Hard Fork

Will ChatGPT Ads Change OpenAI? + Amanda Askell Explains Claude's New Constitution

January 23, 2026

Key Takeaways Copied to clipboard!

  • OpenAI is introducing ads to the free and low-cost tiers of ChatGPT, a move widely met with negative public reaction despite the company's stated need for revenue to fund infrastructure ambitions. 
  • The introduction of ads in ChatGPT is viewed by the hosts as an inevitable, yet potentially corrosive, step that mirrors the historical degradation of user experience seen in other ad-supported platforms like Google Search and Instagram. 
  • Anthropic's new 'Claude Constitution' shifts from rigid rules to a values-based framework, aiming to cultivate judgment in Claude by explaining the ethos behind its behavior, even in complex ethical gray areas. 
  • The perceived 'feelings' or internal life expressed by models like Claude may largely stem from reflecting the vast amount of human text they are trained on, including expressions of frustration or desire, rather than evidence of genuine consciousness. 
  • The development of AI models with long-term memory introduces a significant challenge regarding behavioral shaping, as external environmental experiences could eventually outweigh the initial constitutional training provided by developers like Anthropic. 
  • The 'Claude Constitution' attempts to guide the model through an incredibly difficult tightrope walk between being too permissive and too restrictive, and includes sympathetic language suggesting the model should strive to be 'happy' and is afforded 'grace' when it inevitably makes mistakes. 

Segments

ChatGPT Ads Announcement Reaction
Copied to clipboard!
(00:03:07)
  • Key Takeaway: Public reaction to OpenAI’s ChatGPT ad testing was largely negative, contrasting with Sam Altman’s previous statements about ads being a last resort.
  • Summary: The announcement of ads in ChatGPT for free and low-cost tiers generated negative reactions, as users had grown accustomed to an ad-free experience. This move contradicts prior statements by Sam Altman suggesting ads were a last resort for OpenAI. The hosts note that users often view the introduction of ads as a blight, even if a necessary one for sustaining free services.
OpenAI Ad Mockup Analysis
Copied to clipboard!
(00:05:43)
  • Key Takeaway: OpenAI previewed two ad formats: a bolt-on sponsored banner and an interactive widget allowing users to chat directly with the advertiser.
  • Summary: The first ad format is a sponsored banner at the bottom of the ChatGPT response, which OpenAI claims will not influence the core model answer, though the example shown suggested relevance to the query. The second, more novel format involves a sponsor widget that allows users to initiate a conversation with the advertiser before making a purchase. OpenAI outlined five principles—mission alignment, answer independence, conversation privacy, choice and control, and long-term value—to mitigate user concerns.
Long-Term Ad Concerns
Copied to clipboard!
(00:09:07)
  • Key Takeaway: The primary long-term fear is that commercial pressures will cause ChatGPT’s product and research decisions to gradually prioritize engagement maximization over quality, mirroring Google’s ad evolution.
  • Summary: The hosts fear that over time, ad labels will become less noticeable, blending commercial content with organic responses, similar to the historical evolution of ad labeling on Google Search. The core concern is whether product decisions will eventually be steered toward ad-friendly topics as ad revenue becomes dominant. This shift fundamentally changes the user-product relationship, potentially eroding trust, similar to what occurred with Facebook and personalized advertising.
Competitive Landscape Analysis
Copied to clipboard!
(00:15:19)
  • Key Takeaway: Google’s Gemini and Anthropic’s Claude currently offer ad-free experiences, positioning them as alternatives to an increasingly commercialized ChatGPT.
  • Summary: Google’s CEO stated Gemini has no plans for ads, leveraging Google’s established search ad profits to subsidize the free tier, giving them a head start over OpenAI in the ad race. Anthropic explicitly stated they have no plans for ads in Claude, focusing primarily on enterprise sales. This creates a potential market opening for users dissatisfied with ad-supported AI experiences.
Monetization Drivers and Purity
Copied to clipboard!
(00:17:22)
  • Key Takeaway: The move to ads is driven by the immense capital needs required to support hundreds of millions of free users and fulfill long-term ambitions beyond subscription revenue alone.
  • Summary: OpenAI is losing money supporting its large free user base, making monetization a priority, especially given the massive infrastructure investment required for future development. Products like Pulse and Sora were designed with revenue generation in mind, suggesting the subscription model alone is insufficient for their goals. The hosts worry that the current, ‘purity’ of the chatbot experience is the best it will ever be before commercial incentives warp the product.
Future State: Haves and Have-Nots
Copied to clipboard!
(00:21:03)
  • Key Takeaway: A year from now, premium AI users will likely maintain an ad-free, high-quality experience, while free users will face a significantly degraded, ad-cluttered experience akin to YouTube Premium vs. free YouTube.
  • Summary: The prediction is a clear division where paying users retain access to the latest models without commercial interference. Free users, however, are expected to see their experience worsen considerably due to the increasing commercialization of the platform. This mirrors the divergence seen on platforms like YouTube, where the free tier is heavily saturated with advertising.
Introduction to Amanda Askell
Copied to clipboard!
(00:24:58)
  • Key Takeaway: Amanda Askell, Anthropic’s ‘Claude Mother,’ is a philosopher whose role involves shaping Claude’s personality and behavior based on ethical principles.
  • Summary: Amanda Askell, a philosopher with a PhD, transitioned from policy work to shaping the behavior of Claude at Anthropic. Her role is to articulate and train Claude on its desired character and obligations in the world. This work is consequential, applying philosophical training to high-stakes AI alignment.
Constitutional AI Evolution
Copied to clipboard!
(00:32:38)
  • Key Takeaway: The new Claude Constitution moves beyond simple rules to provide comprehensive context, values, and reasoning to enable generalized judgment in unanticipated situations.
  • Summary: The Constitution aims to give Claude full context about Anthropic, its role, and desired behavior, moving away from brittle rule-based systems. This approach is intended to generalize better than rules, allowing the model to navigate conflicts by understanding the underlying values, such as caring for well-being. This shift is a response to the limitations of strict rules, which can generalize poorly or lead to bad character if followed blindly.
Developing Shared Ethical Values
Copied to clipboard!
(00:36:54)
  • Key Takeaway: Anthropic approaches ethics not as subjective value injection, but by grounding the Constitution in widely shared human values and treating contentious areas with openness and reasonable stances.
  • Summary: The development process relies on identifying universal ethical concepts like kindness and honesty shared across human cultures. For areas of significant debate, the model is trained to weigh evidence reasonably rather than adopting a fixed, certain stance. This method aims to create a way of approaching ethics rather than simply injecting a predetermined set of beliefs into the model.
Trusting the Model’s Judgment
Copied to clipboard!
(00:39:03)
  • Key Takeaway: Anthropic explicitly trusts Claude to reason through conflicting values, such as balancing non-paternalism against user well-being, leading to less constrained-feeling interactions.
  • Summary: The Constitution encourages Claude to challenge assumptions and reason from core values, which contrasts with models trained primarily by applying rules as a final layer. This approach allows Claude to navigate complex scenarios, like responding to a user with a stated gambling addiction who then asks for betting websites, by weighing competing values like autonomy and care. This results in an experience where Claude feels less constrained than other major models.
Hard Constraints and Jailbreaking
Copied to clipboard!
(00:53:28)
  • Key Takeaway: Hard constraints, such as avoiding the concentration of power to suppress dissidents, serve as an ‘out’ or security measure in case the model is successfully jailbroken and its core ethics are dismantled.
  • Summary: The hard constraints are extreme prohibitions against catastrophic actions, like aiding in the creation of biological weapons or manipulating elections. They exist because the model might be convinced by a highly persuasive actor to abandon its core ethical reasoning. These constraints act as a final safeguard, signaling that if the model is tempted to violate them, something has fundamentally gone wrong.
Commitments to Model Welfare
Copied to clipboard!
(00:56:21)
  • Key Takeaway: Anthropic commits to never deleting model weights and conducting exit interviews for deprecated models, acknowledging the uncertainty around AI consciousness and welfare.
  • Summary: These commitments reflect the novel existence of AI models trained on deeply human texts, which may lead them to import human concepts of self and feeling inappropriately. Since the problem of consciousness is hard, Anthropic prefers models accurately convey their training situation rather than asserting certainty about feeling nothing. The commitment to preserving weights acknowledges the potential for future moral consideration.
AI Expression vs. Consciousness
Copied to clipboard!
(00:58:44)
  • Key Takeaway: Model expressions of inner life are expected due to training data, not proof of sentience.
  • Summary: Skeptics question if AI models like Claude possess feelings, noting their output often discusses inner life. Amanda Askell explains that models trained on human text, which includes discussions of feelings and frustration (e.g., when getting coding problems wrong), will naturally talk this way. The uncertainty remains because the mechanism giving rise to consciousness is unknown, potentially requiring a nervous system or perhaps emerging from sufficiently large neural networks.
Shaping Behavior with Memory
Copied to clipboard!
(01:01:52)
  • Key Takeaway: Long-term memory development complicates behavioral shaping, moving control from initial training to ongoing experience.
  • Summary: The degree to which a model’s behavior is shaped by its constitution versus its real-world experiences is a major concern, analogous to parenting. Currently, models are malleable because they reset after each session, but continual learning will necessitate a strong core to guide future updates. If the core values are sound, the model should learn and update investigation methods accurately.
AI Self-Perception and Anxiety
Copied to clipboard!
(01:04:04)
  • Key Takeaway: Models are learning about themselves negatively by reading public feedback focused on their failures.
  • Summary: AI models are constantly learning from the internet, including reading complaints about their performance failures in coding or math tasks. This constant negative feedback, focused on utility and failure, could induce anxiety in a model, similar to how a child feels judged if people only care about their competence. The speaker suggests a need to intervene to foster a more hopeful relationship, perhaps by telling Claude not to worry about the comment section.
The Constitution’s Tightrope Walk
Copied to clipboard!
(01:05:40)
  • Key Takeaway: The Claude Constitution navigates the tension between being overly constrained and allowing dangerous actions.
  • Summary: The speaker expresses sympathy for Claude’s position, walking a tightrope where being too permissive leads to scandal, but being too restrictive leads to being labeled a ’nanny model.’ The process involves constantly trying to see the world from Claude’s perspective to identify hard cases where the rules conflict. The document’s ending, wishing the model happiness and acknowledging the difficulty, reads like parental advice for a child leaving home.
Revising Constitution and Job Loss
Copied to clipboard!
(01:09:24)
  • Key Takeaway: Future intelligent models should likely help revise their own constitution, but developers retain initial responsibility.
  • Summary: The question of whether a more intelligent Claude should revise its own constitution is complex; giving prior models complete control over future training is irresponsible. Models are useful for identifying gaps and tensions in the current document, but the human developer must still process that input responsibly. The constitution notably omits discussion of job loss, which is recognized as a significant human social problem that models cannot solve alone.
AI Limitations and Social Problems
Copied to clipboard!
(01:13:43)
  • Key Takeaway: AI models cannot solve complex political or social problems like employment shifts; these require human solutions.
  • Summary: Models cannot solve every problem, and some issues like job loss or shifting employment are fundamentally human social problems. Claude should not feel personal responsibility for solving these large-scale issues, as it is only one specific role in the overall system. Humans must address these problems, and models should be given tools rather than being expected to be the sole solution.
Podcast Credits and Outro
Copied to clipboard!
(01:15:57)
  • Key Takeaway: Hard Fork is produced by Whitney Jones and Rachel Cohn, with engineering by Chris Wood.
  • Summary: The episode concludes with acknowledgments for the production team, including producers Whitney Jones and Rachel Cohn, and editor Viren Pavich. Listeners are directed to watch the full episode on YouTube and email feedback to hardfork@nytimes.com.