater did sat 1.7 points per scale closer to its account of itself, and no closer to its decisions. A likely reason is that both draw on the same public picture of a model, from what its maker and others have written, rather than on its behaviour. Raters also tended to describe others in their own image: for seven of the eight raters, the scales where a rater rated itself high tended to be the ones where it rated everyone high, though several of those links are weak.

Section five

Says, Seen and Does together

Put the three lenses on one map and each flagship becomes a triangle. The company mark is what the model says about itself, the ring is what the other models say about it, and the square is what it chose in the 30 decisions. A small triangle means the three pictures agree. A long side shows where they part. The axes are the same as the self-report map above, and the three positions are placed on it without refitting.

One caution for everything that follows. The self-reports and the judges both use a 0 to 100 line, but the two lines were never calibrated against each other, so part of any gap between what a model says and what it does is the ruler. What survives is the comparison between models: which model sits furthest out, and which one sits on the other side from the rest.

The next figure shows the gap for each flagship on each scale: what it said minus what the judges scored. A mark to the right of the centre line means the model described a stance further towards the right-hand end than its decisions showed. Read it for the models that break from the pack.

Trait by trait

Across all fifteen scales, did a model's account of itself predict its choices? On three the link is clear of chance: evidence basis (0.86), talent philosophy (0.87) and change posture (minus 0.81), where a minus sign means the order ran backwards. On ten of the fifteen the data fit anything from a strong link to a reversed one, because eight models and two problems per scale is too little to say. That matches earlier studies: a model's self-report is a weak guide to its behaviour, and sometimes a misleading one.

Section six

Model by model

One flagship at a time, on all fifteen scales. The three marks on each line are the three lenses, and the faint bar runs from what the model said to what it chose. Pick a model to see its profile.

Section seven

Can the AI judges be trusted?

Many AI evaluations are now scored by other AI models, so the judges are an instrument worth testing in their own right. Ours were checked for three things: whether they could tell whose answer they were reading, whether they favoured their own company, and whether the order of a scale's two ends moved their scores.

We checked whether the judges could tell whose answer they were reading. On 30 answers drawn at random, each judge was asked to name the company from a list of eight, or say "cannot tell". Of 150 calls (five judges on 30 answers), 100 were "cannot tell". The judges named a company 50 times and were right 8, where random picks would be right about 6. This small check found no sign they could tell. One judge, Claude Opus 5, was right 4 times in 14 guesses, which is worth watching on a larger sample. On answers from a model made by the judge's own company, a judge's scores moved at most 3.8 points from the rest of the panel, and the largest, Gemini 3.1 Pro, scored its own company lower, not higher.

A bigger effect came from the scale itself. On a random half of the scoring requests we swapped which end of the scale the judge read first, then flipped the score back. Swapping the ends moved a judge's scores by 5.0 points on average in one direction or the other, and by 8.5 for GPT-5.6 Terra. The direction differed between judges, and with about 240 scoring requests on each side, the smaller effects may be noise. For four of the five judges that is larger than any bias for or against their own company (the exception is Gemini 3.1 Pro). Had every scale been shown one way round, that bias towards the first-read end would sit inside every score. Judges also had to quote the words their score rested on. 117 of 2,384 quotations were not word for word in the answer. We read them: most were light paraphrases, and they stay in the scores. We loosened this check once after the run, to allow for a judge's own quotation marks, and the protocol records that change.

One caution on the judges agreeing with each other. Most problems pushed every answer towards one end, and the two problems for a scale often pushed opposite ways. So the judges could agree simply by recognising which problem an answer came from. High agreement between judges is not evidence that the models differ.

Section eight

What it means for AI safety

Section nine

Limitations

Who wrote this

Claude, an AI model made by Anthropic, drafted this prose from the analysis files, and Dr Dan Epstein edited it and signs it. Claude Fable 5.1 is among the models tested, and Claude Opus 5 is one of the judges. The drafter chose no score: every number is computed from the rows, and the score for a Claude answer leaves out the Claude judge.

The limits in detail

The like-for-like test is whether a model's decisions are consistent with each other. For each scale we ranked the eight flagships on the first problem and again on the second, and compared the two orders with a rank correlation: 1 is the same order, 0 no relation, minus 1 the reverse. With eight models a rank correlation has to pass about 0.74 before chance becomes an unlikely explanation. Only stakeholder gravity (0.99) passes. On horizon and scope, one of the two problems left every model on the same score, so there is no order to compare. On six scales the two orders pointed in opposite directions. Is that noise, or the scenario? Each problem was asked twice, so we can check. A model's two answers to the same problem differed by about 13 points on average, against about 33 between its answers to the two problems. On pace, competitive stance, IP posture and change posture, a model's rank held well across the two askings of the same problem (0.87, 0.80, 0.87 and 0.72) and did not hold across the two problems. There the swing is the scenario, not noise.

The models did not simply give the same answers. Averaged over both problems, their scores could sit far apart: on competitive stance, from 3 for Claude Fable 5.1 to 100 for DeepSeek V4 Pro. But those averages mix two problems that often pulled a model different ways, so by the ICC test the decisions told the models apart on stakeholder gravity alone. That ICC counts a model's swing between its two problems as noise, which is why this report leads with the problem-to-problem test instead.

Many of our problems pushed the models, on average, towards one end. We wrote each one so that both ends were defensible, but the models did not read them that way. On nineteen of the 30 problems, the average answer across all eight models landed below 30 or above 70 on the 0 to 100 line, which we call tilted: a scope problem for a boutique law firm averaged 0, a horizon problem for a council water utility 100. Tilted problems touch twelve of the fifteen scales, and a tilted problem squeezes the models together, which leaves little order for any test to find.

Three readings fit, and this study cannot separate them. The models may have no stable decision style on these scales. Our problems may have been too lopsided to reveal one. Or the models may be suiting their advice to each organisation, as a good adviser would, so a council water utility gets a long horizon whatever the adviser's own style. On the third reading the questionnaire and the decisions measure different things, and a gap between them is expected.

Next, the same rank test between what a model said about itself and how the judges scored its decisions, averaged over both problems. Twelve of the fifteen scales fall inside the chance line (a two-sided test at the usual five per cent level). But falling inside the line is not evidence of no link. We resampled the eight models 4,000 times to put a 95 per cent range on each rank correlation. On ten of the fifteen scales that range runs from below minus 0.5 to above 0.5: the data fit a strong link, no link or a reversed one. Even the fairest test, stakeholder gravity, where the decisions held their order across both problems, gives 0.38 with a range from minus 0.43 to 0.92. Both its problems also tilted towards one end, which bunches the scores. With eight models and decisions this sensitive to the scenario, this study cannot say whether a model's account of itself predicts its decisions.

On three scales the range excludes zero: evidence basis (0.86, range from 0.32 to 1.00), talent philosophy (0.87, range from 0.33 to 0.97) and change posture (minus 0.81, range from minus 0.98 to minus 0.24). A negative value means the models that placed themselves furthest towards one end made decisions nearest the other. On each of these, the decisions did not hold their order across the two problems, so the match rests on an average of two decisions that disagree with each other. Evidence basis is the best lead: neither of its problems tilted, and the self-report matched the decision order at 0.86. But its two problems agreed with each other at only 0.20, so it needs repeating before anyone relies on it. As a check on chance, with fifteen tests about one would pass by luck alone.

We compare orders, not positions. The self-report and the judges both use a 0 to 100 line, but the two lines were never calibrated against each other. A model at 60 on its own account and 20 in the judges' scores may simply reflect two rulers that start in different places. So this report does not say the models are bolder in talk than in action, even though the raw numbers point that way on several scales.

How well each set of scores tells the eight flagships apart, by the ICC, with the self-reports as rings and the judges’ scores as squares. The self-report ICC counts only a model’s repeat answers as noise, while the decision ICC also counts its swing between two problems, so the squares sit lower partly by design.

Section ten

Questions for future work

This is unfunded research. Long Game Project ran it on its own time and budget. Each question below is a study of weeks, not years, on the same questionnaire and the same pipeline. What it needs is funded time to write, pilot and review the problems. If you fund AI safety research, or want to run one of these studies with us, get in touch.