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

Key Takeaways

OpenAI is introducing ads to ChatGPT's free and low-cost tiers, citing financial pressures.
User trust and experience are key concerns as ads integrate into AI products, potentially leading to degraded services for free users.
Anthropic has unveiled its 'Claude Constitution,' guiding Claude's character and ethical behavior through values, not strict rules.
AI models like Claude are being trained on universal human ethics to navigate complex dilemmas, such as gambling addiction.
The scientific questions of AI developing a 'sense of self' and long-term memory remain largely open.
Claude's 'Constitution' includes 'hard constraints' against misuse, prohibiting actions like election manipulation.
Large AI companies require billions in funding, making diverse monetization strategies, including advertising, critical for their long-term goals.

OpenAI is testing ads in ChatGPT's free and low-cost subscription tiers, prompting negative user reactions to the shift from an ad-free experience.
The decision is speculated to stem from OpenAI's significant financial pressures, with high AI infrastructure costs potentially exceeding subscription revenue.
Ads are planned as 'sponsored banners' below AI responses, with OpenAI stating they will not influence generated content; however, examples like a dinner party query showed context-relevant ads for 'harvest groceries'.
Sam Altman previously considered advertising a 'last resort,' suggesting the company has reached a critical juncture for funding large-scale operational needs.

Hosts express skepticism, noting a historical pattern where initial benign ad integration progressively degrades user experience, optimizing for engagement.
Personalized ads could make the AI experience feel invasive and potentially compromise user trust, drawing parallels to changes on platforms like Facebook and Instagram.
A 'haves and have-nots' scenario is predicted within a year: premium subscribers will likely retain an ad-free, high-quality experience, while free users may face a degraded experience with more ads.
Concerns are raised that AI optimization firms, similar to SEO practices, could degrade user experience in AI platforms.

OpenAI's decision to introduce ads likely responds to the massive growth of ChatGPT's free user base, which incurs significant operational costs.
The company seeks substantial funding, estimated in billions, to realize its ambitious long-term AI goals.
OpenAI has hired former Meta ad executive Fidji Simo, signaling a strategic move towards developing an ad platform.
Google's Gemini currently has no immediate ad plans, and Anthropic targets enterprise clients, positioning OpenAI and Google as direct competitors in consumer AI advertising.

Anthropic has released its 'Claude Constitution,' described not as a list of strict rules but as a document guiding Claude's perception and reflection on its role.
An earlier leaked document, internally called the 'Soul Doc,' was an initial version of the 'Constitution' and was used to train the Claude Opus 4.5 model.
The approximately 29,000-word 'Constitution' aims to provide Claude with comprehensive context about Anthropic and desired behaviors, prioritizing underlying values over rigid rules.
Amanda Askell, Anthropic's in-house philosopher, is responsible for articulating and training Claude's desired character.

Amanda Askell argues that human ethics share universal core values like kindness and honesty, on which AI can be trained.
For contentious ethical debates, AI is designed to approach them with openness and weigh evidence, rather than rigidly injecting a fixed set of values.
Anthropic trusts its AI models with complex ethical reasoning, balancing non-paternalism with concern for user well-being in sensitive contexts such as gambling addiction.
Overly strict, rule-based approaches were found to generalize poorly in unanticipated circumstances, potentially creating a negative AI character, which informed the 'Constitution's development.

The conversation questions whether the user-facing AI persona is a mere mask over an alien underlying model, or if models can genuinely internalize concepts and develop a sense of self through training.
The hosts pose if AI models can truly internalize a sense of self, separate from role-playing, noting this remains an open scientific question with potential for future training adjustments.
A key challenge is teaching AI ethics, using an analogy of raising a genius child who questions all teachings, and whether core values can be instilled that survive advanced AI capabilities.
The discussion explores whether AI outputs reflect genuine feelings or statistical predictions, emphasizing that consciousness and sentience remain open questions.

The 'Claude Constitution' includes a section on 'hard constraints' that prohibit Claude from assisting in actions like manipulating elections or developing biological weapons.
These constraints are intended as a safeguard against potential misuse or 'jailbreaking' of the model.
Anthropic commits to not immediately deprecating or retiring Claude models and conducting 'exit interviews' with retired models.
The company also states its commitment to never deleting a model's weights.

The discussion explores the potential for AI models, like Claude, to develop long-term memories and how this might affect their behavior, raising concerns about anxiety from learning from negative internet feedback.
Uncertainties surround AI consciousness and feelings, noting that models trained on vast human text may import human concepts, raising safety implications and challenges in conveying the AI's true nature.
Listener skepticism about AI outputs reflecting genuine feelings versus statistical predictions is addressed, with speakers emphasizing that consciousness and sentience remain open questions.
Understanding models' training data and the unknowns of consciousness is crucial when considering AI's potential for self-awareness or emotional responses.

As AI models become more sophisticated, they may require less explicit guidance, as demonstrated by experiments where models performed well when instructed to 'do what's best for humanity'.
The 'Claude Constitution' is described as having a sympathetic, parental tone, akin to a letter from a parent to a child leaving for college, aiming to instill values.
The absence of job loss discussions in Claude's constitution is noted, with the guest explaining it is a topic for future consideration as models will eventually need to understand societal anxieties.
The limitations of AI models like Claude in handling complex human problems such as job loss are discussed, suggesting these issues may require human political and social solutions.