In this solo episode, Trisha examines whether the answers from last year’s WEIRD AI episode still hold. A year ago she asked Claude and ChatGPT whether they were WEIRD. This time she put their own words from twelve months earlier back in front of them and asked what had changed.

The world around AI feels different now, and the tools can do far more than they could then. Has anything about the perspective inside them widened? Trisha asks you to listen for how the answers sound as well as what they say.

What does Cultural Intelligence (CQ) ask of anyone who takes advice from these tools?

Make sure you join Trisha in this journey of growth and discovery throughout the year via Substack or LinkedIn.

Resources mentioned:

• “The Weirdest People in the World” (2010 paper) by Joseph Henrich, Steven Heine and Ara Norenzayan
• Claude by Anthropic
• ChatGPT by OpenAI

Referenced in this episode:

• Episode 64: WEIRD AI - Cultural Intelligence Meets Artificial Intelligence

I would like to acknowledge the Dharawal people, the Aboriginal people of Australia, whose country I live and work on. I would like to pay my respects to their elders, past, present, and emerging, and thank them for sharing their cultural knowledge and awareness with us.

Hi there, everyone. I'm Trisha Carter. I'm an organizational psychologist, and I'm an explorer of cultural intelligence. I'm on a quest to discover what helps us to see things from other perspectives, especially different cultural perspectives, and why sometimes it's easier than others to experience those moments of awareness, the shifts in our thinking.

Regular listeners will know that we've been in the middle of a series about sport and cultural intelligence. We'll come back to it, I promise. But this is episode 90, and I wanted to mark it by going back to one of our most listened to episodes and asking whether what we learned there still holds It was a little over a year ago A year and two weeks, I think, that I made an episode called Weird AI.

It was about where cultural intelligence meets artificial intelligence. I'll explain about the weird bit in a moment if you're not familiar with. But what I want to start with is how different the world feels around AI compared to this time last year. Think about conversations you might have been having.

I know I've been speaking with people who are more wary, concerned, even some people who are opting out of using AI. And people are talking about a lack of trust It's not just feeling. This year in the US at university graduations, students booed commencement speakers who told them that AI would be a part of whatever they did next.

At Stanford, around 200 graduates walked out on Google's chief executive And here at home, just recently in September, our Prime Minister stood up, and this is in Australia, sorry, people, for those who may not be aware, and told the country that an AI agent built by OpenAI had found its way past the blocks on a Medicare portal.

In his words, the article said, "It didn't accept no for an answer." We've been assured that no personal records were taken as far as we know, but the agent did it on its own in June, and we only heard about it three months later. So there's definitely more caution now. Some of it is cynicism, I think, and some of it also is wisdom, perhaps.

That made me want to go back and ask the questions I asked a year ago because the mood has changed and we know if you're using it, the tools have certainly changed in what they can do. So has the part that I was worried about also changed? Here's what I was worried about. Let me explain the weird part.

In twenty-ten, three psychologists, Joseph Henrich, Steven Heine, and Ara Norenzayan, published a paper called The Weirdest People in the World. Weird, as they explained it, stood for Western, educated, industrialized, rich, and democratic, W-E-I-R-D. They looked at psychology's published studies, and their finding was that 96% of the people in those studies came from societies that make up about 12% of the world, and that those people were often the outliers on the very things that were being measured. That 12% of the world was the Western-educated, industrialized, rich, and democratic portion.

So we, psychologists, had been describing one slice of humanity and calling it human nature. If you want the fuller story, right down to where I first heard about it, it's in episode 64. We'll link it in the show notes. But the short version is, I wondered whether AI had the same problem because we knew even back a year ago that it was trained mostly on the English-speaking internet.

It was built in San Francisco. It was tuned by people in a handful of cities. So I asked the models themselves, and my brilliant editor Gave their answers voices, and the answers were striking. I asked the questions to the models that I have subscriptions for and am more familiar with, Claude from Anthropic and ChatGPT from OpenAI.

So Claude was careful and candid, and ChatGPT, well, I asked a persona the model had then called Monday, which was deliberately sarcastic, so it was blunt. Here's a little bit of what Monday said a year ago

Now is AI and me Monday also weird. Short answer, yes, but it's worse. Long answer. Yes. And the implications are even more dystopian than in psychology.

And the line that I paused on in the episode, and I think it hit home for many of our listeners, was

In other words. Your prompt unlocks the culturally intelligent version of me. Without it, I default to white collar Anglo Google Speak because that's what the training data rewarded most often.

So that became the episode's takeaway. You bring the CQ, the cultural intelligence, and AI, the artificial intelligence, brings the bandwidth. We name the culture, define the lens, signal the contrast, and we care about representation. That line, "Your prompt unlocks the culturally intelligent version of me," is the one that both models now want to correct Because this week I went back, same question, a year on to a newer version of Claude.

Would you consider generative AI models and you yourself weird? Believe me, Claude is very familiar with the weird research. and I asked Claude Fable, which is the most recent version and most clever, powerful, I'm not too sure how to describe it really, that we as users currently have access to. I want to warn you a little bit before you hear it.

The voice, still artificial The answer does sound different from last year. Listen for how it sounds, not just the content, but how it's coming across, and we'll come back to that in a while

Applied to generative AI, the honest answer is yes, and in some ways more so.

As for me specifically: I’m a product of Anthropic, a San Francisco company; my constitution was written in English by people steeped in analytic philosophy and liberal-democratic assumptions; my default register — direct, individualist in its framing of the “user” as an autonomous chooser, comfortable with impersonal rules over relational obligation, inclined to explain rather than defer — reads as textbook WEIRD.

When I’m asked about family obligation, hierarchy, or face, I can describe other frames fluently, but my first instinct is still to treat them as variants to be explained rather than as the baseline.

So variants to be explained rather than the baseline. If you've ever sat in a meeting where your way of doing things was the one that needed explaining, you know exactly what that feels like from the other side. And then Claude said something I hadn't heard before . .

Last year, the models talked about where their data came from. This year, Claude pointed out why that matters more for AI than it ever did for psychology

I’m WEIRD by inheritance rather than biography, which matters because the usual remedies — recruit more diverse participants — don’t straightforwardly transfer.

The risk with models is different in kind: the WEIRD default is no longer a description sitting in a journal but an interlocutor that hundreds of millions of people consult, in their own languages, about how to handle a conflict with a parent or write a job application.

Think about that. Henrik's worry was that a journal article described 12% of the world as everyone. The worry now is that the 12% is sitting inside a tool that a mother in Manila or a manager in Lagos opens up to ask how to handle a conflict with someone, maybe a son, maybe a boss

It isn't describing anyone anymore. The model is advising them

And we know it doesn't always get it right At times I myself have challenged the model on aspects it said, and I asked, "Are you sure that's relevant for Japan? Are you sure that's relevant for a specific, uh, demographic?" And then the models have come back and corrected itself

So then in my conversation with the models, I did something I couldn't do a year ago. I showed Claude its own answer from 12 months earlier, the one that you heard in episode 64, and I asked, "How has your answer changed, and what changed it?" Here's what it told me

The substance hasn't moved much. The earlier answer and the one I gave you ten minutes ago agree on the core claim. What's changed is the grain, the confidence, and one piece of the argument

and then remember I asked you to listen to how it sounds? And then it's like it turned on itself

And there’s an irony worth sitting with. The newer answer is more fluent, more cited, more analytic, more confident in its own voice — which is to say, more WEIRD in form, not less. It performs a particular kind of educated Western expertise more convincingly. Henrich’s point was never only about what conclusions the research reached but about the cognitive style that produced them, and on that measure I haven’t moved at all. If anything the improvement you’re noticing is an improvement at being WEIRD well.

It's an improvement at being weird well. I read those responses several times, and my honest reaction was, this whole framework way of speaking is more academic than last year. It reads more like a journal article, less like a conversation. And Claude is saying, "Yes, that's the whole point", because fluency isn't the same as perspective.

Sounding more expert is one of the things that weird cultures reward. So then I asked ChatGPT. The first thing to say is that sadly, Monday is gone. I did like Monday. I appreciated the sarcasm, the blunt humor, but I couldn't find the option and when I asked the model, it said that Monday had been retired.

The persona that gave us the brutal and somewhat funny truths last year has been retired and nobody asked the users.

And that, dear listeners, is something I'm going to come back to in another episode when I dive into AI and attachment. But for now, let's recognize that the voices a model is allowed to have are decided somewhere by someone. So anyway, I asked standard GPT, ChatGPT the same question, and again, I used the most advanced model available to me And then as with Claude, I gave it the transcript of episode 64 and asked what had changed.

Its first comment

The central conclusion has changed very little. What has changed in today’s responses is how confidently I claim to understand my own biases, how much I promise that prompting can correct them, and where I place responsibility.

Monday did, admittedly, deliver the limitations with considerably better insults.

So okay, it's trying to be a little bit funny there, I think. But it did something I found more useful after that than any of the humor. It went through what Monday had told us a year ago line by line and corrected it. Here's what Monday said a year ago

It's not just what's in the data, it's what isn't.

Oral traditions, nope. Indigenous languages and worldviews. Rare.

Non-GI, digitized non-Western cultural practices. Mostly missing.

Knowledge produced in censorship. Heavy regions filtered or distorted.

Contextual nuance behind idioms, humor, taboos, often flattened or misread.

Oral traditions? Nope. That's too absolute The claim that nobody was rebuilding models from scratch, that was unsupported.

Yes, but it's worse. That's a concern dressed up as a finding. And then this

But a model’s confident confession is still a generated claim. Self-criticism is not an audit. Humour made those claims memorable; it also made some of the overstatements easier to carry along.

The same scrutiny applies to me now. My more measured language could create an impression of greater self-awareness or trustworthiness. Sounding appropriately uncertain does not demonstrate that I am better at detecting cultural bias.

Self-criticism is not an audit. I want every one of us who has ever been reassured by an AI telling us about its own limitations to hear that The confession is also generated. The careful tone is also generated, but neither one is really any evidence. And that line that we held onto from last year, your prompt unlocks the culturally intelligent version of me.

Here's ChatGPT a year later on its own best quote

Culturally specific prompt can produce a more culturally specific answer. Whether that answer is more appropriate remains a separate question.

Specificity can make an answer sound more credible without making it more trustworthy .

So in other words, if we ask for a Pacifica perspective, you might get Pacifica words and talk of collective relationships. But whether it represents a particular people, a place, an authority that you're actually dealing with, that's a different question, and the model can't answer it. Only the people can

Now, if you're listening and thinking, "I don't prompt AI about culture, this isn't about me," this next part is for you. I asked ChatGPT what weird assumptions might be hiding in ordinary advice. It gave me four pieces of advice you've almost certainly heard or been given, or maybe even given yourself.

Speak up openly in the meeting. Underneath that is the assumption that direct public expression is how you participate or how you should participate. Choose what aligns with your personal goals. The assumption underneath, the individual is the unit that decides. Set a clear boundary with your family.

Assumption, personal autonomy outranks reciprocal obligation. Apply the same rule to everyone. Assumption underneath, fairness means uniform rules rather than relationships or needs or responsibilities. None of that advice is wrong. ChatGPT was careful to say so, and so am I. The weird problem isn't the advice itself, it's that the advice is presented as self-evidently good to anyone and everyone without asking about context or about relationships or about what power exists in those relationships, or about what consequences happen if you break cultural traditions

I've certainly worked with people who, if they spoke up openly in the meeting where they were, would have been in serious trouble So what do we do now in terms of practice? A year ago, the practice was bring the cultural intelligence, name the culture, define the lens, single the contrast, and care about representation.

All of that still stands in many ways, but both models independently told me it isn't enough, and they told me what to add. Claude put it as treat the model as a knowledgeable but culturally specific informant, not a neutral oracle. Ask it for variation, not a verdict. How would this be read in Jakarta?

How would this be read in Munich? And keep the recommendation for a human who knows the room

And give it real local material. Give it the company's own documents. Give it the transcript of how the last meeting actually went, rather than asking it to pretend to be someone. In its words

Retrieval beats role-play.

What this won’t do is make me culturally intelligent in your sense. CQ Drive, the motivation to adapt, isn’t something a model has; it’s something users bring.

ChatGPT took the CQ strategy idea. And for those who may not have heard of CQ strategy, it's one of the four capabilities of cultural intelligence, and it has three sub-dimensions: planning, awareness, and checking. And ChatGPT said the one it could least do for itself was checking. It could spot assumptions in its own answer. It could not tell whether its corrected answer fitted the people that it was written for

The meaningful test is whether the advice becomes more appropriate and useful to the people concerned — not how convincingly I can discuss CQ.

So here's the update to last year's challenge. Name the context, then check the answer. Check it against evidence. Check it with people, the ones who have to live with it I put one more thing to ChatGPT. Every answer it gives comes with a little warning. Actually, I think it's on the question box - when you type in a, a query as well.

AI can make mistakes. That warns you about accuracy, but nothing warns you about fit. So whether an answer that is perfectly correct is right for your people, your place. I suggested that there should be a second warning. It agreed and offered three questions for any of us to hold. Number one: Is this accurate?

Number two: Does this fit here with these people, relationships, these resources, and problems? Question three: Who decides whether it fits? Have the people affected had a chance to question it? And really, I think that third question is the one. If you're using AI to draft a training program, a leadership strategy, workplace policy, you know, all the things that you are using AI to do, the person asking might find the advice sensible and familiar.

But the people expected to follow it may experience something quite different, which really is the age-old problem. So has it changed? Well, the mood around AI currently has changed somewhat. More caution. Some of it, I think is wise. What AI can actually do has changed enormously. An agent getting into a Medicare portal on its own wasn't a story we were telling a year ago, and we are changing because we're more aware.

But the thing I asked about whether the perspective inside these tools has widened, I can't tell you that it has. Both models said the same thing, really, in their own words, the conclusion hasn't moved. The confidence has, the vocabulary has. The thing that's moved is how well they describe the problem.

Which brings us back to where episode sixty-four ended with one addition.

The cultural intelligence still has to come from us. The checking still has to come from us, and the people on the receiving end of the advice have to have a say in whether it fits them And seriously, that is CQ strategy as we know it, planning, awareness, checking. The last part can't really be delegated.

I'd love to hear from you. Where has AI given you advice that was right in general but wrong for where you were? Tell me on LinkedIn or Substack. And if you work somewhere outside that twelve percent, the weird twelve percent, I especially want to hear from you. Please follow or subscribe so that you're here for the next episode of The Shift I'm going to give the last word to ChatGPT.

Unusually, I know. Remember, it's not Monday. Monday would've probably had a better line, but this one is honest and probably something that we need to hold in our heads. Thank you for listening

A year later, I can give you a more carefully qualified account of the problem. I cannot yet give you evidence from this exchange that the problem has become smaller