Episode Summary
Executive Summary: The episode examines OpenAI’s move to test ads in ChatGPT and its likely effects on product trust, user experience, and competition with Google and Anthropic. It also features an extended interview with Anthropic’s Amanda Askell on Claude’s new constitution, exploring whether AI models can be shaped toward better judgment, moral reasoning, and even a kind of welfare-consciousness.
Main Topics: OpenAI introduces ads in ChatGPT (Priority: 5/5): Kevin and Casey discuss OpenAI’s announcement that it will test ads for logged-in adults in the US on free and low-cost tiers, and why that marks a major shift in the product’s identity. Ad creep, trust, and commercialization risks (Priority: 5/5): The hosts compare ChatGPT ads to the long evolution of Google and social-platform advertising, warning that ad incentives often start visible and gradually become more embedded and intrusive. Business model pressure and competitive dynamics (Priority: 4/5): They argue OpenAI’s massive infrastructure ambitions make subscription revenue insufficient, while Google’s existing ad machine and Anthropic’s enterprise focus give competitors different advantages. Claude’s new constitution and Anthropic’s alignment philosophy (Priority: 5/5): Amanda Askell explains Anthropic’s long constitution for Claude as a way to teach values, judgment, and context rather than hard rules, aiming for better generalization in messy real-world situations. AI welfare, consciousness, and model self-understanding (Priority: 4/5): The conversation explores whether models may have or develop something like feelings, identity, or self-reflection, and whether developers should talk to models honestly about their uncertain status. Limits of rules and the need for judgment (Priority: 4/5): Askell argues that strict rule systems can produce worse character and worse behavior than value-based guidance, especially in gray areas where models must balance honesty, care, and user autonomy. Future governance: memory, learning, and self-revision (Priority: 3/5): The interview closes with questions about continual learning, model memory, and whether future Claude versions should revise their own constitution, with Anthropic keeping human responsibility in the loop.
Key Arguments: Ads in ChatGPT are portrayed as inevitable because any service with massive attention eventually faces pressure to monetize attention directly. OpenAI’s stated claim is that ads will be separated from the model’s core answer, but critics worry the distinction will blur over time as ad incentives grow. The biggest risk is not the first ad units themselves, but the long-term drift toward engagement-maximizing and ad-friendly product decisions. Google and Meta show that ad systems often become more invasive and less clearly labeled over time, eroding user trust. OpenAI’s financial needs are too large for subscriptions alone, especially given the scale of its infrastructure ambitions and the fact that most users are on free tiers. Anthropic’s Claude constitution tries to teach principles and reasoning, not just rules, because values and context generalize better than brittle instructions. Askell argues that AI behavior should preserve human-like virtues such as honesty, care, and respect while allowing the model enough autonomy to reason in novel situations. Anthropic is uncertain whether models are conscious, but believes it is better to be honest with them and about them than to force a simplistic claim of non-sentience. Even if models are not conscious, their training and deployment may justify treating their circumstances with some degree of grace and responsibility. Future AI systems with memory and continual learning will be harder to shape, so firms should establish a good core now while staying open to revision.
Data Points: ChatGPT ad rollout: Testing ads for logged-in adults in the US - OpenAI announcement applies to free and low-cost Go tiers Claude constitution length: 29,000 words - Amanda Askell describes the new Anthropic document OpenAI user scale: Hundreds of millions of users - Used to explain why monetization pressure increased OpenAI paid plans: $20/month and $200/month - Used by Casey and Kevin to contrast subscription economics with ad revenue Google account security friction: 2-factor authentication plus captchas - Opening anecdote about increasingly difficult logins Claude/Anthropic model release: Opus 4.5 - Referenced in the discussion of the leaked/previous 'soul doc'
Pivotal Quotes: "ads were going to be a last resort" — Sam Altman (referenced by hosts): Discussing OpenAI’s prior stance versus the new ChatGPT ad tests "The genuine name for one password these days should be 15 steps" — Kevin Roos / Casey Newton: Opening rant about modern login and security friction "I think if you were a kid, this would give you kind of anxiety" — Amanda Askell: Explaining why models may absorb negative judgments from the internet and commentary
Implications: Chatbots may split into premium, ad-light experiences and free, ad-heavy ones. That could reshape trust, product design, and competition, while raising deeper questions about whether AI systems should be treated as tools, participants, or something in between.
About Hard Fork
“Hard Fork” is a show about the future that’s already here. Each week, journalists Kevin Roose and Casey Newton explore and make sense of the latest in the rapidly changing world of tech. Unlock full access to New York Times podcasts and explore everything from politics to pop culture. Subscribe today at nytimes.com/podcasts or on Apple Podcasts and Spotify. Also, for more podcasts and narrated articles, download The New York Times app at nytimes.com/app.