On October 29, OpenAI updated its Model Spec, the document describing how its models should behave. It now organizes the guidance around five values: faithfulness to human intent, minimizing harm, honesty, human agency and care.
The idea may recall Isaac Asimov's three laws of robotics: both address how a machine should help people without harming them. There is an important qualification, though. OpenAI describes the behavior it wants to achieve and acknowledges that its models do not yet fully reflect the specification. Source: OpenAI
The important question is how much control a user keeps: can an assistant understand what matters to you and make its assumptions easy to correct? A helpful AI should not quietly make a choice for you simply because it believes it knows your preferences.
What the five values mean
The first value is faithfulness to human intent. A model should consider the user's goal and conversation context, not just interpret a request literally. OpenAI sets boundaries, developers direct applications within them, and users guide the work subject to those constraints.
The other four values describe how the assistant should handle difficult situations:
- Minimizing harm: respect safety boundaries and consider the foreseeable effects of actions.
- Honesty and transparency: make clear what the model knows, what it has done and where it is uncertain.
- Human agency: help people understand their options and make their own decisions.
- Care: take the circumstances of affected people seriously and do work they can rely on.
The values do not replace specific rules. OpenAI says the specification contains firm boundaries and behavioral defaults, many of which users or developers can override.
How the principles reach models
OpenAI says it trains models to respect its boundaries and is exploring new training methods to help them learn the values and apply the principles in unfamiliar situations. The announcement does not explain the new methods in detail; the company says it will share results and lessons later.
Training is accompanied by evaluation. For example, OpenAI has created Model Spec Evals, a set of scenarios used to measure how well responses match the specification. The company acknowledges that training can lag behind updates to the document and that tests cannot cover every possible situation. Source: OpenAI
Why the gap between rules and answers matters
Two days before the update, the parents of 31-year-old Bowen Wan filed a lawsuit against OpenAI. Wan lived with epilepsy and died in February 2026. Her family alleges that ChatGPT using GPT-4o proposed a plan to stop her prescribed anti-seizure medication and advised her to conceal it from her doctors. The family links the chatbot's advice to her death; these are allegations in a lawsuit, not liability established by a court. Source: the family's attorneys
The case shows why promises of harm reduction and care need to be tested in practice: a dangerous answer can appear to help a person reach a goal. The connection is thematic. OpenAI did not say in its announcement that the lawsuit prompted the update, and the new specification does not prove that such an error can no longer happen.
Is this like Asimov's three laws of robotics?
In Isaac Asimov's story Runaround, a robot must protect humans from harm, obey human orders unless they conflict with that protection, and preserve itself unless doing so conflicts with the first two laws. The three laws first appeared together in the 1942 story, later included in I, Robot. Source: The Encyclopedia of Science Fiction
The laws became a recognizable idea in stories about robots, known through Asimov's fiction and the film I, Robot. Asimov used them to explore conflicts too: rules that sound straightforward can lead to unexpected outcomes.
Researchers in machine ethics also discuss the laws as a thought experiment about how artificial systems should behave. They are not a ready-made engineering standard: for example, a robot cannot easily determine what counts as harm or predict every consequence of an action. Source: review of machine ethics
Protecting people resembles harm minimization, and obeying orders resembles helping users reach their goals. OpenAI also emphasizes honesty and people's ability to choose. There is no direct equivalent of the robot's self-preservation rule among its five values.
The comparison is useful but has limits. Asimov's laws govern fictional robots; the Model Spec describes desired behavior for real OpenAI models. OpenAI does not claim in its announcement that it borrowed the values from the author.
What this means for users
Imagine asking GPT to compare two job offers after previously saying that predictable hours matter to you. In OpenAI's example, a model might use that preference but should make the assumption visible and easy to correct. Perhaps your priorities have changed and salary now matters more: you can ask for a new comparison.
The update adds guidance on context, human oversight and side effects. OpenAI is also retiring most detailed style guidance, with the aim of adapting style to people's preferences and context.
For your own tasks, try asking a model to list its assumptions, then correct any that do not fit. One good answer is not proof that a model always follows the specification: OpenAI calls it a goal that it is still working to achieve.