Skip to content
OpenAIOpenAI

Episode 13 - The Thinking Behind Ads in ChatGPT

How should advertising work in an AI product? Asad Awan, one of the ad leads at OpenAI, walks through how the company is approaching this decision and why it’s testing ads in ChatGPT at all. He explains how ads are built to stay separate from the model response, keep conversations with ChatGPT private from advertisers, and give people control over their experience. Chapters 00:00:29 — Mission and principles 00:04:01 — Separation between ads and answers 00:07:31 — Who will see ads 00:08:52 — Internal input and decision-making process 00:11:06 — Controls and how ads will work 00:15:53 — Guardrails for sensitive conversations 00:17:33 — Skepticism about ads 00:20:26 — Helping small businesses 00:24:13 — Future of ads

Asad Awanguest
Feb 9, 202625mWatch on YouTube ↗

CHAPTERS

  1. 0:00 – 1:55

    Why ChatGPT is adding ads: funding broad access to the best AI

    Andrew Main frames the episode around how ads will work in ChatGPT while preserving user trust. Asad Awan explains the mission-driven rationale: ads are a proven way to keep high-quality AI broadly accessible, especially for large-scale free usage.

    • Episode focus: ad design, eligibility, and protecting trust
    • Mission link: bringing beneficial AI to everyone at scale
    • Ads positioned as a practical model to fund higher limits and better models
    • Goal: ads should be helpful to users and businesses, not just monetization
  2. 1:55 – 4:01

    Ad principles: independence, privacy, transparency, and incentive alignment

    Asad lays out the core principles OpenAI wants to commit to before scaling ads. He emphasizes that the product’s foundation is trust, so ads must not compromise answer quality, privacy, or user understanding and control.

    • Answers must remain independent from ads (system and training separation)
    • Conversations are private; sensitive conversations will never show ads
    • Chats are not shared with advertisers; matching happens internally
    • Transparency and user controls are required to avoid “creepy” relevance
    • Company incentives should prioritize user value over time-spent or revenue
  3. 4:01 – 5:03

    The “wall” between ads and answers: technical + visual separation

    The discussion drills into how ChatGPT answers won’t turn into ad copy and why the model won’t be influenced by what appears in ad space. Asad describes a strict separation so users can clearly distinguish model output from ad content.

    • Model does not know what ad is being shown alongside the chat
    • Ads displayed in a clearly distinct area (e.g., bottom banner)
    • No implicit product recommendations injected into the model’s response
    • Separation designed to prevent perceived collusion or bias
  4. 5:03 – 5:28

    User-initiated ad context: “Ask ChatGPT about this ad”

    Asad explains the only way the model will discuss an ad: the user must explicitly opt in. This makes ad discussion functionally similar to pasting a link into ChatGPT and asking questions about it.

    • Explicit button: “Ask ChatGPT about this ad” to bring ad into context
    • Without opting in, the model will respond that it doesn’t know the ad
    • Design goal: don’t make it harder than asking about any web content
    • Maintains separation while enabling usefulness when users want it
  5. 5:28 – 7:28

    Preventing long-term drift: why trust is the core business model

    Andrew challenges whether the separation will hold years later when ad revenue grows. Asad argues that OpenAI’s long-term success depends on maintaining trust across consumer, enterprise, and future device experiences, so drifting would undermine the business.

    • Trust positioned as OpenAI’s “core business,” not ads
    • Assistant-style relationship requires safe handling of personal data
    • Enterprise context reinforces trust requirements (high-stakes customer data)
    • Incentives and identity: trust isn’t optional for this product category
  6. 7:28 – 8:53

    Who will see ads: tiers, subscriptions, and enterprise separation

    Asad clarifies eligibility: ads will appear for Free and Go tiers, while paid plans and Enterprise remain ad-free. He connects this to serving different user segments with different business models.

    • Ads shown on Free and Go tiers
    • No ads on Plus, Pro, or Enterprise
    • Subscriptions offer an ad-free upgrade path
    • Ads help avoid making the free tier overly limited
  7. 8:53 – 11:07

    How decisions get made internally: debates, roundtables, and a priority rubric

    Asad describes OpenAI’s internal process as research-culture-driven and debate-heavy, involving broad input across the company. He introduces a decision rubric that ranks trust above all else, explicitly rejecting “creepy but effective” targeting.

    • Company-wide input via many roundtables shaped the ad principles
    • Simple but strict rubric: trust > user value > advertiser value > revenue
    • Example: even a “good” ad fails if it triggers surveillance fears
    • Governance forums for privacy and safety guide product decisions
  8. 11:07 – 12:50

    Personalization and controls: seeing, limiting, or clearing ad-related data

    The conversation turns to what users can control and how personalization can still be trustworthy. Asad describes transparency into data used for ads and user controls ranging from limiting chat usage to turning personalization off entirely or upgrading to remove ads.

    • Personalization can improve relevance (e.g., trip planning → camping gear)
    • Transparency: users can see what data is used for ads
    • Controls: exclude past chats, allow only ad clicks, or disable personalization
    • Sensitive chats are never used; users can clear data entirely
    • Ad-free option remains via Plus/Pro
  9. 12:50 – 13:39

    Ad frequency and quality bar: show fewer ads, and only when helpful

    Asad explains that the system aims to be conservative with ad load, especially during rollout. The guiding rule is usefulness: if there’s no good match, it’s better to show nothing than to degrade the experience.

    • Early rollout will show very few ads to learn placement and experience
    • Primary criterion: usefulness and relevance, not maximizing impressions
    • Quality standards apply to both ChatGPT content and ad content
    • If no good match exists, no ad is shown
  10. 13:39 – 14:36

    Guardrails for sensitive conversations: definitions, detection, and enforcement

    Andrew asks how OpenAI will identify sensitive contexts and prevent inappropriate ad placement. Asad describes rigorous policy definitions and model-driven classification to filter out sensitive topics (health, politics, violence, etc.) from ad matching and display.

    • Sensitive areas include health, politics, violence, and other high-risk topics
    • Policies defined rigorously by specialized teams and reviewed internally/externally
    • Models classify conversations with high precision to apply protections
    • Sensitive contexts are filtered out—no ads shown and no matching done
  11. 14:36 – 17:34

    Ad design direction: balancing “native” feel with clear separation

    Asad discusses the product design tradeoff between ads that feel integrated versus ads that remain unmistakably distinct. He says the initial approach is conservative and clearly separated, with room to evolve formats as learning accumulates.

    • Design spectrum: native/non-jarring vs. clearly separated and obvious
    • Initial approach: conservative and clearly distinct to protect trust
    • Format will evolve based on data and user feedback
    • Core constant: answers remain clearly separate from ads
  12. 17:34 – 20:27

    Addressing “no ads” skepticism: rebuilding confidence in online advertising

    Asad responds directly to audience objections, acknowledging industry-earned suspicion around privacy and incentives. He argues OpenAI must earn trust through stronger principles, transparency, control, and an upgrade path for those who prefer no ads.

    • Skepticism is valid given the history of online ads
    • OpenAI’s responsibility: clearer rules, better transparency, stronger controls
    • Ads should be powered by strong AI to be genuinely useful
    • Users can opt out by upgrading if ads aren’t acceptable
  13. 20:27 – 24:14

    Helping small businesses: simplifying ads into an AI-guided workflow

    The conversation shifts to advertisers, especially small businesses that lack performance marketing expertise. Asad envisions an agent-like experience where owners describe goals and constraints, and the system runs experiments and optimizes campaigns without requiring specialized teams.

    • Today’s ad platforms are complex; SMBs often must hire specialists
    • Vision: conversational setup—state goals, region, budget, constraints
    • System runs experiments, recommends bids, and iterates with the owner
    • Better targeting could help niche products find their audiences
  14. 24:14 – 25:34

    Future of ads in an agentic world: conversational discovery and deal aggregation

    Asad closes by imagining more agentic ad experiences over time, from richer conversational ads to behind-the-scenes discovery that finds the best products, discounts, or deals based on user preferences. He reiterates that future evolution must remain controllable, understandable, and trustworthy.

    • Near-term: more conversational ads that explain products better
    • Longer-term: agents that aggregate best deals and surface relevant discoveries
    • Marketplace dynamic: matching user needs with businesses seeking discovery
    • Non-negotiables: relevance, control, transparency, and trust as systems evolve

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.