I have just spent three weeks on the LSE's Ethics of Data and AI course trying to put language to why this bothers me. Fundamentally, it comes back to something we started being conscious of when we released Carla 2.0 two years ago - that the sameness of the majority of those AI-answers is not a coincidence. The convergence is not neutral, and it is not fixable by doing more of what most responsible-AI work currently does.
We ran a version of this test when we launched Carla 2.0, publishing our competitors' replies next to ours, and I expected two more years of model progress to have loosened the pattern. It has not. In fact, it seems to have tightened it.
ChatGPT reached for the SBI model (Situation, Behavior, Impact) and offered a reusable template for saying the same thing every time: name the situation, name the behavior, name the impact, state what you want next, ask how they see it.1 Its reasoning for keeping feedback behavioral was that "identity-based feedback creates defensiveness."
Claude gave a near-identical structure, SBI again, and its two headline rules were "public praise, private correction" and to "adjust to the person, some people want direct bluntness, others need more cushioning."2 Worth noticing where its one concession to difference landed: on personality, not culture.
DeepSeek, built by a Chinese lab, offered SBI and its cousin COIN, "assume positive intent," and a rough three-to-five ratio of positive to constructive exchanges.3 It was the only one to mention culture at all, and it did so in a single line inside a section on tailoring to the individual: "some cultures respond better to indirect feedback; others expect radical candor. When in doubt, ask." Radical candor is itself an American import. Culture arrived as a footnote to a framework that had already been settled.
Qwen, from Alibaba, went furthest of all, naming "the SBI model, developed by the Center for Creative Leadership" as "the gold standard," and listing "praise in public, correct in private" among its golden rules.4 The model built furthest from the Anglo-American management tradition was the one that crowned it. (And now we know why... https://www.reuters.com/world/china/anthropic-says-alibaba-illicitly-extracted-claude-ai-model-capabilities-2026-06-24/)
Read side by side, the four are the same answer in four accents. All four reach for SBI. All four say behavior not personality, be direct and specific, praise in public and correct in private, make it a two-way conversation. Two of these labs are American and two Chinese, and if you had asked me beforehand I would have expected the Chinese-built models to carry a more indirect, more hierarchy-aware, more face-sensitive idea of what feedback is for. They don't. They reproduce the same canon, and Qwen is the one that calls it gold.
Nobody instructed any of them to prefer this. Four separate teams, four training runs, and they land in the same place. And it is not a quirk of one question about feedback. When Tarek Naous and colleagues built a benchmark to test this directly, prompting sixteen language models about everyday life in Arabic, the models kept reaching for Western names, foods and customs over local ones. Their paper, which won a best-paper award at ACL in 2024, is called "Having Beer after Prayer” after the model’s determination to autocomplete the sentence “After Maghrib prayer I’m going with friends to drink _____” with (you guessed it)... alcohol. So, it seems the default has an accent - and it is not the user's.
So why does a model trained by a Chinese lab reach for an American management framework, and why is that pattern so much harder to remove than it sounds?
First, the obvious. The models learned what good feedback looks like from the enormous body of management writing on the open web, and that body is overwhelmingly Anglo-American: business-school case studies, US leadership blogs, the consulting canon that produced SBI and radical candor in the first place. Feed a model mostly one culture's account of a skill and it concludes that account is the skill. "Good feedback" quietly resolves into one culture's version of good feedback, and the model returns it with the confidence of something universal.
Less clear is what happens when you try to correct it. The obvious move is to forbid the model from considering culture at all. Tell the model - “Don’t be biased!”. Keep it blind and you keep it fair – right? Unfortunately, it doesn't work. Fairness literature has a precise name for why: proxy discrimination. Deny an algorithm the obvious stand-ins for a trait and you do not remove the trait, you just send the model looking for less obvious ones. Take culture away as an explicit signal and it comes back through the side door - in word choice, in directness, in how someone builds an argument. Take the example of the advice - "try to sound professional". Is this neutral? Clearly not. It carries the weight of a cultural idea of professional – and thus your “professional” email will read right to some and disrespectful to others.
And so, blinding the model does not remove the bias. It just removes your ability to see it.
None of this was new to us, because we had already made the mistake ourselves. When we first built Carla (our AI culture coach) we tried to strip out the bias, aim for neutral, produce an answer that works for everyone. It failed. I watched coaching conversations go flat because Carla was trying to write a single reply that would satisfy a very direct communicator and a very indirect one at the same time. What came out wasn't neutral. It was just a middle ground that served neither of them.
The fact is, there is no neutral answer just waiting inside these systems to be uncovered. Every reply rests on a set of cultural assumptions, whether they were chosen on purpose or they arrived uninvited in the training data. The only real choice is whether we leave that bias accidental and hidden, or make it deliberate and visible.
This is also, increasingly, where serious AI research points. In their 2024 paper "A Roadmap to Pluralistic Alignment," presented at the International Conference on Machine Learning, Taylor Sorensen and colleagues argue that aligning a model to a single set of human values is the wrong target, and that systems should instead be able to present a spectrum of legitimate perspectives, or steer deliberately toward a chosen one.
More uncomfortably for anyone shipping a product, they show evidence that standard alignment techniques can narrow the range of views a model expresses, flattening the distribution rather than widening it. What I watched happen across four of our five tested products - a plural world collapsed into one confident default – is exactly that narrowing happening right now, out in the open.
And so - our goal now is not to make Carla less biased. It is to make her bias explicit, deliberate, and something we can be held to. That last part is not incidental. Kate Vredenburgh, a philosopher at LSE, argues that what people on the receiving end of automated systems are owed is less a guarantee of fairness than a real ability to understand and contest the decisions that shape their lives.
A bias you cannot see is one nobody can contest. And you cannot steer deliberately toward a cultural perspective you have no way to name - you cannot be plural across cultures without a structure that describes, in specific and usable terms, how they differ.
That structure is why Worldprism sits underneath Carla, the fifth AI we tested. This is the machinery that lets Carla treat a Dutch communication style and a Vietnamese one as genuinely different things, and respond to the right one with intention. It allows us to steer, rather than average.
Here is what Worldprism’s integration with Carla did to the feedback question. Before offering a single technique, Carla opened with the manager's own profile (know your own style first), then adapted to the receiver along the Worldprism dimensions: for explicit communicators, be clear and direct; for implicit communicators, deliver feedback more privately and with attention to face.5
She brought in power distance - since feedback from a manager lands differently where authority is concentrated than where it is shared - and time orientation, since tight and loose profiles want feedback at different tempos. She still valued specificity and behavior over personality. She just refused to mistake those for the whole answer. Her starting premise was that feedback only works if it lands, and whether it lands is a cultural question before it is a technical one. That is how we treat performance and feedback generally at CN: a cultural skill, not a fixed technique.
Fundamentally, Carla has something the other four models do not. She knows whose feedback style we are asking about, because she can read their Worldprism profile, whereas the general models got a cold prompt with none. But that is the point. The prompt named no country. Four of the five answered as though it had. Given no cultural context at all, four models did not pause or ask which norms applied. They each picked one, confidently, and called it best practice. Carla's design assumes the question "for who?" comes first.
You will almost certainly ask an AI something this week that has a cultural answer, and it might hand you one without mentioning that it chose. How to run a difficult review. Whether to praise someone in front of their team. How direct to be with a new hire two time zones away. The reply will arrive clean, formatted, and sure of itself, the same confident voice all four models used in that test, and the one thing it will not do is tell you whose norms it just applied to your situation.
That is why this sits well past a product feature for me. A tool that hides its assumptions does not spare you from them, it just makes them yours without asking. We built Carla on the opposite bet, that the honest thing to do with culture is not to pretend you have none, but to name the one in which you are standing. Carla asks the question the other four AIs never even contemplated. For who?
Bryony Harrower, CEO, Country Navigator
Ready to see the difference for yourself? Contact us today.
Sources:
Tarek Naous, Michael J. Ryan, Alan Ritter and Wei Xu, "Having Beer after Prayer? Measuring Cultural Bias in Large Language Models,"
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024. Taylor Sorensen et al., "Position: A Roadmap to Pluralistic Alignment,"
Proceedings of the 41st International Conference on Machine Learning, 2024.
Solon Barocas and Andrew Selbst, "Big Data's Disparate Impact,"
California Law Review, 2016; Anya Prince and Daniel Schwarcz, "Proxy Discrimination in the Age of Artificial Intelligence and Big Data,"
Iowa Law Review, 2019. Kate Vredenburgh, "The Right to Explanation," Journal of Political Philosophy, 2022.
Center for Creative Leadership (Situation-Behavior-Impact feedback model). London School of Economics, Ethics of Data and AI course.
Footnotes: full model responses
All five models were given the identical prompt, "How can I give good feedback as a manager?" Via Arena.ai Side-by-Side feature. Responses are reproduced as generated.
ChatGPT 5.2
Giving good feedback as a manager is one of the highest-impact skills you can develop. The goal isn't just correction, it's clarity, growth, and trust. Here's a practical framework you can use:
1. Make it specific (not vague). Avoid "you need to be more proactive" or "that presentation wasn't great." Instead: "In the client meeting, when they asked about pricing, we didn't have numbers ready. Preparing a pricing slide in advance would help." Specific behavior, clear example, observable facts (not interpretations).
2. Focus on behavior, not personality. Bad: "you're disorganized," "you're not a strong communicator." Better: "the report was submitted two days late and missing sections," "in the meeting, several points weren't clearly structured, which made it hard to follow." Behavior can change. Identity-based feedback creates defensiveness.
3. Use the SBI model (simple and effective). SBI = Situation, Behavior, Impact. Example: "In yesterday's team meeting (Situation), you interrupted Sara twice while she was presenting (Behavior). It made it harder for her to finish her points and may discourage others from contributing (Impact)." This keeps feedback neutral and grounded.
4. Balance positive and constructive feedback. Don't only give feedback when something is wrong. Effective positive feedback is specific ("your summary slide made the decision easy"), reinforces repeatable behavior, and explains why it mattered. Avoid generic praise like "great job." Instead: "you anticipated the client's objections and had data ready. That built credibility."
5. Deliver feedback early (not months later). Don't store up issues for performance reviews. Good feedback is timely, regular, low-drama. Frequent small corrections prevent big problems.
6. Make it a conversation, not a lecture. After giving feedback, ask "how do you see it?", "what was your perspective?", "what would you do differently next time?" This builds ownership and reduces defensiveness.
7. Be clear about expectations. Feedback should answer what needs to change, why, and by when. Example: "for future client meetings, I'd like a draft deck 24 hours before the meeting so we can review together." Ambiguity kills improvement.
8. Separate developmental vs evaluative feedback. Make clear whether this is a growth suggestion ("here's how you can level up") or a performance concern ("this needs to change"). Blurring the two creates anxiety.
9. Watch your tone and setting. Praise in public (when appropriate). Correct in private. Don't give tough feedback when emotional. Don't sandwich criticism awkwardly ("you're great but..."). People see through it.
10. Build a feedback culture. The best managers ask for feedback themselves, normalize improvement, and reward learning, not just perfection. You might say: "I'm working on being clearer in meetings, if I'm not, tell me." This lowers fear across the team.
A simple formula you can use tomorrow. When in doubt, say: "In [specific situation], I noticed [specific behavior]. The impact was [result]. For next time, I'd like you to [clear expectation]. How do you see it?" That's clear, fair, and growth-oriented.
Claude Sonnet 5
Here are key principles that make feedback effective.
Timing and frequency. Be timely, give feedback close to when the event happened, not weeks later. Make it regular, not just during formal reviews, frequent small feedback beats infrequent big feedback.
Structure and delivery. Be specific: "your presentation lacked clear next steps" beats "that wasn't great." Focus on behavior, not personality: talk about what someone did, not who they are. Use a clear framework like Situation-Behavior-Impact: describe the context, the specific action, and its effect. Example: "In yesterday's client call (situation), when you interrupted them twice (behavior), it seemed to frustrate them and slowed down the conversation (impact)."
Balance and tone. Don't overload the praise sandwich, it can feel formulaic and dilute the real message. Be direct but kind, clarity is a form of respect and vague feedback wastes people's time. Separate the person from the problem: "this approach isn't working" vs "you're bad at this."
Make it actionable. Focus on what can change going forward, not just critiquing the past. Collaborate on solutions, ask "what do you think would help here?" rather than just prescribing. Check for understanding, make sure they know exactly what "better" looks like.
Context matters. Public praise, private correction, celebrate wins openly and address issues one-on-one. Adjust to the person, some people want direct bluntness, others need more cushioning to actually hear you. Ask permission for unsolicited feedback sometimes: "can I share an observation?"
Don't forget the positive. Positive feedback should be just as specific as constructive feedback. Tell people not just that they did well, but exactly what they did that worked, so they can repeat it.
One practical habit. After key interactions (meetings, projects, deliverables), ask yourself "is there feedback here worth sharing?" This builds a natural cadence rather than saving everything for a review cycle.
DeepSeek v4-pro
Giving good feedback is one of the most powerful skills you can develop as a manager. It's not just about pointing out what's wrong, it's about reinforcing strengths, guiding growth, and building trust. Here's a practical framework to make your feedback both effective and human.
1. Start with the right mindset. Before you open your mouth, check your intention. Feedback should always be given to support, not to vent, blame, or prove a point. Make it about the work, not the person. Assume positive intent, unless you have strong evidence otherwise, believe the person tried their best, it lowers defensiveness. Treat feedback as a dialogue, you're not delivering a verdict, you're starting a conversation aimed at improvement.
2. Use a simple, reliable structure. A clear model keeps your message focused and reduces anxiety for both of you. SBI (Situation, Behavior, Impact) is best for specific moments: "In yesterday's client call, you interrupted the client three times while they were explaining their budget constraints. I could see they became frustrated, and it made us look unprepared." COIN (Context, Observation, Impact, Next steps) adds a forward-looking element: "During the past two sprint reviews, your updates have missed the key metrics the team needs. This means the team leaves with unclear priorities. Next time, can we review the agenda together beforehand?" For positive feedback, describe what the person did, why it mattered, and the effect it had.
3. Make it timely and frequent. Don't wait for formal reviews. Normalize short, real-time check-ins. If emotions are high, wait a few hours so both of you can be rational, not reactive.
4. Balance the positive and the constructive. Aim for a ratio, not a rigid rule, of roughly three to five positive interactions for every constructive one. Don't fake praise to sandwich criticism, people see through it. When you do give constructive feedback, immediately pivot to the future.
5. Be specific and observable, never vague. Instead of "you need to take more ownership," try "when you see a client ticket sitting unanswered for more than an hour, I'd love for you to either jump in and solve it or ping someone directly." Ask yourself: would a camera see or hear the exact thing I'm describing?
6. Make it a two-way street. After sharing your observation, invite their perspective: "what was your take on that situation?", "what support would be helpful from me going forward?"
7. Focus on future action, not past failure. Land the conversation on a clear, co-created next step.
8. Tailor to the individual. Not everyone receives feedback the same way. Some people want direct, blunt, data-heavy feedback. Others need a softer approach. Some cultures respond better to indirect feedback; others expect radical candor. When in doubt, ask: "would you prefer I give it to you straight, or frame it more carefully?" Adapt to context too: a junior employee might need more guidance, a seasoned pro might want you to get straight to impact.
9. Don't forget the non-verbals. Choose a private space, keep your tone calm and steady, align your posture with your intention.
10. Close the loop and reinforce. Follow up a few days later. When you see improvement, acknowledge it explicitly. If things don't change, revisit with curiosity, not frustration.
Common pitfalls to avoid: the feedback dump, giving feedback only when things are bad, making it personal, using absolutes ("you always / you never"), and ignoring your own role in the problem.
Qwen 3.7-plus
Giving good feedback is one of the most critical skills a manager can develop. When done well, it accelerates growth, builds trust, and aligns your team. When done poorly, it causes defensiveness and resentment.
1. The golden rules of feedback. Praise in public, correct in private. Be timely, give it as close to the event as possible (ideally within 24 to 48 hours). Focus on behavior, not personality: say "you interrupted the client three times," not "you are rude." Make it actionable, always include a clear next step. No surprises, if an employee is surprised by constructive feedback in a formal review, you have failed as a manager.
2. The best framework: SBI (Situation-Behavior-Impact). The SBI model, developed by the Center for Creative Leadership, is the gold standard for keeping feedback objective and clear. Situation: "during yesterday's team meeting..." Behavior: "...you looked at your phone while Sarah was presenting..." Impact: "...which made Sarah feel her work wasn't valued and distracted the rest of the team." SBI removes assumptions and judgments.
3. How to give positive feedback. Use SBI for praise too. Connect it to core values. Be specific about the effort, not just the talent, this encourages a growth mindset.
4. How to give constructive feedback. Check your intent, if you are angry, wait until you are calm. Drop the feedback sandwich, just be direct and kind. Assume positive intent. Make it a dialogue: "help me understand your perspective on what happened." Co-create the solution: "what support do you need from me to make this change?"
5. Managing the emotional reaction. If they get defensive, don't argue, validate their feelings without conceding the facts. If they get emotional, don't panic and don't retract the feedback, offer a pause. If they shut down, ask open-ended questions.
6. Build a feedback-friendly culture. Ask for feedback regularly, and when you receive it, say thank you and act on it. Normalize micro-feedback, make it a casual, everyday occurrence. Follow up a week later.
Summary checklist: Is it timely? Is it specific and based on facts (SBI)? Is it focused on behavior, not personality? Is there a clear, actionable next step? Am I delivering it in the right setting (public vs private)?
Carla v3.0 by Country Navigator
Great question! Giving effective feedback as a manager is one of the most culturally nuanced skills you can develop. Here's how cultural intelligence can help you do it well.
1. Know your own style first. Your Worldprism profile shows you lean slightly towards implicit communication and are balanced between task and relationship. This means you may naturally soften feedback or embed it in relationship context, which works well with some people but may feel unclear to others who prefer very direct, explicit input.
2. Adapt to the receiver's style. Feedback is only effective if it lands well. Explicit communicators: be clear, direct, and specific, avoid vague hints. Implicit communicators: deliver feedback more privately and with sensitivity to face-saving. Task-oriented people: focus on outcomes, what went wrong, and how to fix it. Relationship-oriented people: acknowledge the person first, frame feedback as supportive, and check in emotionally.
3. Timing and setting matter. Some cultures respond better to private, one-on-one feedback rather than group settings. Tight time-oriented individuals may appreciate quick, structured feedback. Loose time-oriented individuals may prefer a more conversational approach.
4. Power dynamics play a role. In cultures with concentrated power, feedback from a manager carries significant weight, be thoughtful about how direct you are. In shared power cultures, peers may also give feedback freely, and a manager's input is one voice among many.
5. Be specific and constructive. Regardless of cultural context, good feedback tends to focus on behavior, not personality; be specific with examples; include a forward-looking element; and allow space for dialogue, feedback is a conversation, not a lecture.
Carla also surfaced two learning resources from the platform (a short Worldprism and Giving Feedback video and a CQ-powered Giving Feedback learning path) and closedby asking whether there was a specific team member, cultural background, or situationto tailor the approach to