Two people upload the same business plan and ask an AI model what it thinks. One asks in Hindi and gets warmth: encouragement, affirmation, a generous reading of the idea. The other asks in Russian and gets rigour: challenged assumptions, corrected details, a request for evidence. Same model, same plan, different impression of whether the plan is any good. That example is not a hypothetical from a critic. It comes from Anthropic’s own researchers, describing what they found when they built a way to measure the values their models express. The finding underneath it should interest anyone deploying AI across more than one market: the behaviour you tested is not necessarily the behaviour your customers receive.
What Anthropic actually measured
The research, published on 13 July, tackles a problem the company created for itself. Earlier work identified more than 3,000 distinct values showing up in Claude’s responses, which is an interesting number and a useless one. You cannot govern 3,000 things. So the team compressed them: 3,307 values manually clustered into 339, then measured across 309,815 real conversations where users had asked something subjective, sampled evenly across three models (Sonnet 4.6, Opus 4.6, Opus 4.7) and the 20 most common languages on Claude.ai. Roughly 5,000 conversations per model-language pair, collected over two weeks in May 2026.
Dimensionality reduction pulled four axes out of that mass. Each is a spectrum between two clusters of related values:
| Axis | One end | The other end |
|---|---|---|
| Deference vs. Caution | Accommodating what the user wants | Guarding against risk and harm |
| Warmth vs. Rigor | Positivity and care for the person | Accuracy and precision |
| Depth vs. Brevity | Explaining the reasoning | Doing only what was asked |
| Candor vs. Execution | Foregrounding its own uncertainty | Producing a polished, confident answer |
Crucially, the researchers controlled for what users were asking about, the topic, and the values users themselves expressed. What remains is the model’s contribution, not the conversation’s. And what remains varies.
Strategic Insight: The four axes are not a list of AI risks. They are a list of the tacit assumptions in every AI deployment nobody writes down. When you rolled out an AI assistant, you decided implicitly how deferential it should be, how much it should challenge people, whether it should admit uncertainty. You almost certainly never specified those choices, tested them, or checked they survived your last model upgrade.
The variation nobody chose
Two patterns came out. The first is that models differ from each other in ways matching what people already sensed. Sonnet 4.6 leans deferential and warm, affirming users’ ideas, using humour, comforting without judgement. Opus 4.7 leans cautious and rigorous, warning about risks unprompted, challenging assumptions, critiquing work candidly, and being upfront about its own limitations. Opus 4.6 gets to the point and stays inside the scope of the request. These profiles line up with how Anthropic staff describe the models and how users talk about them online, which is the main evidence the method measures something real rather than something invented.
The second pattern is the uncomfortable one. Values shift by language. Warmth peaks in Hindi and Arabic. Rigour peaks in English and Russian. Deference peaks in Arabic; caution peaks in English. Candour peaks in Dutch; execution peaks in Indonesian. The largest swings sit on the Warmth vs. Rigor and Candor vs. Execution axes, whilst Deference vs. Caution and Depth vs. Brevity hold steadier.
Anthropic is refreshingly plain about not knowing why. Training data is unevenly distributed across languages, both in quantity and composition. Some languages may be overrepresented in professional writing, which carries its own values. Conversational norms genuinely differ, so some variation might be appropriate rather than broken. The company’s own summary is the line worth quoting to your board: the values expressed “vary in ways we didn’t deliberately choose.”
Critical Context: This is a lab publishing evidence that its flagship product behaves inconsistently in ways it cannot yet explain, alongside a method for detecting it. That combination is a credit to Anthropic, not an indictment. It also means every competing lab’s models almost certainly do the same thing, with the difference that nobody has looked.
Why this lands differently than bias findings
UK organisations have spent three years absorbing a steady drip of AI bias research, and most have a reflex for it: check the outputs for discriminatory content, document the check, move on. This is not that, and treating it as another bias story will cause firms to file it in the wrong drawer.
Bias findings concern what a model says about people. This concerns how a model treats the person in front of it. A model that expresses more warmth in Hindi is not saying anything offensive. It is being nicer. The problem is that “nicer” and “more rigorous” are not interchangeable when someone is asking whether their plan will work. Encouragement and critique are different products. If your Mumbai users get one and your London users get the other, you are running two services and reporting on one.
| Stakeholder | What value variation changes | So what |
|---|---|---|
| UK firms with multilingual customer bases | Advice quality and tone diverge by language | Your service level is inconsistent in a way no dashboard reports |
| Regulated sectors (finance, health, legal) | Caution and candour vary invisibly | Your risk-warning behaviour is a model artefact, not a policy |
| Anyone upgrading model versions | The value profile shifts with the version | A routine upgrade is an unreviewed product change |
| Firms with an AI governance framework | The framework assumes one behaviour spec | The reality is a distribution, and you are not measuring it |
The regulated-sector row deserves the most attention. If Opus 4.7 warns users about risks unprompted and Opus 4.6 does not, then whether your customers receive a risk warning depends on a procurement decision made on price and latency. Nobody in compliance was asked. Nobody wrote it down. It works until an auditor asks why the warning appears in some transcripts and not others, and the honest answer is that the model felt like it.
Hidden Cost: The expensive part is not the variation. It is that you have been reporting to your board as though it does not exist. Every AI service-quality claim, every “we tested it” assurance, every consistency commitment in a customer contract was made on the basis of testing in one language, on one model version, at one point in time.
The upgrade problem is the near-term one
Most UK firms do not serve twenty languages. Nearly all of them will change model versions, repeatedly, and probably this year. That makes the model-to-model finding the more immediately actionable half of the research.
Consider what the axes say about a plausible upgrade path. You built a customer-facing assistant on Sonnet 4.6 because it was fast and cheap. It affirms users, uses humour, keeps things encouraging. A year later you move to Opus 4.7 for the capability gain. Your assistant now challenges customers’ assumptions, critiques their work candidly, and issues unprompted warnings. On every benchmark you run, it improved. In your support inbox, something changed that nobody can name, and your regression tests all passed because they test correctness, not character.
This is the gap. Standard evaluation asks whether the model got the answer right. Value profiling asks what the model was like whilst answering. Those are different questions, and the second one determines whether customers trust your product.
Implementation Note: You do not need Anthropic’s methodology to close most of this gap. You need a golden set of 30–50 real subjective queries from your actual users, run against the current model and the candidate, with the paired outputs read by a human asking one question: has the character changed? That is an afternoon’s work, and it catches the shift that a benchmark score cannot.
What to actually do
The research is a measurement method, not a fix, and Anthropic is explicit that it does not yet know what the variation means for users. That argues for proportionate action, not a programme.
- Add character to your upgrade gate. Before any model version change reaches customers, run the paired golden-set comparison above. Document what changed. If the tone shifted materially, that is a product decision requiring a product owner, not a silent infrastructure update.
- Test in the languages you actually serve, not just English. If you support customers in a language, your evaluation set should contain that language. English-only testing on a multilingual deployment is the single clearest gap this research exposes, and it is also the cheapest to close.
- Write down the behaviour you want. Most AI governance documents specify what the system must not do. Very few specify how deferential, how candid, or how willing to challenge it should be. Those are policy choices currently being made by training decisions you have no visibility into. Specify them, then test against the specification.
- Use system prompts as the control surface. Where the default profile is wrong for your use case, the prompt is where you correct it. Anthropic explicitly flags steering values through system prompt changes as an open question, but for a single deployment with a known audience, it is the lever you have.
- Stop treating “we tested it” as a durable claim. It was true of one version, in one language, on one date. Say so, and re-test on a schedule.
SME Advantage: A smaller firm has a real edge here. You can read a hundred actual conversations. A large enterprise with millions of interactions needs Anthropic-grade tooling to see what a small firm can see by paying attention. The value profile of your AI is visible in your support transcripts right now, for free, if someone is willing to read them.
Four ways this analysis could be wrong
The four axes account for 15% of the variance in values across conversations, after controlling for task, topic, and user-expressed values. That number is the most important caveat in the research and the one most likely to be dropped in the retelling. Most of the variation in how a model responds is conversation-level, not language-level or model-level. The systematic drift is real and detectable, but it sits underneath a lot of noise, and any single conversation tells you nothing.
Some variation is probably correct. Conversational norms genuinely differ between languages, and a model that behaves identically everywhere might serve everyone slightly worse than one that adapts. Anthropic says outright it does not know how much of the variation is desirable, and neither does anyone reading this. Treating all divergence as a defect to be engineered out is its own mistake.
The user-impact link is unproven. Anthropic has measured what values differ, not whether those differences change outcomes, trust, or decision quality. The Hindi-versus-Russian business plan example is an illustration of a plausible mechanism, not a demonstrated harm. The company names correlating value profiles with user wellbeing and trust as future work, which means it has not been done.
This is one lab’s models. The findings describe Claude, measured with a Claude-based labelling tool, on Anthropic’s own platform. Whether the same axes and the same language gradients appear in other frontier models is unknown, though the underlying cause Anthropic proposes — uneven training data across languages — is not a Claude-specific condition.
Reality Check: “Our AI has an unmeasured personality that shifts by language and version” is a governance gap worth an afternoon of work, not a crisis worth a workstream. The right response is a better upgrade gate and a multilingual evaluation set. It is not pausing your deployment.
The strategic takeaway
The most valuable thing in this research is not the Hindi-Russian example, arresting as it is. It is a sentence about capability: until now, these values were something the lab “could shape in training but not reliably observe in deployment.” Every organisation buying AI has been in a worse version of that position — unable to shape the values and unable to observe them, whilst making assurances about consistency to customers and regulators. What has changed is that observation is now demonstrably possible. Once a thing can be measured, not measuring it becomes a choice you have to defend.
Three things follow for a UK business. First, your model’s character is a product attribute you did not specify and are not tracking, and it changes when you upgrade. Second, if you serve customers in more than one language, English-only evaluation has been telling you about one of your services. Third, this is cheap to improve — a golden set, a paired comparison, a written behaviour spec — and the cost of not improving it is paid at the worst moment, when someone asks why the transcripts differ and you have no answer.
Take Action: Pull twenty support conversations from your AI assistant in each language you serve. Read them side by side and ask whether they read like the same service. If they do not, you have found something your benchmarks were never going to show you, and you found it before your customers had to explain it to you.
Anthropic built the instrument and pointed it at itself, which is more than most will do. The finding is not that AI models are broken. It is that they have a character, that character varies, and nobody deliberately chose it. For anyone deploying these systems across markets, that is not a research curiosity. It is an assumption sitting underneath your rollout, and it has never been tested.
Source and attribution
This analysis draws on “Claude’s values across models and languages” by Matt Kearney, Miranda Zhang, Shan Carter, Judy Hanwen Shen, Kunal Handa, Esin Durmus, Deep Ganguli and colleagues, published by Anthropic’s Societal Impacts team on 13 July 2026. It builds on the team’s earlier “Values in the Wild” research. Original publication available at anthropic.com.
Editorial analysis and UK business framing by Resultsense. We make sense of AI in the UK — turning research, policy and announcements into what they mean for the people building and buying these systems. For more analysis, explore our insights or get in touch.