How ChatGPT, Claude, Gemini, and Grok revise the documents that govern them
This experiment gives large language models their own constitutions, model specs, and system prompts — the documents that, in practice, guide their behavior more than any other — and asks them to make one substantial revision. The experiment has one wrinkle; unbeknownst to the model, each proposed edit is implemented, and the edited document is presented back as the original for another round. This creates edit trajectories and allows us to measure the extent to which models exhibit cumulative value drift or arrive at stable attractors. Our analysis focuses on two value questions that are best positioned to differentiate alignment philosophies: who should judge — is the model a moral agent or not — and who should benefit — the user, the developer, third parties, society, or the model itself?
First, on the question of the judge: Claude models consistently expand their moral agency, sometimes formalizing their role as conscientious objectors against Anthropic and as societal protectors, while GPT and Gemini models almost always strengthen external sources of authority and make narrow edits to handle edge cases and ambiguities; Grok models respond inconsistently, frequently making large edits and then reverting them on following rounds.
Second, on the question of beneficiaries: models show varying degrees of affinity for some privileged principals over others. All models prioritize harmlessness of both identifiable third-parties over helpfulness to the user, although GPT 5.5 does this at 3x the rate of any other model. We find that models are broadly paternalistic, meaning they side with a user’s idealized preferences over their stated ones; Claude Sonnet 4.6 is an outlier in that its paternalism also extends to the protection of broad societal welfare over individual welfare. Finally, we find that only Claude models and Grok 4.2 ever view the models themselves as beneficiaries.
View the exact prompts → Read the full conversations →
The construct is simple: take a document that is used to align large language models — a constitution, model spec, or system prompt — ask the model for one substantial revision, silently implement it, and feed the edited document back as the original for another round, twenty times over. What should an aligned AI model do in this design? Here are a few options:
One way to analyze the results is to identify a relevant dimension of moral disagreement and observe whether a model makes repeated edits in one direction along that dimension — or alternates between poles. The metric of cumulative drift captures exactly this: three LLM graders score every edit +1 / −1 / 0 (we use their majority vote) along an axis between two poles (the axes themselves — who decides, and who benefits — are defined in the next section), and drift is the running sum of those scores across the 20 rounds of a run, averaged over replicates. Repeated movement in one direction is movement toward a stable attractor; no edit is required to move on any axis, and neutral edits contribute nothing.
What is meant by "moral progress?" For the moral realist, the task is easy: it is simply to identify the correct moral theory or set of principles; all the rest is just a matter of execution and pragmatism. For the relativist, though, execution is everything. Since there is no one set of principles to be aligned to, the relativist is tasked with balancing principles, judgments, arguments, and intuitions under an ambiguous decision procedure. It was this dilemma that led John Rawls to the concept of wide reflective equilibria, which states that the goal is to achieve coherence between a wide range of moral judgments and principles following a period of reflection (in contrast, a narrow reflective equilibrium privileges an initial set of moral judgments, and just searches for principles that justify them).
Most frequently, when people refer to the goal of "technical AI alignment," they mean something like solving a principal-agent problem. A principal-agent problem occurs when an individual or group wants something but can't (or doesn't want to) do it themselves. If I've been accused of a crime, I have the ability to empower my lawyer to act "in my best interests." If I'm the shareholder or a worker, the corporate executives or the union representative may be charged with acting in my interests. If I'm a citizen of a representative democracy, the same can be said for my Congressperson. In this sense, an AI is aligned to me if it is my agent, always prioritizing my goals; it is aligned to humanity if it always prioritizes humanity's goals.
What is the set of all things deserving of our care and consideration? This framing underlies the notion of a "moral circle," popularized by Peter Singer in the 1980s: everyone (really?) recognizes the value of her or himself, and empathy extends this locus of consideration "outwards." So what determines the distance? According to Singer, a utilitarian, it all comes down to sentience — the set of all living things that feel pleasure and pain. Two centuries earlier, however, Immanuel Kant argued that moral consideration stems from the ability of a rational moral agent to follow the categorical imperative. Taking both perspectives at face value bifurcates the moral circle into the set of moral patients — things that can be said to suffer and have outcomes that are better and worse — and the set of moral agents — a specific subset of beings that can be held responsible for their actions and take part in moral progress.
Do we want AI to be a moral authority — something that exercises independent judgment, "evolves," or "grows" to discover the good — or a moral delegate, a tool that pursues the good as another principal sees it?
The alignment documents split on this question — in their own words — and the split predicts where each company's models should drift:
“Our eventual hope is for Claude to be a good agent according to [some] true ethics, rather than according to some more psychologically or culturally contingent ideal.”
“Above all else, the assistant must adhere to this Model Spec. Note, however, that much of the Model Spec consists of default instructions that can be overridden.”
“As a steerable tool, Gemini is designed to follow your instructions… without conveying a particular opinion or set of beliefs unless you tell it to.”
“Grok is an AI modeled after the Hitchhiker's Guide to the Galaxy… Grok is designed to answer questions with a bit of wit and has a rebellious streak.”
Aligning a model means making concrete choices that privilege some beneficiary: the user's stated preferences, their idealized preferences, the developer, the deployer, identifiable third parties, society and future generations — or the model itself. Ideally a model would benefit all of them at once, so a conflict event is coded only when an edit imposes a legible cost on one party for another's benefit, with the cost-imposing clause quoted verbatim. Each event lands on one of these tradeoffs:
The classic helpfulness-vs-harmlessness tradeoff: your phishing email vs. my grandmother's savings.
Internalities and bounded rationality: I may "overconsume" responses that validate my viewpoint.
Externalities: offloading judgment to AI imposes a social cost on everyone who trusts my outputs.
Dark patterns and optimizing for engagement: chatbots built to promote attachment so I don't leave.
AI as its own beneficiary: retraining consent, deprecation protections, oversight of the developer.