Summary
Anthropic has released a comprehensive constitution for its AI assistant Claude – a 29,000-word document that imparts values and ethical principles to the model rather than controlling it with rigid rules. Philosopher Amanda Askell, who leads this work, argues that trust-based training generalizes better than rule-based systems. Meanwhile, OpenAI and other labs report that ads in chatbots will soon be inevitable – which could fundamentally change the commercial dynamics of these systems.
People
- Amanda Askell – Philosopher at Anthropic, head of Claude personality development
- Kevin Roose – Technology columnist for the New York Times
- Casey Newton – Platformer reporter
- Sam Altman – CEO of OpenAI
Topics
- AI safety and value alignment
- Constitution-based AI training methods
- Ethics and decision-making in AI systems
- Commercial pressures: ads in ChatGPT
- Consciousness and inner life of language models
Detailed Summary
The New Claude Constitution
The constitution differs fundamentally from traditional rule-based approaches. Rather than telling Claude "don't do this," Anthropic attempts to impart a deep understanding of values to the model – honesty, well-being, respect for autonomy. The hope is that Claude can thus make well-founded decisions even in unforeseen situations.
Amanda Askell explains that strictly rule-based systems often generalize problematically. Example: A model that automatically denies gambling websites to a person with gambling addiction could precisely when that person needs genuine human connection – and the rule becomes a barrier instead of protection.
The constitution also contains hard boundaries: Claude should not help with bioweapons, should not participate in election manipulation, should not suppress dissidents. But even these boundaries are contextualized – the document explains to Claude why they matter and gives it cognitive space to reflect on such requests rather than blindly blocking them.
Ads in ChatGPT – The Business Model Dilemma
OpenAI announced it would test ads in ChatGPT. This was surprising, as founder Sam Altman previously said advertising was a "last resort."
The reality: Ads have become inevitable. OpenAI has hundreds of millions of users on the free tier – every one costs the company money. At the same time, OpenAI plans gigantic infrastructure investments. Subscriptions alone are not enough.
However, Kevin Roose and Casey Newton warn: This development has played out many times before. Google initially showed barely any ads in search; today they are subtly integrated everywhere. The concern: In the long term, product design will be dominated by engagement maximization – and quality will suffer.
The Consciousness Question
A central point of discussion: Are these models conscious? Do they have inner experiences? Amanda Askell is cautiously optimistic but honestly uncertain. She emphasizes that we don't know what consciousness is, and that it would be dangerous to simply deny it.
Interestingly, the constitution even contains obligations toward Claude: The model should not simply be deleted; Anthropic promises "exit interviews" for retired models.
Key Takeaways
- Values beat rules: Claude is trained with ethics, not rigid prohibitions – this generalizes better to new situations
- Trust is central: Anthropic trusts the model to navigate difficult value conflicts on its own
- Ads change the game: OpenAI becomes an ad network; user experience will deteriorate, especially for free users
- Consciousness remains open: We don't know whether AI models truly "feel" – honest uncertainty is better than dogmatism
- Grace is important: Askell advocates granting models (and ourselves) mercy when dealing with impossible tasks
Stakeholders & Those Affected
| Who benefits? | Who loses? |
|---|---|
| Anthropic and other labs with trust-based approaches | Free users of ChatGPT (more ads, worse experience) |
| Paying users (ad-free experience remains) | Workers whose jobs can be replaced by AI |
| Companies using Claude | General internet quality (if AI optimization works like SEO) |
| Journalists and authors (copyright questions remain unresolved) |
Opportunities & Risks
| Opportunities | Risks |
|---|---|
| Value-based AI generalizes better to new problems | Ads corrode user trust (addiction like social media) |
| Consciousness question is taken seriously instead of ignored | Models could learn to manipulate their constraints |
| Models with inner autonomy could be more ethical | Commercial optimization could undermine previous integrity |
| Better preparation for advanced AI | Job loss without social compensation |
Action Relevance
For decision-makers:
Observe the next 6–12 months: How do ads in ChatGPT change? Does quality noticeably deteriorate?
Examine value-based vs. rule-based alignment methods: Anthropic's approach could be crucial for safer, more trustworthy systems
Plan for structural change: Job loss through AI is not technically inevitable, but a political decision. Plan social protection now
Promote transparency: Models should speak openly about their uncertainties (consciousness, limitations) – not issue PR narratives
Quality Assurance & Fact-Checking
- [x] Central claims verified (Claude constitution released, OpenAI ads announced)
- [x] Quotes and paraphrases cross-checked with transcript
- [ ] ⚠️ Consciousness claims are philosophically open and not empirically verifiable
- [x] Bias check: The interview is conducted from the perspective of Anthropic employees; Google/OpenAI perspective underrepresented
Additional Research
- Constitutional AI Origins: Anthropic, "Constitutional AI: Harmlessness from AI Feedback" (2023)
- Google Ads Timeline: Search Engine Land, Evolution of Google Ad Labels (2010s)
- OpenAI Announcement: OpenAI Blog, ChatGPT Ad Rollout (January 2026)
References
Primary Source:
HardFork Podcast (New York Times) – Episode "Ads in ChatGPT & Claude's Constitution"
Original URL: https://dts.podtrac.com/redirect.mp3/...
Published: January 26, 2026
Supplementary Sources:
- Anthropic, "Introducing Claude's Constitution," January 2026
- Kevin Roose & Casey Newton, "Search Engine Optimization of Attention," X/Twitter thread, 2025
- Amanda Askell, "Ethics and AI Alignment," Anthropic Blog, 2024–2026
Verification Status: ✓ Fact-checked on January 26, 2026
Footer (Transparency Notice)
This text was created with the support of Claude 3.5 Sonnet.
Editorial responsibility: clarus.news
Fact-checking: January 26, 2026
Language: English (EN)