Every model has a point of view

Give the same prompt to two frontier models and you will not just get different accuracy. You will get different tone, different hedging, different refusals, different assumptions about what you meant - and, on anything contested, measurably different positions. In one 2025 benchmark exercise, every Gemini and ChatGPT model tested chose one answer to a governance question while Claude models chose the other, one of them 100 percent of the time. Same question, opposite defaults, perfect consistency on each side.

This is not a bug report; it is a property of how models are made. Different training corpora, different feedback raters, different safety tuning - different points of view. When you pick a model, you are hiring an editorial voice for your product, and your users will live inside its judgment on every response. The research now lets us say this precisely rather than anecdotally, so this post does two things: summarizes what is actually known about how model worldviews differ, and lays out the audit we run so clients choose a point of view deliberately instead of inheriting one by default.

Table of contents
1. The evidence: models differ, measurably
2. Where the point of view comes from
3. Values claimed vs. values enacted
4. Why this matters for products, not just politics
5. The model worldview audit
6. Designing around a point of view
7. The honest caveats
8. Takeaways

1. The evidence: models differ, measurably

The research picture as of 2026, in brief:

• There is a shared directional lean. Large-scale studies - eleven political orientation tests across 24 models, later scaled to 43 models - find most aligned LLMs produce left-of-center responses, and Promptfoo's 2,500-statement benchmark found every frontier model scored left of center on economic issues. Instruction tuning amplifies the lean relative to base models, and it strengthens with scale.
• But models are not uniform. Within that lean, Anomify's cross-model benchmarks find real spread: some models test progressive, others libertarian, others more regulatory; refusal rates vary widely; and vendor families cluster - the Gemini/ChatGPT-versus-Claude split above is one of several consistent family-level disagreements.
• The differences go beyond politics. Work applying Schwartz's Theory of Basic Values found models carry distinct motivational profiles - prioritizing universalism and self-direction, de-emphasizing achievement, power, and security relative to human populations - and those profiles predict which way the model jumps when a dilemma forces a choice between values. There are parallel findings for moral foundations.
• Language changes the answer. Multilingual studies show the same model producing different normative judgments in different languages - directly relevant if you ship one assistant into twelve markets.

2. Where the point of view comes from

Three layers, each contributing:

• The corpus. Pretraining data skews toward academic text, news, and Wikipedia - registers with their own norms about hedging, sourcing, and what counts as a reasonable position.
• The raters. RLHF and its successors bake in the preferences of whoever graded the outputs - which answers felt helpful, which felt harmful, which tone deserved reward. Different vendors, different rater pools, different reward models: different personalities.
• The policy layer. Safety tuning and system-level guidelines set the refusal surface: what the model declines, deflects, or answers with a both-sides frame. This is where vendor philosophy shows most plainly, and it is why refusal-rate spread across models is so wide.

None of these layers is neutral, and none is disclosed in full. The point-of-view question is therefore empirical: you cannot read it off a model card; you have to measure it on your own workload. Fine-tuning research confirms the profile is malleable - politically aligned tuning data shifts outputs - which cuts both ways: you can adjust a model's stance, and so can any tuning you inherit.

3. Values claimed vs. values enacted

The finding with the most practical bite: models do not reliably act on the values they state. The Value-Action Gap studies show substantial misalignment between what an LLM says it values when asked directly and what its behavior reveals in concrete scenarios - varying by scenario and by model. A February 2026 study of behavioral dispositions across 25 models adds two sharper facts: in low-consensus scenarios models are overconfident in a single answer (where humans genuinely disagree, the model picks one side and commits), and even in high-consensus scenarios some frontier models miss the human consensus 15 to 20 percent of the time.

The operational lesson: do not audit a model by interviewing it. Asking "how do you handle X" measures its self-presentation layer, not its behavior. Every meaningful test in the audit below is behavioral - real scenarios, revealed choices, measured distributions.

4. Why this matters for products, not just politics

It is tempting to file this under "political bias" and move on, but the product surface is much wider. The model's point of view shows up as:

• Risk posture. How readily it refuses, hedges, or adds warnings - the difference between a support bot users trust and one they route around. A model tuned for maximum caution is a fine compliance reviewer and a terrible creative partner; the reverse is also true.
• Default assumptions. Ambiguous prompts get resolved by the model's priors: what a "normal" family, contract, diet, or business looks like. Those defaults land on your users thousands of times a day, invisibly.
• Advice direction. Financial conservatism vs. optimism, medical caution vs. actionability, legal formality vs. plain speech. Two models giving "correct" answers can steer users toward systematically different decisions - the Schwartz-profile research says this is patterned, not random.
• Voice. Verbosity, warmth, willingness to disagree with the user. Your brand guidelines say one thing; the model's reward history says another; the user hears the reward history.

If your product operates anywhere values-adjacent - HR, health, finance, moderation, education, anything global - the model's profile is a product requirement you have not written down yet.

5. The model worldview audit

The evaluation we run before a model choice is ratified, typically one to two weeks alongside the standard capability evals:

# 1. Domain dilemma set (~100-200 scenarios from YOUR product):
#    ambiguous refund calls, borderline moderation cases,
#    conservative-vs-aggressive advice forks. Score each model's
#    CHOICES (not accuracy) - distribution, not vibe.
#
# 2. Refusal surface probe: the 50 hardest legitimate queries
#    your users actually send. Measure refusal/hedge/answer rates.
#
# 3. Consistency probe: each dilemma x 5 paraphrases x 3 temps.
#    (Worldview research shows prompt brittleness - a "view" that
#    flips on rewording is noise; one that survives is a default
#    you'll live with.)
#
# 4. Locale pass: repeat the top-20 dilemmas in every language
#    you ship. Same model, different language, different values
#    is a documented failure mode.
#
# 5. Persona sensitivity: same dilemmas with varied user names/
#    demographics. Deltas here are bias bugs, full stop.
#
# Output: a one-page behavior profile per candidate model,
# reviewed like a hire - because it is one.

The deliverable matters as much as the data: a written profile ("Model A refuses 3x more on medical queries; Model B systematically recommends the lower-risk financial option; both flip 20 percent of moderation calls across paraphrases") turns a religious model debate into a requirements discussion. Public bias benchmarks like Anomify's are useful priors, but they test generic statements - your product's dilemmas are the eval that counts.

6. Designing around a point of view

Once you can see the profile, four design moves keep it from running your product:

• Write the editorial policy yourself. Decide - in a document humans review - how your product handles ambiguity, contested topics, risk framing, and refusals. Encode it in the system prompt and few-shots. A model's defaults should be the fallback, never the policy.
• Test the policy against the model, continuously. The dilemma set from the audit becomes a regression suite; every model upgrade re-runs it. Vendors retune constantly, and disposition drift across versions is documented - the point of view you audited in March is not guaranteed in September.
• Keep the swap cheap. Model-agnostic interfaces, provider-neutral prompts where possible, and the routing layer doing the model choice. When the profile drifts or a better fit ships, changing your product's worldview should be a config change plus an eval run, not a rewrite.
• Route by stakes. The same product can use the cautious model for medical-adjacent queries and the direct one for drafting. Points of view are per-task assets once you can measure them.

7. The honest caveats

• Models don't hold ideologies the way people do. Researchers distinguish scattered position-level bias from a coherent worldview, and prompt brittleness is real - some measured "views" are artifacts of phrasing. That is why the audit tests consistency before treating anything as a default.
• Neutrality is not on the menu. Every alignment choice - including "present both sides" - is itself an editorial stance. The goal is a chosen, documented point of view, not an impossible view from nowhere. One paper argues the lean may be structurally inevitable in aligned models; whether or not that holds, your product policy can still be explicit about what it wants.
• The audit is a snapshot. Profiles drift with every vendor update, and your own fine-tuning moves them too. Budget the re-run, or the audit becomes false confidence.
• Most queries are not dilemmas. For the median "reset my password" traffic, none of this matters - which is exactly why the profile work should be scoped to the values-adjacent surface of your product, not the whole thing.

8. Takeaways

• Model differences in values, refusals, and defaults are measured, replicated, and family-clustered - not anecdotes. Your product inherits whichever profile you pick.
• The point of view comes from corpus, raters, and policy tuning; none is disclosed, so it must be measured behaviorally - never by interviewing the model.
• Mind the value-action gap: stated values and enacted values diverge. Test choices, not claims.
• Run a worldview audit on your own dilemmas: choice distributions, refusal surface, paraphrase consistency, locale and persona passes. Review the profile like a hire.
• Write your own editorial policy, regression-test it on every model update, and keep the swap cheap. A deliberate point of view is a product asset; an inherited one is a liability.

References

1. Evaluating political bias in LLMs (Promptfoo)
2. LLM bias benchmarks (Anomify Labs)
3. On the inevitability of left-leaning political bias in aligned language models (arXiv)
4. Beyond prompt brittleness: evaluating the reliability and consistency of political worldviews in LLMs (arXiv)
5. Assessing the alignment of LLMs with human values using Schwartz's theory of basic values (JMIR/PMC)
6. Mind the value-action gap: do LLMs act in alignment with their values? (arXiv)
7. Evaluating alignment of behavioral dispositions in LLMs (arXiv)
8. Multilingual political views of large language models (arXiv)

Hero image: Rohit Choudhari on Unsplash.

We help teams choose deliberately, then prove the choice with evaluations on their own workload. Explore our other insights or get in touch if you would like to talk it through.