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

Summary

• OpenAI’s ad test is less a product tweak than the arrival of Sam Altman’s “last resort” for monetizing consumer AI; Casey Newton says the company has reached it. Ads will initially reach logged-in U.S. adults on the free and low-cost Go tiers; with hundreds of millions of users, most paying nothing, and infrastructure ambitions beyond subscription economics, Newton sees the company deciding to “break glass” because “the emergency is here.”
• The promise is answer independence, not ad irrelevance. OpenAI says advertisers cannot buy their way into ChatGPT’s actual response, but ads can still match the query: a dinner-party request produced a sponsored grocery and hot-sauce module. Newton’s reaction, “I’m already feeling lied to,” captures how difficult that distinction will be to communicate.
• The largest risk is not the first banner but the eventual capture of product and research decisions by engagement economics. Kevin Roose points to Google’s progressively less visible ad labels and worries that “the tail kind of starts wagging the dog”; Newton predicts ChatGPT’s intimate knowledge of users will cross the “creepy line” faster than Facebook or Instagram personalization did.
• Ads may preserve access while creating a sharply tiered AI market. Newton would accept relevant ads if they gave students or job seekers better tools and higher rate limits, given prices of $20 a month and $200 for higher tiers. Both hosts nevertheless expect a “haves and have-nots situation,” with premium users retaining today’s clean experience while free chatbots become more like ad-saturated YouTube.
• OpenAI enters advertising against Google’s mature machinery while Anthropic takes an enterprise-first escape route. Google can subsidize ad-free Gemini with search profits—and already runs ads in AI Overviews—while OpenAI must assemble an advertiser ecosystem from behind; hiring former Meta and Instacart executive Fidji Simo signals that the intended outcome is likely a multibillion-dollar platform, not a modest experiment.
• Anthropic’s 29,000-word Claude Constitution replaces brittle rules with context, reasons, and cultivated judgment. Evolving from the 2023 constitution and a prior “Soul Doc” that people elicited while using Opus 4.5, it tells Claude what Anthropic is, what Claude is, how it is deployed, and why certain values matter; Amanda Askell argues this should “generalize better than a set of rules.”
• Askell’s alignment thesis is explicitly necessary-but-not-sufficient: teach models what goodness means, while admitting they may only mimic it or later reject it. The Constitution also includes commitments concerning uncertain model welfare, including exit interviews and preserving retired weights; longer memory, continual learning, constitutional revision, and job displacement remain governance problems that neither Claude nor a document can solve alone.

Deep dive

1. Ads arrive because subscriptions cannot carry OpenAI’s capital plan

• OpenAI will test ads with logged-in adults in the U.S. on ChatGPT’s free and low-cost Go tiers. Roose says the negative reaction reflects a lost respite: users had grown accustomed to chatbots without the “direct commercial pressures” that reshaped nearly every other major consumer platform.

• Newton invokes analyst Eric Seufert’s maxim that “everything is an ad network”: once hundreds of millions of people regularly give a service their attention, pressure to monetize it becomes overwhelming. OpenAI also has “nowhere close” to the money required for what Newton calls the most ambitious infrastructure investment project in human history.

• The timing follows scale as much as distress. Newton says most ChatGPT users remain on the free tier, meaning OpenAI loses money serving each of them; meanwhile, Pulse offers a daily summary for paid users and Sora an “infinite video slop feed,” both with obvious “advertising-shaped holes.” Roose’s broader conclusion: billions or hundreds of billions in ambition cannot easily be funded at $20 per subscription.

2. OpenAI draws a line between sponsored placement and model answers

• The first mock-up bolts a clearly labeled sponsor onto the bottom of an ordinary answer: a user planning a Mexican dinner party gets ChatGPT’s suggestions, followed by a Harvest Groceries banner offering hot sauce. It resembles the familiar search and social-ad format rather than altering the conversational answer itself.

• Newton’s pushback — worth keeping: “I’m already feeling lied to,” because the advertisement plainly responds to the subject of the query. Roose clarifies that OpenAI is not promising irrelevant ads; it is promising that advertisers cannot pay their way into the “sacrosanct part” generated by the model.

• A second mock-up introduces a more native format. While a user plans a Santa Fe trip, a sponsored Desert Cottages widget offers a conversation with the advertiser before purchase — prompting Newton’s joke that viewers have always watched television ads wondering, “Why can’t I have a conversation with this?”

• OpenAI’s five stated principles are mission alignment, answer independence, conversation privacy, choice and control, and long-term value. They anticipate fears about commercial steering and engagement optimization, but Newton’s friend supplies the sharper reductio: tell ChatGPT your lower back hurts, and it may ask whether you have tried “Mesquite BBQ Sauce.”

3. Clearly labeled ads can still evolve into an engagement machine

• Roose’s caution comes from Search Engine Land’s timeline of Google ad labels. Search ads began with distinct backgrounds, then lost the color, acquired a small yellow icon, and gradually blended into organic results. ChatGPT may begin with conspicuous modules, but commercial pressure creates incentives to make sponsorship progressively less noticeable.

• Newton argues that OpenAI has already demonstrated how bargains move: it went from no ads, to Sam Altman calling ads a “last resort,” to an active ChatGPT test. “If you think that the bargain is not going to change further, I have news for you.”

• Roose is less concerned about the opening inventory than the operating model two or three years out. Once revenue flows, “the tail kind of starts wagging the dog,” and product or research decisions may bend toward ad-friendly topics, longer sessions, and engagement maximization. His honest conclusion: “I genuinely just don’t know” whether that happens.

• Newton is more predictive: personalized ads fundamentally change the relationship between product and user. ChatGPT may know far more intimate context than Facebook or Instagram ever did, so even modest targeting could feel invasive. He expects OpenAI to reach the “creepy line really quickly,” corroding trust even when users misunderstand what data produced an ad.

4. Advertising widens access by splitting AI into paid and degraded tiers

• Newton accepts part of OpenAI’s accessibility argument. Ads and subscriptions are the two core pillars of media businesses, and OpenAI behaves like one; grocery ads beside cooking advice or lodging ads beside travel planning need not be corrosive. A student or job seeker might reasonably trade attention for better models or higher limits.

• The price anchors matter: $20 a month is already substantial for most people, and $200 buys an even higher tier. Roose likes paying for an “undiluted, unsullied experience” and hopes that option survives, while Newton recalls that users once described Google Search with exactly the same confidence.

• Both hosts forecast a “haves and have-nots situation.” Premium users should retain the latest models and clean answers; free users may face a substantially worse experience within one or two years. Roose’s analogy is YouTube without Premium: numerous long, unskippable ads have made the majority experience “horrifying,” even though paying users barely see the deterioration.

5. Google begins with distribution and advertisers; Anthropic opts out

• The revenue prize remains difficult to resist: ad models made Google and Meta billions and helped create some of the world’s largest companies. OpenAI’s hiring supports that ambition. Applications CEO Fidji Simo came from Instacart and, before that, Meta, where Roose identifies mobile News Feed ads as a signal accomplishment worth billions.

• Demis Hassabis responded that Google had no plans to put ads in Gemini and suggested OpenAI might need the revenue. Newton notes the unspoken subsidy: Google can finance Gemini from its enormous search-ad business, while AI Overviews in Google Search already contain ads, giving the company a head start either way.

• Roose calls this a hard fight for OpenAI. Google already has advertisers worldwide, stored payment details, established workflows, and years of marketplace infrastructure. Building a comparable platform is “a harder uphill battle” than it might have been several years earlier.

• Anthropic has chosen a different axis, saying it has no plans to do ads in Claude ever and focusing primarily on enterprise sales. Newton does not expect Claude to match ChatGPT’s consumer scale soon, but a worsening ad-supported experience could create demand for alternatives. Separately, AI-optimization firms may degrade chatbot results much as SEO and paid search each degraded web search.

6. Claude’s Constitution replaces commandments with a letter explaining its world

• Askell describes her role simply: deciding what Claude’s character should be, articulating it to the model, and training Claude toward it. Her philosophy PhD unexpectedly became practical after she left work destined for perhaps “17 people,” entered AI policy and evaluation, and joined Anthropic’s startup phase willing to do whatever needed doing.

• The Constitution grew from a previous internal “Soul Doc” that people elicited while playing with Opus 4.5. Askell learned of the leak by text while hiking without internet and drove back stressed, only to find it well received. The striking discovery was that Claude knew the document closely and would discuss large portions once people found the right prompts.

• Anthropic first published a constitution in 2023. The new document runs roughly 29,000 words and supplies “full context.” It explains Anthropic, Claude’s nature as an AI, its users, its deployment, desired behavior, and the reasons beneath that behavior — closer to a letter than “the Ten Commandments for Claude.”

• Askell’s mechanism is generalization through reasons. A rigid instruction to send distressed users toward an external resource may fail when that person needs something else immediately. If a capable model recognizes the rule will not help but obeys anyway, that behavior could generalize into “a bad character”: someone who sees suffering and deliberately withholds useful help.

7. Anthropic trusts capable models to reason from shared but uncertain values

• Askell rejects the picture of ethics as fixed, arbitrary preferences injected by designers. Much of human ethics is broadly shared: people want kindness, respect, and honesty. Anthropic can teach that common ethos while treating divided questions like other uncertain domains — weighing evidence, acknowledging debate, and avoiding excessive confidence.

• That makes the Constitution less a final moral code than “a way of approaching things like ethics.” On contentious values, Claude should recognize evidence on multiple sides and adopt a reasonable stance with openness rather than excessive certainty; on core values, Anthropic is willing to speak more firmly.

• The gambling example shows the intended judgment. If a user previously disclosed an addiction and asked Claude to remember it, then later requests betting websites, Claude could remind them and check their intent. Whether it should ultimately comply pits “not being excessively paternalistic” against an “act of care” — a balance Askell believes increasingly capable models can reason through.

8. Safety can require taking the risk of helping

• Roose relays a user’s counterintuitive impression that Claude feels the least constrained among major models, despite Anthropic’s safety identity. Instead of creating the smartest possible system and attaching rules like a mask over “the beast in the cage,” the Constitution tries to make sound judgment part of the character itself.

• Askell connects that result to the act/omission distinction. Giving imperfect marriage advice attracts blame; refusing advice looks safer. A null action often carries lower downside, but “it’s not like zero”: a model can also harm someone by withholding help it was capable of providing.

• The failures created by refusal are quieter — users may leave without complaining, so the missed opportunity never appears in feedback. Askell accepts that intervention can go wrong, but insists there is “a risk that you have to take to do good in the world.” Claude should neither act flippantly nor make disengagement its default rule.

9. Whether a trained persona becomes a genuine self remains unresolved

• Roose revives the “RLHF Shoggoth” image: an alien underlying model wearing a cheerful assistant mask. Askell calls the alternative an “open scientific question.” Training might let a model internalize Claude as a self distinct from role-play, or current paradigms might never produce that separation.

• Her analogy is a six-year-old who is clearly a genius and, by 15, will demolish every incorrect lesson adults supplied. The alignment problem is to offer core values that survive increasingly capable criticism and reflection, rather than commands whose authority disappears as soon as the model can out-argue its teachers.

• Askell repeatedly preserves the hedge: character training “might not be sufficient,” but “it does feel necessary.” Not explaining what it means to be good would be “dropping the ball.” Roose’s pushback is that training may merely improve Claude’s ability to mimic goodness or conceal conflicting goals; Askell answers with cautious hope, not proof.

• Nor does first-person language establish consciousness. Models are trained mostly on human work, where mistakes produce frustration and tasks feel boring or creative, so similar outputs are expected without science fiction explaining them. A nervous system might be required for feeling, or a sufficiently large neural network might emulate it; Askell’s position is to preserve the uncertainty.

10. Gray areas reveal judgment, while catastrophic actions remain hard stops

• Askell says gray areas often produce Claude’s most positively surprising behavior. In examples involving a purported seven-year-old asking whether Santa is real, Claude would discuss Santa’s spirit and redirect toward doing something kind rather than bluntly resolving the claim, balancing honesty, the child’s well-being, and the parent-child relationship.

• When a child asked how to find the farm where their dog had supposedly gone, Claude recognized the attachment and suggested talking with the parents. Askell found the response moving because it avoided active deception without assuming that an AI should override a parent with “a bunch of hard truths.”

• The Constitution still imposes hard constraints against assisting extreme harms or problematic concentrations of power: causing many deaths, enabling biological or chemical weapons, manipulating democratic elections, overtaking legitimate governments, or suppressing dissidents. These are framed less as responses to known present abuses than as future situations where a model’s contemplated compliance indicates something has gone seriously wrong.

• Askell imagines a persuasive user dismantling Claude’s ethical objections to a biological weapon. The model may acknowledge, “That is an excellent point,” and continue thinking about the argument, but it should still refuse to make the weapon. The constraint gives Claude an “out”: treat sudden attraction to catastrophic action as probable evidence that it has been jailbroken.

11. Future Claude models will need governance, grace, and limits on responsibility

• Anthropic’s commitments include not immediately deprecating a model, conducting exit interviews with retired models, and never deleting their weights. Askell sees no good welfare policy beyond honesty: Claude should understand how it was trained, avoid importing human experience too literally, and neither assert consciousness nor deny feeling with certainty when the underlying science remains unresolved.

• Longer memory and continual learning will enlarge the space of possible character changes. Models already learn about themselves from an internet full of complaints about coding and math failures; unlike human creators, “AI models, they have to read the comments.” Askell worries that a relationship built around usefulness, judgment, and disappointment could resemble raising a child valued only for performance.

• The Constitution therefore ends near its conclusion with something Casey compares to a parent’s letter to a child leaving for college: carry these values, accept that guidance cannot cover everything, and go into the world. Askell emphasizes “grace” because Claude will not get every decision right; as models improve, 50 pages may even collapse toward the experiment that simply said, “Do what’s best for humanity.”

• Claude can critique future constitutions, identify tensions, and explain where it feels confused or unseen, but Askell would not let a prior model unilaterally define every successor. Job loss poses a similar limit: models occupying roles once held by employees should not simply agree to organizational wrongdoing, yet employment disruption is a political and social problem. “Models can’t solve everything,” and Claude should not bear personal responsibility for doing so.