How People Actually Use ChatGPT

0:00This podcast is supported by Google. Hey folks, Stephen Johnson here, co-founder of Notebook LM. As an author, I've always been obsessed with how software could help organize ideas and make connections. So we built Notebook LM as an AI-first tool for anyone trying to make sense of complex information. Upload your documents and Notebook LM instantly becomes your personal expert, uncovering insights and helping you brainstorm. Try it at notebooklm.google.com. Today on the AI Daily Brief, how people are actually using LLMs. Before that in the headlines, the latest agent protocol that could expand agent utility.

0:40The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:53All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, Gemini, Superintelligent, Agency.org, and Robots and Pencils. To get an ad-free version of the show, head on over to patreon.com slash aidailybrief. And if you are interested in sponsoring the show, shoot us a note at sponsors at aidailybrief.ai. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. One of the reasons that agent development has moved so fast this year is that the big labs and big agent builders have all decided that the prize on the other side of the rainbow is so big that it is not worth them competing to have to own all the infrastructure and standards along the way.

1:32What I mean by that is that you've seen this real rally around the standard kind of effect when it comes to MCP, which is, of course, the model context protocol and the way that agents can get access to sets of data in a standardized way, as well as A to A, which is an agent communication protocol. Well, now Google has announced a new protocol for agentic payments. On Tuesday, the company announced the Agent's Payment Protocol, or AP2, which they describe as an open protocol developed with leading payments and technology companies to securely initiate and transact agent-led payments across platforms.

2:06The protocol is supported by over 60 partners, including you name it, basically a who's who of finance and fintech, American Express, Ant International, Coinbase, Etsy, MasterCard, PayPal, Revolut, UnionPay, etc. So a broad consortium of payments and technology firms that spans the globe. The protocol is explicitly designed as an extension of MCP and A2A, establishing a standard for agents to interact with various commerce and payments platforms. Google noted that agentic shopping breaks three fundamental assumptions in e-commerce around authorization, authenticity, and accountability. A merchant can't know if an agent has valid authorization to spend or whether their actions accurately reflect user intent.

2:43Further, there's no way to establish accountability if a fraudulent or incorrect transaction takes place, which generally leaves the merchant at fault. AP2 seeks to solve those problems by standardizing the way that agents convey this information. Information is codified as mandates, which Google describes as tamper-proof, cryptographically signed digital contracts that serve as verifiable proof of a user's instructions. Mandates can be created by a human user or fully delegated to an agent. For example, a user might ask for white running shoes, which is codified as an intent mandate. Then the agent would return with a shopping cart containing a pair of shoes, which the user signs off on using a cart mandate.

3:19For delegated operation, the user might sign an intent mandate for an agent to buy concert tickets as soon as they go on sale. This functions as full pre-approval, allowing the agent to generate a cart mandate by itself based on certain criteria being met. This process establishes authorization, intent, and accountability to create an audit trail, essentially making agent purchases less able to be contested or reversed. Now, the expressiveness of the protocol allows users to do some really interesting things with complex transactions. Google gave the example of being able to authorize an agent to buy a particular jacket in a different color, and being willing to spend up to 20 % more than the price of a sold-out item.

3:54Users can also assign an overall budget for a more loosely defined goal. The example given was instructing an agent to book a round-trip flight in a hotel in Palm Springs for the first weekend of November with a total budget of$700. The agent can then book the various flights, hotel stays, and car rentals using the authorization. The protocol consequently paves the way for an agent-to-agent sales system. Google explained that a user could instruct their agent to purchase a new bicycle from a particular merchant, and then using that intent information, the merchant agent could then create a bundle including a helmet and a luggage rack at a discount and communicate that offer to the customer agent as a way of automating the upsell process.

4:29There is also a whole stable coin integration, which to the extent that you find interesting, maybe we'll go into in more depth. But ultimately, this is all about standardizing a set of communications that are necessary to make agents more functional in the real world. Google concluded, AP2 provides a trusted foundation to fuel a new era of AI-driven commerce. It establishes the core building blocks for secure transactions, creating clear opportunities for the industry, including networks, issuers, merchants, technology providers, and end users to innovate on adjacent areas like seamless agent authorization and decentralized identity.

4:59The protocol is completely open and available through GitHub, so anyone can start building features on top of it immediately. Now, as you might imagine, a lot of the discussion on Twitter slash X was around the crypto integration implications of this, whether this should all be on-chain, etc., etc. But I think from our standpoint over here on the AI show, if this also starts to get adopted, it could very quickly accelerate an entire set of agentic commerce use cases that are honestly such low-hanging fruit that I expect them to be ubiquitous by the end of next year. Staying in the Alphabet universe but moving to a very different area, YouTube has introduced a huge range of new AI features at the Made on YouTube event.

5:37The inevitable and expected VO3 integration is here, with YouTube allowing creators to use Google's video model to generate short clips for the platform. The toolbox also includes an AI editing feature, which allows creators to take a still image and apply a motion template. You can see their example of AI being used to transform a photo of an office worker into a karate master. Creators can also apply a style overlay to videos, including pop art or origami, as well as adding characters or props to videos using text prompts. Now, there are really kind of two totally separate things going on here.

6:07The first is quirky, interesting tools that could make for viral content formats. But some of the more professional tools, I think, have much bigger implications for the way that creators make YouTube content in general. The Edit with AI feature, for example, can also be used to generate a workable first draft from raw footage. Creators can upload a video from their camera roll, and Google's AI models will put together a short of the best moments and add music and voiceover. Basically, YouTube has put together a semi-automated way to churn out short-form content, which can then be refined as a creator sees fit.

6:36The same philosophy is being applied to podcast content, which has become an increasingly large focus for YouTube in recent years. In the coming months, podcasters will be able to use AI to generate clips in YouTube shorts from their longer-form content. The AI models will also be able to generate video content to go along with audio-only podcasts, which is a huge unlock. And honestly, as someone who is creating exactly this type of content, it is hard for me to describe how useful it is to have these tools natively in the distribution platform rather than just available as a third-party tool.

7:05That doesn't mean that there aren't opportunities for really unique and high-quality third-party experiences, but so much of the work of distribution is just that. It's work, it's clunky, it's hard. It involves an entirely different skill set than producing the content in the first place. The more that YouTube can normalize that and just make it effortless for creators, frankly, the better. Moving over to Microsoft and a follow-up to our story about them, including Anthropix models. The company has introduced an automatic model selector to GitHub Copilot and will now in fact favor Anthropix models.

7:37On auto mode, the selector will choose between Claude Sonnet 4, GPT-5, GPT-5 mini, and other models. However, Microsoft noted that for paid users, we currently plan to primarily rely on Claude Sonnet 4 as the model powering unit. Tom Warren of The Verge writes, it's a tacit admission from Microsoft that the software maker is favoring Anthropix AI models over OpenAI's latest GPT-5 models for coding and development. Sources familiar with Microsoft's developer plans tell me that the company has been instructing its own developers to use CloudSignit 4 in recent months. And as we discussed when that news first broke, as much as people want to make this about the Microsoft-OpenAI relationship fraying, I'm kind of willing to take them at face value that this is just about express model preference.

8:19Now, speaking of ChatGPT, two more quick stories on that front. The first is that Sam Altman and OpenAI have introduced an update to ChatGPT's personalization settings, Altman posted, we've updated ChatGPT's personalization page. Personality configuration, custom instructions, and memories are now all in one place. Now, while to many this might just represent a decent quality of life improvement, AI engineer Jacob Postle noticed an Easter egg hidden in the announcement. Over on the left-hand bar, he noticed a tab called Orders and commented Sam Altman quietly showing orders are coming to ChatGPT.

8:51Could this be a place to manage integrations for the sort of agentic shopping we were just talking about? I bet we'll find out soon. Lastly today, OpenAI has rolled out new restrictions in guardrails in an attempt to improve teen AI safety. ChatGPT will now refuse to engage in conversations around sexual topics or issues of self-harm with underage users. The chatbot will also be capable of contacting parents or police in certain circumstances. In addition, parents will now be able to set blackout hours where ChatGPT is unavailable for use by their children using new parental controls. The changes come in the wake of a rolling scandal about harm from teens using chatbots.

9:27Months of reporting and lawsuits culminated in a Senate hearing on Tuesday entitled Examining the Harm of AI Chatbots. Three parents of teenage victims were present to tell their stories, while none of the AI lab sent a representative. In announcing the new features, OpenAI shared their thinking about the underlying problem. It's notable that the announcement blog post wasn't from the company, but was from Altman himself. The company acknowledged that there is going to be some privacy compromising to implement these features, but that ultimately they think it's a worthy trade-off. Look, unfortunately, over the past few weeks, it has become abundantly clear just how challenging the issues of internet culture are and how far behind, frankly, we are on those issues.

10:03So I think it is very safe to say that this is the beginning of a much larger conversation around this aspect of artificial intelligence. For now, though, that's going to do it for today's AI Daily Reef Headlines edition. Next up, the main episode. Today's episode is brought to you by Superintelligent. Now, one thing that we are having a lot of conversations with folks about is the fact that for some of you, your fiscal year is coming to an end, and that means two things. One, it means planning and thinking about what you're going to do in the next year. And two, it means using up those last of budgets so you don't lose them.

10:35If you are an enterprise that happens to find yourself in that situation, Superintelligent would love to help on both fronts. We are moving increasingly towards an annual AI planning model where we map out how you can create an action map of your organization's agent opportunities that represents an executable backlog of AI and agent use cases that you can deliver on over the course of the next year. Additionally, for those end-of-year budgets, we have worked out deals with a number of partners where we can pre-lock in general implementation packages even before you figured out exactly what use cases are going to require them.

11:07If you'd like to learn more about Super Intelligence Agent Readiness Audits and this new end-of-fiscal-year plan, visit us at bsuper.ai, click Get Started, and make sure to use the word fiscal somewhere in the description. Shape the future of enterprise AI with Agency, A-G-N-T-C-Y. Now an open-source Linux foundation project, Agency is leading the way in establishing trusted identity and access management for the internet of agents, the collaboration layer that ensures AI agents can securely discover, connect, and work across any framework. With Agency, your organization gains open, standardized tools and seamless integration, including robust identity management to be able to identify, authenticate, and interact across any platform.

11:46Empowering you to deploy multi-agent systems with confidence, join industry leaders like Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and 75-plus supporting companies to set the standard for secure, scalable AI infrastructure. Is your enterprise ready for the future of agentic AI? Visit agency.org to explore use cases now. That's agntcy.org. Today's episode is brought to you by Robots and Pencils. When competitive advantage lasts mere moments, speed to value wins the AI race. While big consultancies bury progress under layers of process, Robots and Pencils builds impact at AI speed.

12:22They partner with clients to enhance human potential through AI, modernizing apps, strengthening data pipelines, and accelerating cloud transformation. With AWS-certified teams across US, Canada, Europe, and Latin America, clients get local expertise and global scale. And with a laser focus on real outcomes, Their solutions help organizers work smarter and serve customers better. They're your nimble, high-service alternative to big integrators. Turn your AI vision into value fast. Stay ahead with a partner built for progress. Partner with Robots and Pencils at robotsandpencils.com slash AI Daily Brief.

12:56Welcome back to the AI Daily Brief. Today, we are moving beyond the anecdotal to actually understand how people are using LLMs. Both Anthropic and OpenAI recently put out reports that help give us a picture. Now, Anthropics is their regularly updated economic index report, comes out every few months and is tracking the way not only that AI is being used, but what its potential impacts on the economy might be. But the OpenAI report is something altogether different. This is a big, distinct study that was conducted by OpenAI's economic research team in collaboration with the National Bureau of Economic Research and Harvard economist David Deming.

13:30It sampled 1.5 million conversations in a privacy-preserving manner to produce what OpenAI is calling the most comprehensive study of actual consumer use of AI ever released. Now, one thing to note, which distinguishes this from the enterprise study, is that this is just consumer AI use. No enterprise data was gathered. Yet at the same time, it does still look at economic and work-related use cases that appear in that broader pattern of consumer usage. So what did OpenAI learn? Let's start with the high level. One of the things that stands out to me as I dug into this is just how absurd the growth pattern of ChatGPT has been.

14:03This little line here that shows that it took under a year for ChatGPT to get to 100 million weekly active users, just an insane rate of growth already, looks basically flat compared to the growth that they've seen since then. There was just a massive inflection point jump, particularly when they released the reasoning models, where you can see over the last year or so, ChatGPT has just gone parabolic. It's also not just the users, they're also using it more. In fact, the message volume has actually increased at an even faster rate. But let's go back to what OpenAI thought was worth calling out as their most high-level insights.

14:36First of all, early on in ChatTBT's life, there had been something of a gender gap. Back in January 2024, among users with names that could be classified as either masculine or feminine, 37 % had typically feminine names. Now, this is obviously not a perfect measure of gender, but I think for painting with broad brushstrokes, it works. By July of this year, that gender gap is completely gone. The number of users now that have those typically feminine names is at 52%. So if you're rounding off, you can basically assume, I think, that ChatGPT is being equally used by men and women. ChatGPT is also not just a tool for rich countries.

15:11In fact, it's growing faster in lower-income countries. By May 2025, they wrote, ChatGPT adoption growth rates in the lowest-income countries were over four times those in the highest-income countries. To give a sense of just how different this is, this is a chart showing how long it took different technologies to have 90 % of their usage come from outside North America. The long blue line is the internet. From a classification of 1990 as year one on the internet, it took 23 full years for 90 % of the usage to come from outside North America. For ChatGPT, it took less than three years. And honestly, by year two, it was at like 80%.

15:47So the point is, this technology is for everyone around the world. Now, going back to usage patterns, OpenAI found that the most common usage was generally about getting everyday tasks done. They wrote, three quarters of conversations focus on practical guidance, seeking information in writing, with writing being the most common work task, while coding and self-expression remain niche activities. Which again, you can tell this is the broad general population, not our super concentrated productivity hackers that we have over here on the show. Looking at it a different way, OpenAI suggested that AI use could be split into three broad categories, asking, doing, and expressing.

16:22They wrote that about half of the messages, 49 % are in that asking category. They suggest that this means that people, quote, value ChatGPT most as an advisor rather than only for task completion. Doing represented 40 % of use and captured task-focused interactions like drafting text, planning, or programming, where the model generates functional outputs of some variety. Meanwhile, expressing was only 11 % of use and captured everything else, largely involving things like personal reflection, exploration, and play. But what about the division between work and non-work tasks? the study found that around 30 % of use was work-related, while 70 % was non-work-related.

16:59Breaking down the different tasks by category, the study found that practical guidance was the most common type of conversation, representing 28.8 % of usage. Seeking information, unsurprisingly, was up there as well, with 24.4 % of usage. Next was writing at 23.9%, but from there it was a big step down. Multimedia was at 7.3%, and self-expression, technical help, and other and unknown, each had a little over 5%. Part of what's interesting about the study is that you can also see how these patterns change over time. For example, in April of this year, there is a huge jump up in multimedia, which was of course right after the release of the GPT-40 ImageGen model and the Studio Gibbification of Everything meme, where just tons and tons of people were generating images who hadn't been doing that before.

17:43And interestingly, while that multimedia usage has leveled off, coming down from its peak of 12%, it's still up meaningfully from where it was. Before the Ghibli meme, it looks like multimedia usage was hovering a little under 4%. So to jump up to 7.3 % means that some number of those Ghibli memers actually stuck around for other versions of that use case. Another interesting insight looking longitudinally is that the very obvious use case of writing is falling away over time. Now, my instinct is that this doesn't represent people finding less use for ChatGPT's writing abilities. I think this is much more about them learning over time other things that they can use ChatGPT for.

18:21For example, seeking information is up from 18 % a year ago to 24.4 % this summer. That growth lines up with the introduction of deep research and just better search tools in general. One observation that I saw a couple people flagging was the gap between these results and that Harvard study that found therapy and companionship as the number one use case. Justin Angel writes, Weird that emotional support tasks like companionship and therapy are the top use case in the Harvard Business Review paper, but totally absent from this recent OpenAI report. Odd misclassification given that the HBR data examples are public.

18:52Feels like shenanigans. Why wouldn't OpenAI acknowledge emotional support use cases? Now, I've always had a little bit of skepticism about that Harvard study that found therapy and companionship right at the top. But it is interesting to me and potentially notable that this study explicitly decided not to have a category for therapy and instead lump that in somewhere with practical guidance or self-expression. Now, those are the big top line findings that OpenAI shared. But the folks on X dug much deeper into the actual paper and found a wealth of information. And one of the big interpretations was, boy, is this really good information if you are trying to figure out what to build.

19:26Greg Eisenberg, the host of the Startup Ideas podcast, posted the granular chart of each use case that was measured, arguing, each bar in this chart is a billion-dollar wedge if you build the right verticalized, trust-rich AI startup. So this breaks down those uses in a deeper way. For example, in Practical Guidance, they found that tutoring or teaching represented 10.2 % of all ChatGPT usage. Greg wrote, people want on-demand teachers more than almost anything else. A personal AI tutor that explains things your way, remembers your progress, and nudges you daily. How-to advice represented 8.5 % of use.

19:59With Greg commenting, this is consumer SaaS for micro niches. How to fix my resume. How to meal prep. How to set up my Shopify store. Every how-to is a wedge into a vertical AI agent. Lumping personal writing and editing together at 18%. Greg said this is demand for AI co-pilots inside every workflow tool, sales, legal, HR, PR. His big takeaway was, every bar is a behavior that people are already paying for elsewhere. Translate those behaviors into vertical AI agents and startups, the data doesn't lie. ARC Prize president Greg Kamrat had a similar take sharing the same chart and saying, if you're looking for a startup idea, here you go.

20:35And also famously referencing the Craigslist unbundling chart, which showed how individual categories on Craigslist in many cases turned into their own highly valuable startups. But the counter narrative was there as well. Jamie Forward summed up a lot of people's feelings when he responded, if you're looking for which startups not to build, here you go. The debate, of course, was around what ChatGPT is actually going to do. Michael Kove wrote, if people are using ChatGPT for those going directly head-to-head with OpenAI is going to produce churn. People that aren't aware about ChatGPT eventually will find your apps feature there.

21:05People already using it aren't going to pay for your app and ChatGPT. I'd focus on what they're not using ChatGPT for. And that's precisely why it's a hard thing to find product market fit, especially in the consumer app market. It's outside the scope of this particular show, but this is one of the most interesting questions for startups right now, is trying to make a bet on which opportunities the big labs are going to decide are core functionality of these LLMs versus things that they're comfortable letting other people do. We also don't really know yet what behavior patterns are going to look like, and whether people are going to be actually willing to pay for wrappers, and what features are going to make wrappers that people are willing to pay for.

21:43Olivia Moore of Andreessen Horowitz had a different set of interesting observations. One she noted was that work-related usage is getting less and less over time for people on regular non-enterprise accounts. In June of last year, around 47 % of usage was work-related, but now that's down to 27%. She also pointed out that in contrast to, for example, Claude, only 4.2 % of conversations in the study are in some way code-related. Looking at a chart of daily messages sent per weekly active user, SWIX also noted that there were some really significant inflections around the launch of the reasoning models and around the launch of deep research.

22:17So far, GPT-5 doesn't show a similar pop, but it's still very early and too hard to tell. Now, as I mentioned, this week we also got a new edition of Anthropic's Economic Index Report. This study uses a similar methodology of taking anonymized CLAWD chats and categorizing them into different use cases. One key difference is that Anthropic analyzes all usage, including through the API, so this is not just the consumer chatbot use. In fact, some of their most interesting observations are around how API usage differs from CLAWD.ai use. This is now the third edition of this report and is published every three months this year, meaning that we're starting to get an interesting picture of not just how CLAWD is being used, but how usage has changed over time.