OpenAI Introduces ChatGPT Ads to Fund Free Access to Top Models

OpenAI推出ChatGPT广告 以资助顶级模型免费使用 该公司认为纯付费商业模式是一种限制获取先进人工智能的精英化途径并予以拒绝，转而选择通过让底层模型对广告数据完全不可见来维护用户信任。 在本期节目中，主持人 A.J. Maine 与 OpenAI 的 Asad Awan 探讨了面向 ChatGPT 免费版和“gold”层级用户战略性推出广告支持模式的举措，该举措旨在践行让全球受众免费获取高性能 AI 提供资金支持的使命。为维护用户信任，OpenAI 实行了严格的优先级机制，将信任与隐私置于收入之上，确保对话数据与广告商完全隔离。视觉设计上对模型原生回答与独立广告横幅进行了明确区分，并配备了完善的用户控制功能，支持清除历史记录或退出个性化推荐。展望未来，该举措旨在通过自然语言提示简化小型企业的广告投放，并朝着值得信赖的对话式产品发现方向演进。 OpenAI 宣布在 ChatGPT 的免费版和 Gold 订阅层中测试广告，而 Plus、Pro 和 Enterprise 企业用户将完全免除广告，以此资金支持向超过 8 亿大众用户提供更高性能的模型和更充裕的使用额度。 (0:00) 系统实施广告与模型回答的严格视觉与架构分离，AI 无法感知展示的广告内容，对话数据由 OpenAI 内部匹配且绝不共享给广告主，健康与政治等敏感领域也将自动排除。 (4:01) 内部制定了“用户信任 > 用户价值 > 广告主价值 > 收入”的严格决策准则，将用户信任视为核心商业模式，杜绝牺牲隐私或引发用户不安的过度商业化尝试。 (8:53) OpenAI 设想未来借助 Agent 化的 AI 技术降低中小企业广告营销门槛，并通过更自然的对话式双向发现机制连接用户与利基市场需求。 (20:26)

OpenAI plans to introduce advertising to ChatGPT's free and gold tiers to fund computational costs and expand access to its top models without restricting usage limits. Guest Asaad Awon maintains that maintaining user trust through strict technical separation and privacy controls is essential to the platform's core business model. Subsidizing Broad Access Across User Tiers (0:00) Facing a user base exceeding 800 million active consumers, OpenAI chose to adopt an ad-supported framework rather than degrading model quality or placing severe usage caps on non-paying users. Awon explains that the company rejected what some call the "high road" of remaining entirely ad-free because doing so would create an elitist product accessible only to paying customers, running counter to OpenAI's stated mission of delivering AGI to all of humanity. Under this multi-tiered architecture, advertising is strictly confined to free and "gold" accounts, while Plus, Pro, and Enterprise tiers will remain completely ad-free. Awon argues that enterprise contracts and consumer subscriptions fund advanced features, whereas advertising revenue allows OpenAI to offer generous usage allowances and high-performance models to the entry-level tier, avoiding the frustrating experience of hard question caps. Technical and Visual Separation of Promotional Content (4:01) To preserve objectivity, OpenAI establishes a structural wall between the underlying AI models and the advertising system. ChatGPT's core models are neither trained on advertising data nor given awareness of the ad displayed on screen; if asked what an ad says, the assistant will state that it does not know. Advertisements are restricted to distinct visual bottom banners to prevent users from mistaking sponsored messaging for organic responses. Interactions between the model and the advertisement occur only at the user's explicit request. By clicking an "ask ChatGPT about this ad" button, the user can manually pass the promotional context into the prompt, operating identically to an external web link pasted into the chat interface. Awon rejects the prospect that commercial pressures will erode this boundary over the coming decade, arguing that the company's value relies entirely on remaining a trusted personal assistant. Governance Rubric, Sensitive Contexts, and Privacy Controls (8:53) OpenAI evaluates advertising decisions against a four-tier hierarchical rubric: User Trust outranks User Value, which outranks Advertiser Value, which outranks Revenue. Awon illustrates this principle by stating that an ad causing a user to suspect the application is secretly recording them represents a failure of trust, summarizing the company's rule as "is creepy okay if it is good? It's not." Consequently, team incentives are directed away from optimizing for time spent on the platform. User privacy is protected through internal matching systems that prevent user chat logs from being shared directly with third-party advertisers. Automated classification models flag and exclude sensitive verticals—including health, politics, and violence—from triggering promotions. For non-sensitive queries, users retain controls to inspect which data informs ad targeting, clear their ad history, opt out of using past conversations for personalization, or disable personalization entirely. Small Business Tooling and Conversational Discovery Forecasts (20:26) Awon argues that current digital advertising systems disadvantage small and medium-sized businesses by requiring specialized performance marketing teams to navigate complex bidding auctions. Using an example of an e-commerce shoe startup, he proposes that generative AI can level the playing field by acting as an autonomous marketing agent, executing campaigns across target regions based on straightforward natural language prompts. Looking ten years ahead, Awon forecasts that digital advertising will transition from static banner placements to agentic, conversational discovery. In this model, an AI assistant with an understanding of individual user preferences could naturally introduce niche products—such as specialized vegan ramen or footwear brands like Allbirds—creating a targeted marketplace that Awon claims will feel useful rather than disruptive. OpenAI acknowledges widespread consumer skepticism toward digital advertising after two decades of intrusive tracking across the web. While the company frames its conservative rollout, strict policy rubrics, and privacy controls as safeguards, the long-term challenge remains balancing commercial monetization against its commitment never to compromise user trust.

OpenAI代表Asaad Awon表示，为了践行让通用人工智能（AGI）普惠全人类的使命并维持逾8亿活跃用户的庞大服务规模，ChatGPT将在免费及“gold”层级引入广告，并通过确立“用户信任高于收入”的决策等级与技术隔离机制，力保模型回答的客观性与用户隐私。 免费层级资金支撑与多层级分配策略 (0:00) 面对超过8亿的庞大消费者基数，Awon指出OpenAI面临战略分水岭：若完全不接纳广告，公司将不得不严重限制免费用户的使用额度，或仅能向非付费群体提供性能较弱的模型，从而导致用户遭遇“提问五次即中断”的不良体验。为了避免形成仅有付费者能享用最强AI的“精英主义”格局，OpenAI选择通过广告模式支持免费层级，以此向最广泛的大众提供顶级模型与充足的使用配额。 在具体的层级划分上，广告将严格限制在免费版和“gold”层级展示，而Pro、Plus以及Enterprise企业级订阅用户将维持完全无广告的纯净环境。Awon强调，依靠企业级与高级订阅支持先进功能研发、同时借助广告收入维持免费端高可用性的多元商业架构，是兼顾普惠接入与可持续运营的最可靠路径。 模型回答与广告展示的严格隔离 (4:01) 为确保广告投放不损害用户体验，OpenAI设立了核心设计原则，即模型生成的答案必须在结构和视觉上与广告绝对分离。Awon解释称，底层的AI模型不会接受广告数据训练，且对广告位展示的内容完全不知情；例如即便用户在探讨健康奶昔时页面底部出现了破壁机广告，若用户追问模型“这个广告在说什么”，AI也只会如实回答“我不知道”。 在视觉设计上，广告目前以明确标记的底部横幅呈现，防止用户将推广内容与模型的原生回答相混淆。用户仍掌握交互控制权，仅当其主动点击“向ChatGPT询问此广告”按钮时，广告上下文才会像用户粘贴的外部网页链接一样输入模型。针对商业利益未来是否会侵蚀这一边界的质疑，Awon回应称OpenAI的根本商业逻辑依托于用户信任，任何跨越界限的举动在商业上都将适得其反。 信任优先的决策准则与隐私红线 (8:53) 在内部经过多轮跨部门辩论后，OpenAI确立了严格的层级化决策准则：用户信任（User Trust）高于用户价值（User Value），用户价值高于广告主价值（Advertiser Value），广告主价值最终高于收入（Revenue）。Awon举例指出，即使系统展示的精准广告促成了用户购买并产生了“用户价值”，但若导致用户怀疑“应用是否在偷听麦克风”，便构成了对“用户信任”的灾难性破坏，因此公司坚决禁止引发恐慌感的广告匹配。 在数据安全与匹配机制层面，用户的对话数据绝不会直接共享给广告主，所有匹配逻辑均在OpenAI内部完成。同时，系统利用高级内容分类模型自动识别并排除健康、政治和暴力等敏感领域，防止在不恰当的对话语境中出现商业推广。此外，平台将赋予用户充分的控制权，允许其查看广告关联数据、清除广告历史、禁止使用过往对话进行广告定位，或直接完全关闭个性化推送。 赋能中小企业与未来代理化广告愿景 (20:26) Awon指出，现存的数字广告竞价拍卖体系高度依赖复杂的数据分析，往往让资源充沛的大企业占据优势，而缺乏专门营销团队的中小企业（SMB）则承担极高风险。OpenAI希望通过AI将广告投放流程简化为简单的自然语言指令（例如直接输入“目标是在中西部多卖出这批鞋子”），由AI代理自动处理竞价、测试和预算分配，使中小企业无需组建营销团队即可实现高效触达。 通过援引鞋类品牌Allbirds早期精准定位硅谷科技设计师并取得成功的案例，以及个人消费特种纯素即食拉面的体会，Awon阐明了智能分发帮助小众产品寻找忠实受众的潜力。展望未来十年，Awon预测广告将向深度对话式与代理化体验（Agentic Ads）演进，AI助手将在充分理解用户偏好且可控的前提下，以自然、真实的方式推荐契合需求的利基商品，构筑双向有益的发现市场。 Awon承认数字广告行业历经约20年发展后，公众对广告侵蚀体验的防备心理完全合理，OpenAI必须以极度保守的投放频率、严格的敏感内容过滤与透明控制来重新建立信任。然而，在未来十年的商业化推进中，这种将广告与模型回答完全剥离的硬性边界能否抵挡持续增长的财务扩张压力，以及对话式代理广告模式能否真正消除用户的侵入感，仍有待市场与时间的检验。

[0:00] Asaad Awon: bringing AGI to all of humanity and to benefit all of humanity. [1:35] Asaad Awon: empty calorie time spent on the platform. [4:01] ChatGPT: I actually don't know. [4:01] Asaad Awon: ask ChatGPT about this ad. [5:43] Asaad Awon: You can't drift when the incentive is set up to be the best at this, [7:28] Asaad Awon: ask five questions and then it just stops. [8:53] Asaad Awon: Was this app listening to me and is the mic on? [8:53] Asaad Awon: Is creepy okay if it is good? It's not. [11:07] Asaad Awon: If it is not [useful], we'd rather not show you anything. [14:16] Asaad Awon: the answers are separate from the model. [15:54] Asaad Awon: one right ad [17:34] A.J. Maine: No ads. No ads. No ads. [17:34] Asaad Awon: third inning [17:34] Asaad Awon: elitist view [20:26] Asaad Awon: My goal is to sell these shoes more in the Midwest and go,

[0:00] Hello, I'm A.J. Maine and this is the [0:02] Open AI podcast. Today we're talking to [0:04] Asaad Awon about ads in Chat GPT, how [0:08] they'll look, who will see them, and how [0:09] the company will preserve user trust. [0:12] Ads are shown to people who are on the [0:14] free and the gold tier. For pro and plus [0:17] and for enterprise, there are no ads. Is [0:19] creepy okay? If it is good, it's not. We [0:22] are in the business of trust. So I think [0:23] if we have to say what is our core [0:24] business, it's like to win users' trust. [0:29] So from a consumer point of view, why [0:32] ads, why now? It goes [snorts] back to [0:34] our mission, which is bringing AGI to [0:36] all of humanity and to benefit all of [0:39] humanity. [0:40] So when you have a consumer product [0:41] which like, you know, 800 million plus [0:43] people were using this, then how do you [0:45] take the best version of that product to [0:47] everyone? [0:49] And and ads is one of the most proven [0:51] models to be able to do that for [0:52] consumer products. [0:54] And I think the other part of that [0:56] mission is how do you benefit all of [0:58] humanity, which is like you want to take [0:59] the best model, you want to give the [1:00] highest limits, usage limits to people. [1:03] You want to [1:04] for the ads to be actually helpful both [1:06] to the users and the businesses as well. [1:08] So so I think it's a very natural fit [1:11] for a company whose ambition is actually [1:13] to take the best AI to all of all of [1:16] humanity. It's a very interesting [1:17] decision because on one hand, you could [1:20] say, "Hey, we're going to take what we [1:22] perceive as the high road and say we're [1:23] not going to do ads, but also we're not [1:25] going to give a really good amount of [1:26] usage for it and limit that and sort of [1:28] maybe use not the most capable models, [1:31] but sort of, you know, say take that [1:32] approach versus embracing it." Yeah. [1:35] Yeah, I think like if if if the goal is [1:37] to truly democratize access, I think ads [1:39] is a good model. I think maybe what is [1:41] hidden in that statement is [1:43] can ads be bad? And and the reality is [1:46] how do we think about the principles of [1:47] ads? How do we actually set a really [1:50] high bar for what ads should be on this [1:52] platform? How do we make them actually [1:54] useful? So when we were starting off, we [1:56] we thought like, "Hey, what would be the [1:58] core principles that we would announce [1:59] to the world that we would be proud of, [2:01] that we would stand behind, and as a [2:03] result, create a really great product?" [2:05] So, so, just to give the example of the [2:07] principles, it's like, number one, the [2:09] answers need to be independent from the [2:10] ads, both visually, but also in how the [2:13] models are trained and how the system [2:15] works, [2:16] so that you can always trust the answer. [2:18] Like, the whole product ChatGPT product [2:20] is based on trust, so actually it needs [2:22] to feed into that. [2:23] >> [snorts] [2:23] >> The second is, your conversations are [2:25] private. If you have a sensitive [2:27] conversation, that will never have ads [2:28] in it. [2:29] And and the conversations are never [2:31] shared with advertisers. So, while we do [2:32] the matching between the best ad that [2:34] can be a useful thing in a conversation, [2:37] the advertisers don't get to see that. [2:38] We do that matching internally. And [2:40] then, of course, like, you know, as as [2:42] you introduce ads, a big question is, [2:44] how did you know about this data? Like, [2:46] that's the difference between the user [2:47] trust and just like doing something [2:48] which is relevant to the user. [2:51] And our goal was, how do we make [2:53] something which users can transparently [2:55] understand, how can they control, and [2:57] and we can go into that, because I do [2:58] think there is a high bar to set over [3:00] there, because you could have some [3:02] transparency, some control, which most [3:04] products have, but what would be a [3:05] really good version of that is something [3:07] that we've been thinking about. And [3:08] finally, once you add ads, I think you [3:10] have to set the incentives for the [3:12] teams, for the company, in a way that [3:14] actually continue to focus on user [3:16] value. So, you don't want to just like [3:18] get empty calorie time spent on the [3:20] platform. You want to build a very [3:22] useful product, and automatically, one [3:25] good ad is good enough, actually. So, we [3:27] don't optimize for time spent [3:29] on on the platform and focus on the user [3:31] value. So, these are the principles. So, [3:32] I think like, [3:33] so, connecting this back to your [3:34] question, is like, how do you why why [3:37] should we add ads and how to do it? I [3:39] think a part of that is, take it to all [3:41] of humanity, and that's the best [3:42] business model to do that, but [3:44] prevent all the negative things that can [3:46] happen if you're not doing it [3:47] thoughtfully, and I think being up front [3:49] with our principles, being very clear [3:50] with that, [3:52] Um, is how we're starting and then how [3:53] we will test, how we will improve, and [3:56] how we become kind of a learning [3:57] organization with respect to this. I [3:59] think that's that's how it will go. So, [4:01] you said basically there's going to be a [4:02] separation. So, if I'm talking to [4:05] Chat GPT about like, "Hey, I want to [4:07] start drinking smoothies and stuff." [4:09] It's not going to all of a sudden blurt [4:10] out like, "Well, here's a blender you [4:11] should buy." Absolutely. Yeah, I think [4:13] it's a both in terms of what the model [4:15] knows, the model doesn't know, whether [4:16] an ad is there or not. If you ask it [4:17] like, "Hey, what is this ad saying?" [4:18] It'll say, "I actually don't know." Uh, [4:21] but you can actually press some buttons [4:22] and add it to the model if you want to [4:24] ask the question. So, it's totally [4:24] totally totally separate. Totally [4:26] whatever is being displayed in the ad [4:27] space, the model has no idea that's in [4:29] it. That's right. And I think in in and [4:31] both visually as well, so that the user [4:33] can very quickly say, "Hey, this is the [4:34] answer that I got from the model." And [4:36] then there's a [4:37] bottom uh, banner which has ad in it, [4:40] uh, which is very clearly distinct. So, [4:42] visually also you don't confuse that. Of [4:43] course, we will learn how that [4:44] experience evolves, but the goal is to [4:47] both keep the system, the models very [4:49] very separate, and ad is kind of [4:51] downstream of. Okay, yeah, that's I [4:52] think that's a very important [4:53] distinction cuz I think some people have [4:55] sort of kind of tried to sort of spin [4:57] that there's some sort of collusion [4:59] between the ad part and the model part, [5:01] but you're saying the model is [5:02] completely separate. [5:02] >> Yeah. So, that's interesting. So, as I'm [5:04] having a conversation, something might [5:06] come up and I can then click and say, [5:07] "Okay, tell the model, 'Hey, I saw [5:10] this.'" Then it knows what's going on. [5:11] >> That's right. Yeah, you have to go like [5:12] in the in the first experience you like [5:14] specifically have to press a button is [5:15] like ask Chat GPT about this ad. And [5:18] that would be as if you took a link from [5:19] the internet and asked a question about [5:21] it. So, it's almost the same. You don't [5:22] want to make that experience harder. But [5:24] but if you say, "Hey, what is this ad [5:26] talking about?" It'll say, "I don't [5:27] know." It's easy to start now and say, [5:29] "Oh, yeah, we're going to do the great [5:30] thing. We'll do it right." But 10 years [5:32] later when there's an entire division in [5:33] charge of ad revenue, you might say [5:35] like, "Well, do we need the wall between [5:37] the model and the ads?" Yeah. Yeah, I [5:39] think like maybe there there there's [5:41] multiple mul- multiple angles to this. [5:43] So, one, we are in the business of [5:44] trust. So, I think if we have to say, [5:46] "What is our core business is like to [5:47] win users' trust and give amazing [5:50] answers to the question that they're [5:51] asking. That's in the in the consumer [5:53] product side. And of course on the [5:54] enterprise side, trust is everything, [5:56] which is like you're entrusting us with [5:57] your most important data. We need to of [6:00] course make maintain that. So, so [6:02] because I think like that the ambition [6:04] and the vision is so expansive, I think [6:06] trust is the central point of it. We [6:08] want to have devices which are helpful [6:09] for you. [6:10] If we truly want to be your best [6:12] personal assistant, then you need to be [6:13] able to share your most important [6:17] information, but know that it will be [6:22] dealt in a way which is like how you [6:23] would treat it yourself. So, I think our [6:26] business model is trust. This is very [6:28] different than many other scenarios. If [6:30] you're just doing like kind of a [6:31] transactional stuff, like again, you you [6:33] gave the question and answer came back [6:34] and that's the end of it. Like a search [6:36] query, I think that's that's okay, but [6:39] it is not a long-term relationship. I [6:41] think if you think of content discovery, [6:43] certainly I I mean like it it is just [6:46] pushing things [6:47] and and trust is not a core component of [6:50] that. For us, I think the whole product, [6:52] whether it's enterprise, whether it's [6:54] consumer, whether it's anything devices [6:56] in the future, they're all centered [6:58] around trust [6:59] trust and [7:01] so so so [7:03] for us it's kind of imperative. For [7:05] others, it could be optional. And I [7:07] think different companies are known for [7:08] different things and we do want to be [7:10] known for trust. And so so I think like [7:13] connecting this to the question is [7:15] will be drift. I think you can't drift [7:17] when the incentive is set up to be the [7:19] best at this and this is the goal [7:22] that we want to achieve. Everything else [7:24] is is there to support that vision, but [7:26] the over principle is trust. [7:28] OpenAI has a very huge number of people [7:30] using the free tier. And then there's [7:32] also paid subscribers and people who do [7:34] that pro users, whatever. How are ads [7:37] going to play out across the platform? [7:38] Yeah, so ads are shown to people who are [7:40] on the free and the go to you. Mhm. Um [7:43] and for pro and plus and for enterprise, [7:46] there are no ads. [7:47] >> Mhm. Um and and I think like that's [7:49] that's an important thing like the [7:50] context in which which the company [7:52] operates is actually like multiple [7:54] missions which come all together to [7:55] bring bring AI to everyone, which is [7:58] when enterprise use it, that's a very [7:59] specific contact. There is no ads over [8:01] there and there is a specific business [8:02] model around that which is very [8:04] powerful. Uh for subscribers who want [8:06] like, you know, the best like, you know, [8:07] highest limits and very advanced [8:09] features. Mhm. [8:10] >> that also works. But for a lot of [8:12] people, a lot of consumers, the best way [8:14] to do that is to have um have high [8:16] limits and free usage uh and then add [8:20] ads which are actually useful. Yeah. [8:22] I've heard people talk about, you know, [8:24] part of the goal of this is to avoid [8:27] making the free tier just like the most [8:29] limited thing available. Absolutely. I [8:31] think like that that is that is the that [8:33] is the most I think frustrating things [8:34] for a lot of real users. Just like you [8:36] ask five questions and then it just [8:38] stops in in other businesses, right? [8:41] We we I think like we want to grow that [8:42] a lot more and I think it fits with our [8:44] overall goal. Um that uh higher usage [8:48] limits is better. Mhm. Uh and how do we [8:50] fund that and be practical about it? So, [8:53] going a little bit behind the scenes, [8:54] how are these decisions made? Like who's [8:56] in the room talking about this? Yeah, I [8:57] think like this is a this is a good op- [8:59] opportunity also to talk a little bit [9:01] about like overall like there is a [9:02] company culture and I think different [9:04] companies have different cultures which [9:05] results in different products and our [9:07] our our company has like this DNA of our [9:10] research team, right? [9:11] Uh so so we have much more I think like [9:14] rigorous debates, rigorous understanding [9:16] of how should we make these principles, [9:18] how does incentives work, how does the [9:20] the model of this going to work in a way [9:22] that it doesn't get corrupted later on. [9:24] So, we have had both a lot of debates on [9:26] that which actually resulted in these [9:28] principles, which resulted in this [9:29] rubric with like hundreds of round [9:31] tables with like folks around the [9:33] company on different areas, not just [9:35] like working on ads, but everyone on [9:37] every different part of the company uh [9:39] giving feedback to create these [9:41] principles. Then we can convert those to [9:43] a very simple rubric. I think like the [9:45] rubric is user trust is the most [9:47] important thing. User trust more than [9:49] user value, which is then more important [9:51] than advertiser value, which is more [9:52] important than revenue. And I think [9:54] while this seems very straightforward, [9:56] it's actually a very very very, I think, [10:00] in-depth decision. So we can go into a [10:02] little bit just like user trust more [10:03] than user value. A good example of that [10:06] is [10:07] if I showed you a really good ad, but [10:10] and you liked it, you clicked on it, you [10:12] bought something. But later on you ask [10:13] the question, was this app listening to [10:15] me and is the mic on? That's not user [10:18] trust. You probably did provide some [10:19] value. So for us, our goal is like we we [10:22] we cannot have that. The users need to [10:24] believe and understand and control [10:26] what's happening. So that's just one [10:27] example, but I think once you set that [10:29] right rubric up, then even bottom up, [10:32] the team thinks like that. [10:34] But of course like you know as we have [10:35] different decisions at different level, [10:36] I think we have a pretty rigorous [10:38] process on how we discuss privacy within [10:40] the company, how we discuss safety [10:42] within the company, and there are very [10:43] very clear forums for that. And then of [10:46] course like as we as we make decisions [10:47] at leadership level, we one go back to [10:50] the simple rubric because although this [10:51] rubric is simple, it's actually pretty [10:53] pretty in-depth and actually is very [10:55] discriminating if you think about these [10:56] kind of questions. It's like, should the [10:58] ad be so good but the users don't know [11:00] where this data came from? Is creepy [11:02] okay if it is good? It's not. So I think [11:06] maybe it follows from that. [11:07] What am I going to see on my end? What [11:09] kind of controls do I have? How's [11:10] personalization going to work? So I [11:13] think like a big part of actually [11:14] delivering really good ads is allow [11:17] personalization so that is so when I say [11:20] I want to [11:21] go on a trip to Yosemite and then that [11:23] shows me camping gear because that's [11:24] what I'd like to do. [11:26] But of course I think the flip side of [11:27] that, how do you gain user trust is like [11:29] how did you know about this? How did you [11:31] learn about this? So one is transparency [11:33] aspect which is like you can see what is [11:34] the data that we have on you which is [11:36] being used for ads. The second is the [11:38] controls [11:40] which which let you see say which part [11:43] of the data from your past chats can be [11:45] used. Of course, sensitive chats, those [11:47] are never used. [11:49] And but you can clear your data which [11:52] actually nobody else does which is kind [11:53] of a crazy concept is like you can clear [11:55] your data so we don't know and we won't [11:57] use that. You could [11:59] say don't use my past chats if that's [12:01] what you you care about or you could say [12:03] turn off personalization fully. [12:06] Of course, there there is the other [12:07] extreme which is like I don't want ads. [12:09] That's a form of control and that's [12:11] where I think upgrading to the pro or [12:12] plus version to completely stop ads is [12:14] also there. So, I think all the way from [12:16] the spectrum of like I really care about [12:17] this. I don't think this is the right [12:19] business model like pro and plus is the [12:21] right business model [12:23] for like hey, I I don't know what we [12:26] were talking about yesterday. I'll just [12:27] clear my history. Great, do that. Or [12:31] it's like hey, I'm more comfortable with [12:33] like you know, clicks on the ads being [12:35] used but not my past conversations. You [12:37] could actually do that as well. Of [12:39] course, people will learn hopefully [12:41] experience like how it improve their [12:43] experience. We have a very high bar and [12:44] how we use these things but in the end [12:46] the users need to know and be able to to [12:49] control that. [12:50] What will be you know, the the kind of [12:52] the expectation for how many ads I'm [12:54] going to see or how often these would [12:56] come up? [12:57] Maybe the Uber principle still goes back [12:59] to in that context is there a good ad to [13:03] show which is useful. If it is not, we'd [13:05] rather not show you anything. In fact, [13:07] like you know, as we roll out this test, [13:09] you'll see that there'll be very few ads [13:11] because like you know, we want to be [13:12] both conservative and we want to learn [13:14] how to [13:15] where to insert those. But but the the [13:18] principle is little bit more around is [13:20] it useful? Is it helpful? Does it add [13:24] to to what the user is doing? [13:26] And can we actually show a really good [13:28] product as well. Um so so keep the [13:31] quality of the the content very high as [13:33] well. Keep the quality of the ad very [13:34] high as well. Keep the relevance really [13:35] high. If we can't find a good match, [13:37] it's fine. You don't need to show an ad. [13:39] You mentioned that sensitive [13:41] conversations. How do you know what [13:42] something is sensitive or not? So that's [13:44] actually one of one of the big strengths [13:47] of open AI is like both for our organic [13:49] work and a lot of research in the [13:51] company has gone into defining [13:54] what's sensitive like it is health, [13:55] politics, like violence, like many [13:57] different kind of verticals. [13:59] Um very very in-depth uh definitions of [14:01] that. And then of course using some of [14:03] the best models to actually predict and [14:05] understand the conversation and saying [14:07] marking it as sensitive or not. I think [14:10] like I've actually never seen such high [14:12] precision in any product so far in my [14:14] career what we have been able to build [14:16] over here by taking in those policies. [14:18] There's a team which works on defining [14:20] those policies very very rigorously [14:23] um [14:23] uh and then actually uh also um [14:26] sharing them with internal external [14:28] partners for review. And then of course [14:30] the enforcement that comes from the [14:32] prediction system that actually says, [14:33] "Hey, like this is matching this policy, [14:35] so don't do that." We've talked a little [14:37] bit about this, but I'd like to touch [14:38] back on this again, design. So where is [14:41] that going to head? What are they going [14:42] to look like? Um as we were designing [14:44] this product, I think like of course we [14:46] set up a very clear principle that the [14:48] that um [14:49] uh the answers are separate from the [14:50] model. And then the question was like [14:52] how does that actually look in the [14:53] product? Um and on that spectrum on one [14:56] side is like how do you make it look [14:57] native so that it's not jarring? [14:59] Um and on the other side there is a [15:01] question which is like, "Hey, how can it [15:03] be very clearly separated out?" Um and [15:06] and I think you can debate with on both [15:08] of them and there is values in both of [15:09] them and we start we we wanted to kind [15:12] of set up the experiment in a way which [15:14] is um that we can learn as we go. So we [15:17] take the conservative option and still [15:20] keep that principle in mind. And as we [15:22] learn through uh building the product to [15:24] getting the data, [15:26] evolve it. But But the idea still is how [15:29] can we maintain that principle of the [15:30] answers being very clearly separated [15:32] from the [15:34] from the ads [15:36] and having a very clear [15:37] understandability and visual [15:38] distinction. I do think we will evolve [15:40] the formats, and I think they will get [15:42] even more useful and better over time, [15:44] but that principle is constant, and [15:46] within the options that we had, we [15:48] started with something which is which is [15:51] clearly separated out and conservative. [15:54] So, you explained kind of a technical [15:55] level how there's a separation, how the [15:57] model doesn't see it, but also for [15:58] guardrails and stuff. And I think you [15:59] mentioned this before, but if I'm [16:01] talking about, you know, you know, [16:02] saying like, "Hey, I'm afraid of this [16:03] trip." And it's like, "Well, hey, how [16:04] about some life insurance?" You know, [16:07] that's not going to happen, but how do [16:08] you guys put in guardrails, and how do [16:10] you decide what's appropriate ads and [16:12] what's not? So, so maybe there's two [16:14] questions in there, like what's [16:15] appropriate ads or not and which context [16:17] is a reasonable one. [16:19] And the second is what are the controls [16:21] in place so that, like, you know, over [16:23] time this doesn't doesn't dissolve. So, [16:25] I think maybe a part of announcing our [16:27] principles and being very clear [16:29] internally for a rubric was to actually [16:31] set that up in the first place. Then [16:33] automatically a lot of the governance [16:35] within the company, how we make [16:36] decisions follows from that [16:38] onwards. So, it's like, "Hey, I want to [16:40] make this change to the product." Do [16:42] this fit with this principle? Do this [16:43] fit with this rubric that we have [16:44] already set up? That's the first pass. I [16:47] think the sensitive context is is [16:49] something that we take very seriously as [16:50] well. Very simple things like, you know, [16:52] conversations around health [16:54] or politics or other contexts where [16:56] where the ads don't fit, and that data [16:59] is not going to be used for making ads, [17:02] like, even even matching ads. You just [17:05] filter it out. So, the first layer is [17:07] really see does an ad belong here. If it [17:10] does and it can be helpful and additive, [17:12] then add it. [17:13] I think like this goes back to the [17:15] principle is like actually you don't [17:16] ruin neither for users nor for [17:19] businesses by showing many ads because [17:21] you don't want advertisers to pay [17:23] randomly for impressions. You don't want [17:25] users to see too many ads. You want to [17:26] share the one right ad and being one of [17:29] like, you know, the [17:30] best AI company, I think that's [17:32] hopefully something we'll do really [17:33] well. Every time we do an episode, we [17:35] get a few people who go in the comments [17:36] are like, "No ads. No ads. No ads." Now [17:38] is your chance to talk to those people [17:40] directly. [17:41] Yeah, I think I think like in in some [17:43] sense the when when people say no ads, I [17:46] I feel like there is a perception and [17:48] it's not wrong that [17:50] uh because I think like maybe how how [17:52] the industry has evolved that there is [17:54] some suspicion around how this works. [17:56] So, I do think it is kind of incumbent [17:57] on us to come up with better principles, [18:00] get a better clarity, better rules on [18:03] how we're going to do this. I think [18:04] there is like we like again this whole [18:06] ad industry, if you think about the [18:08] online ad industry is like maybe 20 [18:10] years old compared to many other [18:11] industries which are hundreds of years [18:13] old. So, I think maybe we are in the [18:13] third inning of this where we are [18:15] saying, "Okay, we have learned from all [18:17] of these questions and problems that [18:18] people have." I think when people say no [18:20] ads, I do believe that they have valid [18:22] questions and concerns around privacy. [18:23] The it's on us to do a really good job [18:26] to earn their trust uh through better [18:28] transparency, through better control, [18:30] through uh [18:31] through building that is also [18:32] delightful. Um [18:34] uh I think there still be skeptics and [18:36] then I think we have a way to upgrade [18:38] because I think that's a valid choice as [18:40] well. Uh but uh [18:42] but enabling really good ads with good [18:44] principles, I think it's possible. I [18:46] think a big part of it is [18:48] having really strong AI to power these [18:50] ads also so that they are actually [18:52] useful and then as a result bring this [18:55] product to so many people without with [18:56] with higher limits. [18:58] Some of your competitors have been [18:59] having a little bit of fun at the idea [19:01] of ads. Yeah, I I think like um [19:04] uh [19:05] different different companies have [19:06] different missions. Our mission is to [19:07] take AI to all of humanity and and um of [19:10] course we have different contexts. So, [19:11] we have the enterprise business, we have [19:13] uh you know, our subscription business, [19:15] and we have a very very huge consumer [19:16] base which is using our product. So, I [19:18] think within that context, we need to [19:20] serve each one of them. We will have a [19:21] really robust enterprise business, and [19:23] there will be no ads over there. And [19:25] then, we'll have a very robust consumer [19:26] business, and ads will help us grow uh [19:29] within that. Uh so, I think if if that's [19:31] not your mission, maybe it doesn't make [19:32] sense, but our mission is to to actually [19:35] build in all of these contacts, and we [19:36] believe they're all actually related uh [19:38] in in how we build the best AI, and then [19:40] actually take it to everyone. And I [19:42] think the good part is that we have [19:43] different verticals in the business [19:45] line. So, it's not just an ads company. [19:47] There are some companies which are [19:48] purely just ads companies, and there the [19:50] incentives are actually different. [19:52] >> Yeah. But, I think we have a much more [19:53] holistic view on on this. And also, when [19:55] you're not serving hundreds of millions [19:56] of free users, it's easier to sort of [19:58] say, "Hey, we don't have to do this." I [20:00] don't think it's like a vision vision [20:02] which is which is set in abstract. This [20:03] is truly a vision which is like, how [20:05] does AI actually help people? And if if [20:09] there is like this elitist view that [20:11] some people get to use it, and some [20:12] don't get to use it based on who can [20:14] pay, I think like that itself is a [20:16] pretty big fork in the road in terms of [20:18] how AI can be valuable to people. [20:21] Uh and I think our position is pretty [20:23] pretty much like everybody needs to have [20:25] access to the best AI. I have friends [20:27] that have small businesses, and they are [20:29] always trying to figure out how to [20:31] promote themselves and do that. Could [20:33] you explain from that point of view like [20:34] what it's going to be like for people [20:35] who are actually trying to reach new [20:37] audiences through ads? [20:38] >> Yeah. Yeah, I think that this that's [20:39] such a good question. Like, literally um [20:42] uh [20:42] I have a few friends who started this um [20:45] e-commerce company um selling shoes, and [20:48] uh [20:49] they did almost everything on their own [20:52] like the founders. They just like uh [20:55] go to the factory, get this done, get [20:56] the logistics done. But, when it came to [20:58] ads, they actually have to had to hire [21:00] like three performance marketers to do [21:02] the work because it's so so cumbersome, [21:05] so analytical. If you don't do it right, [21:07] you could end up wasting a lot of money. [21:09] So, I do think the vision has to be [21:11] where almost as easy as you're prompting [21:13] nowadays for questions, you could say, [21:16] "Buy goal is sell these shoes more in [21:19] Midwest and go." And then it comes back [21:22] and it's like, "Hey, I tried some [21:23] experiments and I think this is the [21:25] right bid given your price point. This [21:27] is the right way to do that. Do you want [21:29] to spend more money on this?" And then [21:30] you continue that conversation and [21:31] almost become an agent for that. But [21:34] today, literally, a small business has [21:36] to hire performance marketers, which [21:37] could be one of the biggest costs. In [21:39] some sense, actually like, you know, [21:41] just that cost of running ads through [21:44] through that is actually one of the [21:45] biggest costs in there. [21:47] Um uh which which of course then makes [21:49] things more expensive. So, I think the [21:50] vision would be that it is as easy as [21:53] just steering and telling what you need [21:56] from your business or describing the [21:57] what but not having to think about how [21:59] it will work and how many campaigns and [22:01] how much dollars and everything else was [22:03] like, "Hey, I want to spend this much. I [22:04] want to grow my business this much. [22:06] These are the constraints and um and and [22:08] ads are created and run to match your [22:10] constraints in some sense." Yeah, it's [22:12] it's a very interesting way to think [22:13] about it because auctions were [22:15] revolutionary. The idea that you just go [22:17] in there, I want to put these words out [22:18] there and pay for that to do it. But [22:20] that created an entire ecosystem of all [22:23] this sort of expertise and stuff that [22:24] you have to do. And it's really hard for [22:26] small businesses to try to play in that [22:28] space. [22:29] >> Yeah, I think as as as the competition [22:31] on that increase, I think like the [22:32] people who had more time and money to [22:35] spend on optimizing that and analyzing [22:37] the data and then running the best [22:39] possible ad [22:41] got the benefit from that versus like if [22:43] I didn't know that like, "Hey, actually [22:46] I think I give an example of an actual [22:48] brand which is Allbirds." It competed [22:50] with really big brands on shoes. But [22:52] somehow they found that every designer [22:54] in the tech company is going to love my [22:56] shoe. And and finding that niche and [22:58] then actually being able to create your [23:01] creatives and your message to focus on [23:03] that made them win in this like if you [23:06] go in Silicon Valley you'll see all [23:07] boards everywhere because of that. So I [23:09] think like but that I think is not [23:11] accessible to everyone if you're very [23:13] analytical and you have a whole team of [23:15] people who think about that you could do [23:16] that but theoretically the best products [23:18] can come to life if you can find out [23:21] where the right niche distribution for [23:23] it is and really go go for that. [23:26] I think another story on this is like [23:29] that there is this company which creates [23:31] a vegan ramen instant ramen which I love [23:34] because I don't have to feel bad about [23:36] eating it. [23:37] But it's such a weird concept vegan [23:39] instant ramen. Like I if I was just [23:42] thinking about it without knowing about [23:43] this company I'm like this can't exist. [23:45] Who will want this? But I think a really [23:47] good product can help you find and [23:49] discover those niche audience then you [23:51] build a really good product. So I think [23:53] it enriches everyone's life from that [23:55] perspective by enabling creation of [23:57] those products, selling of those [23:58] products. [23:59] Maybe it's not a multi-billion dollar [24:00] company but that's great. That's still [24:02] it's like a really big SMB [24:05] which is growing and it serve these [24:06] people who who care about that very [24:08] specific problem. [24:09] So I think it it really enriches people [24:11] life if you are able to create products [24:13] for these niches. [24:14] What does this look like in the future [24:16] where we're using things in a more [24:19] agentic way? [24:20] How do ads even work 10 years from now? [24:23] I think [24:24] next step would be more actual [24:26] conversational ad where you could truly [24:27] kind of understand what this product is [24:29] about. The next version would be kind of [24:30] work behind the scenes and actually [24:32] aggregate the best discounts and best [24:36] deals and the best version of the [24:38] product. Like for example if I I know [24:40] that I like ramen and let's say somehow [24:43] chat GPT has understood that preference [24:45] of mine then it could find that for me. [24:48] I didn't even know that that product [24:49] exists then in behind the scenes it [24:51] could actually say oh actually I found [24:52] this vegan ramen maybe that's something [24:54] that's valuable and of course [24:56] there is a marketplace where somebody [24:57] could say, "Hey, help people who are [24:59] like this to discover." Because Because [25:02] discovery goes from both directions, [25:03] right? Like, of course, I'm searching [25:04] for something, and then people want me [25:06] to discover something, and there's a [25:07] match between those. So, so I think it [25:09] will be more authentic but [25:12] in the future, [25:13] but at least the current modalities, I [25:15] think we start from there, improve it, [25:18] and make it relevant, make it [25:20] controllable, understandable, [25:21] trustworthy. And as I think the systems [25:24] evolve, then native the organic products [25:26] evolve, this will evolve as well with [25:27] that. Excellent. Well, so I thank you [25:29] for explaining this, and I look forward [25:31] to seeing what's going to happen next. [25:32] Awesome. Thanks for having me.