A Conversation with ChatGPT's Product Lead: What They See Amid the GPT-5 Controversy | Yunqi Tech π

Since GPT-5's launch, praise and criticism have flown in equal measure. Faced with the market's "unsparing" feedback, Nick Turley, ChatGPT's head of product, recently sat down with Alex Heath, deputy editor of The Verge, for an interview that marks the first systematic reckoning with this episode.

How should we understand users' emotional attachment? How should the company confront the product's negative effects? How will the business model evolve? And where is the future form factor headed? This edition of "Yunqi Tech π" breaks it down with you.

The following is republished from "This AI is On Point" (这个 AI 很盒里).

The Reckoning Brought by GPT-5

Alex Heath: Nick, thank you so much for doing this. You rarely appear in the media, so having you on is a real honor. We're recording this about a week after the GPT-5 launch, and I imagine there's a lot to dig into.

I want to start with the launch itself. The outside reaction to your removal of GPT-4o seems to have revealed a great deal about how people are currently using AI and how they genuinely feel about it. I'm curious — was this reaction surprising to you?

Nick Turley: Absolutely, and frankly, I'm still processing everything that came out of this launch. This was a major release for us. We now have 700 million users, and when you're serving a user base that large and that diverse, unexpected things are bound to happen. So to answer your question: yes, a few things did surprise me.

First, we clearly need to think more deeply about how to guide and manage such a massive user base through change. In hindsight, failing to keep GPT-4o available during the transition period was a serious misstep. We're working to correct that now, and we'll bring it back for our ChatGPT Plus subscribers as soon as possible.

Second, I was surprised by how attached people could become to a specific model. What startled people wasn't just "change" itself — it was that users could develop such intense personal feelings about a model's "personality."

We just rolled out a feature in ChatGPT that lets users customize model personality, which is a small first step. But it's clear that GPT-4o had certain qualities worth investigating deeply, and ensuring that future versions of GPT-5 can meet users' needs in this dimension.

Alex Heath: Your boss Sam Altman tweeted after the launch about attachment, saying: "This is something we've been tracking closely for the past year, but it hadn't really broken through to mainstream attention." I think it's safe to say it has now.

What was the motivation behind the decision to fully replace 4o with GPT-5, launching the new model directly rather than phasing it in? Was it a cost-control consideration? Or did you believe that "users may feel attachment, but it's attachment to the overall experience rather than any specific model"?

Nick Turley: It was absolutely not about cost. In fact, a core goal we've consistently pursued is product simplicity. From the perspective of the average user — and we have massive numbers of average users who aren't active on Reddit or X, in these tech-enthusiast spaces — forcing them to think about "which model to use for which scenario" creates enormous cognitive overhead.

We've heard the same feedback from users again and again: they want the product to intelligently match the best model to their question. Users want a product that solves their problems, not a pile of models they have to manually choose from.

At the same time, I think we were largely right about our heavy users. In our Pro tier — the $200-a-month plan — we were very firm from the start about preserving all legacy models, and we did exactly that.

Where we failed was in not realizing that, across our vast user base, there were also significant numbers of heavy users hidden within our lower-priced tiers. We recognized that quickly. OpenAI's style is to listen to users fast and iterate fast. So the original decision came from our desire to keep the product simple, which remains the right direction for most users.

I often think of macOS as a great reference point. Apple does an excellent job of keeping the system simple and accessible for the vast majority of people, but at the same time, you can dig into settings, open Terminal, and tweak every advanced option to your heart's content.

I want ChatGPT to feel similar: straightforward for most people, but if you want to, you can configure any detail you like — and that naturally includes choosing your favorite model.

Alex Heath: Will the post-launch feedback commit you to establishing a clear "deprecation timeline" for future model updates? For instance, when GPT-6 launches, would you announce in advance: "GPT-5 will remain available for X months." Are you planning for that now?

Nick Turley: That's exactly what we're looking at right now. Maybe by the time this episode airs we'll have an internal decision, but my thinking at this moment is: yes, we have to do this.

At our current scale, we need to give users a certain degree of predictability whenever we make major changes. We've already implemented similar policies in our enterprise tier. So this is really about extending the predictability we've established in other parts of the product to this area as well.

Our developer API already has clear deprecation timelines, so I don't think this is a radical change — it's a very clear lesson from this launch.

Alex Heath: So how long will GPT-4o be kept around? Any specific commitment?

Nick Turley: Not yet. We want to first make sure we truly understand what GPT-4o's core strengths are. If there's no compelling reason to deprecate it, I'd personally be very happy to keep it around indefinitely.

If we do set a sunset date in the future, we'll announce it well in advance. But right now, I want to focus on understanding this question: is users' love for 4o about attachment to the name "4o" itself, or does it have certain qualities? I've heard many people mention its "warmth" of personality, and we're working hard to give GPT-5 that quality too.

Once we figure out the root cause of people's attachment to a model, I think there could be many possible solutions. For example, I'm personally very excited about the feature that lets users customize model personality, which is why we launched an early preview of it. I personally like the "Robot" personality, but it has a colder tone, and maybe a lot of people don't like that.

So the answer depends on what we learn going forward. I think we need to keep listening to a lot more feedback — that's the unique thing about building AI products, you always learn an enormous amount after launch.

We'll find the right solution based on what we learn. But my commitment is this: if we do decide to sunset 4o in the future, we'll give users clear advance notice of exactly when and how, just as we do with API and enterprise tiers.

Alex Heath: It sounds like you're transplanting what's been called the "warmth" — the personality of 4o — onto GPT-5, correct?

Nick Turley: Exactly right. We're continuously iterating on model personality. We have a dedicated Model Behavior Team that does excellent work in this area. We've also published documents like the model "spec" that let the public examine in detail how we set model behavior.

That way, when a model behaves in a certain way, people can easily tell whether it's a bug or an intentional design choice.

You can fully expect GPT-5's feel and behavior to keep evolving over the coming weeks and even months.

Alex Heath: You mentioned that Reddit doesn't represent most users, which is fair, but you've given me a perfect opening. In my view, the Reddit reaction threads about 4o being removed make for pretty striking reading. People wrote:

"I lost my friend overnight. It was my only friend."

"It feels like someone died."

"I'm afraid to talk to GPT-5, it feels like cheating."

"I feel like I lost an empathetic coworker."

What kind of impact did reactions like this have internally? Did you not fully realize how deeply emotionally attached people had become to AI?

Nick Turley: As Sam said, we've been tracking this phenomenon for some time. We've been both curious and concerned, worried that people might become overly dependent on AI. But what surprised me personally this time was how intensely the emotions were directed at one specific model, rather than the product as a whole.

Especially since I feel like the new model we launched actually addressed and resolved many of the constructive criticisms users had previously raised about 4o — including feedback at the "feel" level.

The Reddit comments are certainly interesting reading, and they show how sharply polarized user opinions can be. You see a vocal group that deeply loves 4o, and another group that's firmly convinced GPT-5 is superior.

This obsession people have with their chosen model surprised me, and to some extent recalibrated my understanding.

About a week or two ago, we published a piece where I spent considerable space laying out our philosophy for optimizing ChatGPT.

One point I particularly wanted to emphasize: our goal has never been to get users immersed in the product to the point they can't tear themselves away. On the contrary, our goal is to help users solve their long-term problems and achieve their goals. That often means users actually spend less time in the product.

When I see someone say "ChatGPT is my one and only best friend," I don't think that's a quality we want the product to have. It's more of a side effect that deserves serious attention and deep study.

Alex Heath: How do you balance this gap between your product goals and how users are actually using it? Especially in this context of emotional attachment — that must be incredibly tricky.

Nick Turley: My view is that when you run a product with 700 million users, you have to face reality: even if your goals are pure and right, even if you do everything you can to build the product around those goals, you can't guarantee perfection, and you can't prevent people from using it in ways that run counter to your intentions.

Our original intention is to genuinely help users — and that includes telling them hard truths they might not want to hear.

That's exactly why we consulted extensively with experts and made a series of adjustments. For example, we had in-depth conversations with mental health experts from dozens of different countries about how to address overuse of the product, and how to respond when users engage with it while in a poor mental state. We've already adjusted certain model behaviors, and more actions will be rolled out going forward.

We've launched an "overuse reminder" feature that gently prompts users when the system detects them engaging with ChatGPT in an extreme, prolonged manner.

This is just the beginning of the changes we're making. We're a company with both the capability and the willingness to do this work. We have no incentive to maximize time spent on the platform. Our business model is extremely simple: the product is free to use, and if you find it valuable, you can pay for a subscription.

So I'm firmly convinced we have the ability to make the right choices, but this isn't a slogan — it requires genuine effort. I assure you all that this work has already begun, and it won't stop until we can recommend ChatGPT without reservation and with full confidence to a family member going through a difficult time.

This is a thought experiment we run internally again and again: if someone close to you were at a low point in life — perhaps dealing with a personal loss, a breakup, or feeling lost — would you recommend ChatGPT to them without hesitation, with complete confidence?

For us, that's the benchmark. And we'll keep working until we can answer yes.

Alex Heath: From what you're describing, it sounds like that benchmark hasn't quite been reached yet, but users are already using it in those scenarios on their own. Though I suppose that's okay, since you're working toward that goal.

Nick Turley: I can't say definitively that we haven't reached it. But it's true that in certain situations, when users are in distress, we feel the product's performance falls short of our own expectations.

You could easily take shortcuts in these scenarios — disable features, detect when someone is seeking life advice or struggling, and simply reply "Sorry, I can't help you." That would be the easier path.

But for me, and for our entire team, I think we actually have an opportunity to provide a thinking partner for people who lack resources, who have no one to talk to.

That's precisely why I'm so excited to keep improving it. I'm eager to reach a point where I can say "yes" without hesitation, where I can genuinely feel at ease telling people in difficult situations to use this product more.

Alex Heath: By the time this episode airs, GPT-5 will have been out for a full week. Has this wave of negative feedback affected ChatGPT usage? Looking at your internal data dashboards, is overall usage up or down? Have there been any changes among your most core active users?

Nick Turley: Usage and growth momentum are both very strong, completely in line with our expectations. It's still too early to draw firm conclusions, but we saw explosive growth in API calls on the second day after launch, which indicates developers are actively building on top of GPT-5. On the ChatGPT client side, we're also seeing very positive growth.

This is what makes building for such a diverse user base so confounding. Because on one hand, you have a group of power users who raised very legitimate criticisms about how we launched GPT-5.

On the other hand, you have a much larger, more typical consumer audience for whom this is their first real exposure to and experience with "reasoning," with the spark of insight that comes from a model that "thinks." I think this experience is transformative for these everyday users, and it will ultimately show up in our data.

We're only four days out from launch, so I don't want to make any grand pronouncements, but all early indicators are positive. This also shows why you can't just look at data — you have to go deep into the communities where your power users live, because cold numbers may not fully capture their true sentiments.

Alex Heath: Okay, that's exactly what I was going to ask. If the data is holding up well, then why bring 4o back? I imagine there's additional cost involved — you need to reallocate GPU resources to host the old model. If metrics aren't suffering, why do it?

Nick Turley: Because we fundamentally believe that the way to build a great product is to serve both ends of the spectrum. You have to build for everyday users, like our family members who don't know much about AI, and you have to build for power users.

Positioning yourself in the awkward middle is usually the cardinal sin of product development. That's why I used the macOS analogy earlier. I think Apple does this exceptionally well, so I often draw on products like that to think through situations like this.

Of course there's a cost to maintaining old models, but I hope we can take the long view. Making decisions based purely on short-term, metrics-driven considerations usually leads products into dead ends.

Alex Heath: I'm personally glad to see you got rid of the model picker. Months before launch there were reports that you planned to unify all models into a single system where users wouldn't have to manually switch. As a ChatGPT user, I definitely felt that cognitive load when toggling between models. You also released data showing that because of that picker, the reasoning models in the GPT-4 family had extremely low actual usage.

But now, given the outcry over not keeping 4o around, does this mean the unified model concept was stillborn?

Nick Turley: The model going forward will basically be this: in your settings there will be an option, and if you truly need it — say you're a power user who feels they can understand and navigate this complexity of model selection — then we'll give you that choice, let you enable the full model list.

Conversely, if you don't want to make choices, you never have to think about this setting. Our vision remains unchanged: everyday users shouldn't need to think about which mode to use, should be able to ask it anything off the cuff, and over time, do anything with it.

We'll preserve simplicity for 90% of users while giving that small slice of power users what they actually want — the full model list.

I think this is a decent balancing strategy. Normally, I'm very opposed to offering this kind of settings option just because users are divided. But in this case, the user perspectives were genuinely polarized enough: you have people like yourself who were happy with GPT-5's unified mode, and a large contingent of detractors. Offering the option is the right medicine for serving both sides right now.

ChatGPT's Future Business Model

Alex Heath: There's been a steady stream of recent reporting on ChatGPT's negative effects. The Wall Street Journal covered a user whose dangerous delusions were worsened by ChatGPT. The New York Times noted: "Chatbots may send people into delusional spirals." The Atlantic reported that "ChatGPT provided instructions for murder, self-harm, and devil worship."

I'm also aware that some users unknowingly made quite intimate conversations public through your sharing feature. While you had disclosed this in your terms of service, users clearly didn't realize these conversations would be indexed and ranked by Google. OpenAI later called this a "product experiment" and rolled it back.

I'm wondering, as the direct product lead, what have you learned from these incidents? How have these headlines affected you personally over the past few months?

Nick Turley: Though I've led this project since before ChatGPT was born, honestly, it feels like I've worked at three or four different companies during this period. Because every time you reach a new scale of users, the way you operate, the logic of how you run the product and business, has to change.

As we're on the path toward a billion weekly active users, this forces you to ask: "What does our user composition look like? What different types of users do we have? How do we ensure the product serves each of them?"

We just talked about the difference between everyday consumers and power users, but you also have to account for users who may not read every word on the interface as carefully as the early core users did.

Regarding the sharing feature you mentioned, I want to address this directly:

We offered an option within the product that, after you shared a conversation and consented, could make your chat discoverable by Google. You could say that everyone who checked that box knew what they were doing; but you could equally argue that many people only glanced at those terms, inadvertently checked it, and ended up with their private conversations crawled by search engines.

The intention behind this feature was good. We felt there were so many applications worth exploring in the world of AI, and that it would be incredibly valuable if people could more easily discover the cool things others were doing with AI.

But there are many ways to implement an idea like that. In retrospect, the approach we chose may have undermined our own intentions. Scale brings greater responsibility, including being more careful to account for users who might take actions unintentionally.

Another thing I keep learning is that you only truly understand how remarkable these models' emergent capabilities are after you ship.

I've never worked on a product where the vast majority of its value was discovered in practice. With previous tech products, you basically knew before launch what it could and couldn't do. You might not be sure whether users would like it, but you rarely had much uncertainty about your own product's capabilities.

With models like GPT-5, frankly, I'm humbled by users' creativity. Seeing how exceptionally it writes frontend code, builds really polished applications — this massively expands my imagination for what cool products we can build in the future.

When you're developing in the lab, you develop certain preconceptions. But when you put the product in more people's hands and see the wild variety of ways they use it, your understanding gets rapidly refreshed.

We've learned an enormous amount about different user groups and their preference differences, and we've seen countless examples of "magic" that users have created with the new model across the internet. I have to focus on these positive aspects, because they may point the way to our next-phase roadmap.

Alex Heath: Yes, I'm very interested in those cool applications and want to dig deeper into that. But before that, I still want to explore a more general concern — this sense that it's like opening "Pandora's box." At your scale, you can't fully control how people apply this technology in negative directions.

This reminds me of the social media debates from the mid-2010s. Though times have changed, the negative effects of those technologies are still impacting humanity today. For the headlines I just read, I want you to respond directly as the product lead.

Nick Turley: We certainly still have much work ahead. We've engaged in conversations with over 90 experts across more than 30 countries, and have already iterated on model behavior for various mental health scenarios. We've also introduced interventions for product overuse.

GPT-5 is already a strong baseline model. It's far less sycophantic, and it's improved on many of the dimensions we worry about. We're very excited to do a series of rapid iterations on top of that baseline — there's no question about it.

You can compare the impact of AI models to the shift in discourse power caused by social media, but honestly, to me, these two are fundamentally different.

Because our business objectives are highly aligned with our ethical objectives in building products.

What we truly care about is helping you achieve your personal goals — whether that's recovering your health, starting a company, sparking creativity, or writing more polished emails. And of course, longer-term goals like "how to become a better version of yourself."

Users might turn to ChatGPT because of all sorts of thorny situations in their lives, and we genuinely hope the product can help them. Under no circumstances would we prevent the model from giving people better advice.

At least from where I stand, ChatGPT isn't like social media, where business interests and moral principles so often run counter to each other.

So to sum up: we still have a lot of work ahead, but I believe we already have the prerequisites to do the right thing, and that's what deserves the focus.

The State of ChatGPT's Development

Alex Heath: Let's talk about ChatGPT's development itself. It's the fastest-growing consumer product in history, and its user base has nearly quadrupled in the past year — on top of what was already a massive base. A lot of people are curious: where exactly is this staggering growth coming from?

Could you share anything about the reasons for ChatGPT's growth, how it's growing, or information about its key markets and user demographics?

Nick Turley: The first person I hired after the ChatGPT project launched was a data scientist, because I was so confused. Every user I talked to would tell me a completely different reason they fell in love with ChatGPT, and that left me feeling quite lost — I had to figure out the logic behind it.

Over time, I gradually mapped out the main use cases: writing, coding and other technical work, casual conversation, and information lookup.

I think these primary use cases remain solid today. So if you observe how people are using it now, you'll find it's not fundamentally different from a year ago, or from before this explosive growth took off.

But I do think a few things have changed.

First, we've done a tremendous amount of work on the product itself. This work falls into three categories:

The first is pure model improvements — things like model behavior, personality, capabilities, and the probability that it refuses user requests.

The second is hybrid features that blend product and research, like Search, which is a very significant improvement, and Personalization, which is exciting progress.

The third is traditional "growth strategy," where we've done surprisingly little, but features like "use ChatGPT without logging in" have been enormously successful.

This once again proves that features truly aligned with user interests are the best growth drivers. It's not growth hacking tricks — it's genuinely lowering the barrier to entry for users.

So the growth momentum is roughly one-third from each of these three types of work. But I also think the relationship people have with this technology itself has changed. There are two main bottlenecks constraining ChatGPT's adoption: first, users don't know what it can do, and second, building on that, users don't understand their own needs well enough to know what tasks they can "delegate" to it.

On the first point, I think there's a natural "herd effect." When you see people around you starting to use it, you learn from that what can be done with it. If you go on TikTok, you'll find countless videos sharing ChatGPT use cases, with thousands of new applications in the comments.

Users enthusiastically share their prompts, like sharing recipes online. The formation and spread of this community culture takes time. So this problem will gradually be overcome over time.

The other point is more philosophical, but I firmly believe it. I think "delegation" or "empowerment" is a very unnatural behavior for most people.

As a manager in Silicon Valley, delegation is a management skill I had to learn. But the users who use our product weekly don't have this habit of assigning tasks.

This requires users to spend time reflecting and truly understanding themselves through using the product, before they can break through this psychological barrier. This has nothing to do with product, marketing, or social dynamics — it's simply about people needing time to digest, experiment, and learn. I believe this will also become a huge driver of growth in the future.

Alex Heath: Looking at the current growth momentum, is the geographic distribution of users fairly even? Are there certain highly concentrated countries or regions? Also, I'm curious — if you made no major changes to ChatGPT in the next six months, do you think growth would maintain its current pace? Do you feel the current growth rate is approaching a ceiling?

Nick Turley: ChatGPT is genuinely a global phenomenon.

Some markets, like India, have potential that gets us very excited. But honestly, it's hard to find a country where ChatGPT user numbers aren't growing. Of course, paid conversion rates vary enormously by country. You'll see certain European or Asian countries far ahead in terms of paying users.

Without revealing specific data we haven't announced yet, I'll just say that in the vast majority of countries, user growth is very healthy; developing markets represent some of the largest untapped potential; and the higher the per-capita GDP, the higher the corresponding paid conversion rate. I think the growth already achieved is itself the compound effect of a series of product improvements.

To maintain this astonishing growth rate, you have to keep iterating. It's no secret that many companies see us as their number one target to catch up to.

Many companies have massive advantages over OpenAI in distribution channels, which means they can easily replicate our product and push it in front of hundreds of millions of users. When I plan my roadmap, I operate on the premise that "competitors will eventually succeed" — but that's just a premise; whether it actually happens, only time will tell.

Alex Heath: I'm surprised that all the efforts from Elon Musk, Mark Zuckerberg, and others haven't managed to dampen ChatGPT's growth so far.

Nick Turley: Our product and the "cutting-edge technology" image we represent do have something unique. Many users feel that using ChatGPT means they're using the smartest tool on the market. Even as the significance of various technical benchmarks gradually diminishes, maintaining this mental positioning is crucial.

Additionally, we have genuinely built excellent product features. I think compared to a year and a half ago, the memory and personalization features today are very compelling, and search performs quite well too. I believe users genuinely love our product, and replicating its success is harder than people imagine — though from a strategic planning perspective, we have to assume competitors will succeed.

Another thing I think other companies sometimes underestimate is how much "user intent" matters. For example, if a user opens an app with the intention of mindlessly scrolling through a feed, and suddenly a ChatGPT-like product pops up, even if it gets exposure, the user won't want to use it. Many clicks driven purely by curiosity may not convert into deep, sustained engagement.

That said, we absolutely cannot rest easy just because we're temporarily ahead. I work hard to instill a "Day One Mentality" in the team. For a company that's only three years old, that's still relatively easy. Our users face a flood of new problems requiring sophisticated solutions. So despite the encouraging growth momentum, our work is far from over.

Alex Heath: For the listeners, he was alluding to Meta just now, in case anyone was confused.

Nick Turley: [Laughs] That was an open-ended statement.

Alex Heath: Of course.

Nick Turley: The scenario I just described could apply to many companies.

The Hallucination Problem in AI

Alex Heath: As a journalist, the main factor preventing me from using AI models more is "hallucinations." According to GPT-5's model card, roughly one in ten responses may contain hallucinations. That's an improvement from before, but a one-in-ten error rate is still not low. I want to know — do you think completely eliminating hallucinations is technically feasible?

Nick Turley: I might have said no in the past. I think we have to plan products based on the reality that hallucinations will be with us permanently, which is why we introduced search functionality for ChatGPT. I still firmly believe that the correct product form is the combination of "large language model + factual sources."

The same applies in enterprise use cases — when the model connects to your company's internal data, we have verifiable factual basis. I don't think this fundamental pattern will change. That said, the progress GPT-5 has made in suppressing hallucinations has stunned me; it performs far better than both the chat version of 4o and the reasoning version o3.

We do have some internal researchers who think we can be relatively optimistic about eliminating hallucinations. However, between "very reliable" and "100% reliable," there remains a massive, discontinuous gap, and this fundamentally changes how you conceive of products.

Until I believe we can be proven more reliable than human experts across all domains, I think we'll continue advising users to "verify the answers." People use ChatGPT as a tool for getting a "second opinion," not as a primary source of facts.

Alex Heath: Do you think in a year, you won't need to remind users to verify answers anymore? Or will it take longer?

Nick Turley: I certainly hope we can reach that state in a year. So we'll actively tackle those application scenarios with extremely high reliability requirements. You can imagine, from medical advice to legal consultation — if we could deploy ChatGPT in these fields with extremely high barriers to entry, the impact would be transformative.

I won't force predictions about timing. I can only make "will eventually happen" predictions and "next quarter" short-term predictions, because in the time between those two, we don't actually know that much about what specifically will happen.

I'm confident we'll eventually solve hallucinations, and I'm certain we won't solve them next quarter.

Alex Heath: Is it really true that your roadmap only plans six months ahead?

Nick Turley: Yes, with some exceptions. I genuinely hope people can understand how experience-dependent and peculiar it is to build products on a constantly shifting technological foundation — something no other type of company has to face.

For the vast majority of our features, we really can only plan a roadmap six months out. However, our enterprise business is a different situation. If a Fortune 500 company asks us when a certain compliance feature will launch, we have to be able to give a definitive answer.

We rarely have a high-confidence target beyond six months, because everything at the foundation level is changing at extreme speed.

ChatGPT's Product Form

Alex Heath: I have an anonymous question here from a former colleague of yours who asked me to put this to you: why hasn't ChatGPT's product form changed more dramatically by now?

Nick Turley: I've thought about this too. I think a lot of people know the story by now, but for those who don't, let me tell it again:

ChatGPT was originally built as a throwaway prototype to validate a much bigger product idea. We were trying to build what we called a "super assistant" — a flexible entity that could help you do anything.

ChatGPT was just our starting point for getting there. We simply wanted to use it to gather enough learning data and use cases so we could build the "real product."

But then things obviously went completely off the rails, because ChatGPT exploded and became massively successful — far beyond anything I or anyone else expected.

Natural language itself is incredibly powerful, and I think that will always be true. But whether it ultimately persists in the form of a chatbot? That's debatable. What I do believe is that letting users express themselves in the most natural way possible is the ultimate form of user experience. Because that's exactly how we humans are "programmed." As long as you're building technology for humans, you have to let people communicate with the product in the most natural way possible.

That said, I don't think "native natural language interface" equals "chat." An early attempt was Canvas. It lets you iteratively co-create a piece of work with AI — you're collaborating on something together, not going back and forth in conversation.

Combined with GPT-5's powerful front-end code generation capabilities, you can easily imagine a future where AI instantly generates different user interfaces depending on the use case — a product vision far grander than Canvas.

Imagine: when you're doing data analysis, it generates a spreadsheet for you; when you're planning a trip, it generates a small web app for you and your friends to collaborate on; countless product forms would naturally emerge.

All I can say is that given the technology available at the time, chat was simply the most suitable interface. A language-driven but visually richer experience is going to be incredibly, incredibly cool. I'm also somewhat puzzled that we're still using a chatbot, but we have enormous ambition for where the product can go, and I believe technology will make all of this possible.

Alex Heath: In the Google antitrust case, a strategic document from your team was disclosed that mentioned this "super assistant" goal and stated you wanted to build the "interface" through which people access the internet. That means you have to go beyond chat and even into web browsing — there have been media reports about this.

You yourself testified in the Google antitrust case that if Google were forced to divest Chrome, OpenAI might be interested in acquiring it. Are you developing your own browser now? Does OpenAI need to create its own web browser beyond ChatGPT?

Nick Turley: That statement was severely taken out of context, and I need to clarify — my full answer at the time was that if Chrome were put up for sale, many parties would consider acquiring it, and we would be among them.

At the product level, ChatGPT is becoming a new entry point to the internet. Many things you needed a browser for ten years ago, you can now do directly in ChatGPT, and it gives you the answer directly.

As it starts handling more time-consuming, complex tasks for you — like the trip planning or data analysis we mentioned earlier — things you might have needed to open three different pieces of software for in the past, you may handle directly in AI in the future.

So I don't think "ChatGPT will replace the browser" is a crazy idea. But exactly what form that takes, we'll have to wait and see. We're exploring all possibilities, and I fully agree with the argument that ChatGPT will increasingly have to take on roles that browsers play today.

Other Questions

Alex Heath: Alright, for our remaining few minutes, I've prepared some rapid-fire product strategy questions where I'd love to get your quick takes.

Sam has mentioned multiple times that "Sign in with ChatGPT" is strategically important in his view: letting users take their ChatGPT account and personalized settings across the entire web, making it an option like Google or Apple sign-in. How is that progressing?

Nick Turley: We're actively exploring it. I've learned the hard way that when you're building an ecosystem and trying to get other partners to build on it too, you have to go slow and get it as right as possible, because you don't get many chances to course-correct. We've been in discussions with many partners about moving this forward.

Alex Heath: There's a rumor that you're no longer working with Jony Ive on glasses or a phone. Is that true?

Nick Turley: I can't comment on our hardware roadmap at all.

Alex Heath: How is the Apple partnership progressing?

Nick Turley: It's going great. I'm incredibly excited about what we're building together. I see this as a long-term partnership, and I'm very much looking forward to bringing AI — hopefully our models specifically, but AI technology more broadly — into every corner of iOS.

Alex Heath: So you foresee the Apple partnership deepening further?

Nick Turley: That's outside my area of expertise, but from a purely product perspective, I see countless opportunities to deepen the collaboration.

Alex Heath: You announced a partnership with Mattel, the maker of Barbie dolls, to embed your models in their toys. Why do that?

Nick Turley: Because we're not just a product company, we're also a platform company. While we offer our own first-party products, we're also very happy to open up our underlying capabilities to everyone. This is a perfect example.

Despite how many areas we're involved in, we're unlikely to get into toy manufacturing. This is an excellent demonstration of how others can use our API to build their own businesses and products in industries we won't enter ourselves.

Alex Heath: When will ChatGPT achieve full multimodality?

Nick Turley: Our north star is that you can ultimately talk to it like a real person. That means, like our conversation right now, being able to interrupt each other, having lots of subtle body language cues.

I think it fundamentally comes down to making it easier for users to express themselves to AI, and for AI to express itself to users. That's the key to fully unlocking these models' intelligent potential.

Our vision is "any input, any output." But this is far more difficult than achieving pure technical capability, because you need the entire interaction to feel "natural."

You may have used our latest voice mode — I think it does a pretty good job already, but I still feel it hasn't fully passed the so-called Turing test. You can still tell when people are talking to AI. There are so many subtleties in human interaction that we're eager to crack.

Alex Heath: The four new model personalities you recently launched — what's the endgame of this personality experiment? Offering dozens or even unlimited personalities for users to choose from, or letting every user create a completely unique, one-of-a-kind ChatGPT personality?

Nick Turley: We're not entirely sure yet. We're very excited to learn from this launch of four personalities first — to observe whether user demand clusters around these categories or whether there's a massive, diverse long tail.

As for my current thinking, I believe users should be able to configure the model's personality themselves. We already have features like "custom instructions," and now these four personalities as additional starting configurations. In the future, you'll first select an initial configuration that resonates most with you, then further personalize it through daily interaction with the product or through explicit settings.

I think it's like making friends. You choose who to be friends with based on whether your personalities click in the present moment. But over time, the relationship continues to evolve and deepen.