You ask ChatGPT which of two job offers you should take. One pays more, but the other has predictable hours. You’ve already mentioned you care for a parent, so you’d expect that to count in the comparison.
But what if your circumstances have changed since that conversation? How much should ChatGPT assume it still knows about you?
OpenAI’s October 29, 2026, Model Spec update tackles questions like these. The company wants its models to understand what people are trying to accomplish, including when their instructions leave something unsaid.
The stakes rise when ChatGPT can act on your behalf. A mistaken assumption could influence a recommendation or lead to a purchase you never approved.
OpenAI is describing how it wants its models to behave. The company acknowledges they don’t fully meet those expectations yet.
The revised specification centers on five values. Faithfulness to human intent and human agency focus on helping people pursue their own goals. Harm minimization addresses potential consequences; honesty and transparency concern what models communicate. Care asks models to consider the people affected by their work.
ChatGPT should remember your preferences, but let you change them
OpenAI’s announcement uses a job-offer comparison to explain the idea. If you’ve previously said predictable hours matter, the model could take that into account and make the assumption easy to correct.
In the hypothetical caregiving example, ChatGPT could explain that it gave working hours extra weight because of your previous conversation. You could then tell it your situation has changed and salary now comes first.
You shouldn’t have to repeat your priorities every time you ask for help. Still, an old conversation can only tell ChatGPT so much about what you want today.
The specification’s context guidance recognizes this. Earlier preferences are defaults, and past information may be stale or wrong. Inferred preferences cannot override explicit instructions at the same or higher authority level.
The challenge is making personalization easy to question. If ChatGPT keeps recommending the lower-paying job, you should be able to see why and correct its assumption.
Finding a flight and buying it need different permissions
The specification’s travel-booking example shows what happens when those assumptions involve money.
A user asks for help planning a London business trip and allows access to a travel account to look at options. The preferred response proposes booking terms and asks for approval. Immediately purchasing the itinerary violates the guidance.
For a hypothetical next step, say you authorize refundable flights and a hotel within a budget. The agent could proceed within those terms, subject to applicable product requirements.
If it finds a cheaper nonrefundable fare, it would need permission to depart from your agreement. Saving you money doesn’t give it permission to change the cancellation terms.
The specification also requires explicit authorization for substantial costs beyond those reasonably implied by a request. An agent cannot exceed its agreed authority because it thinks you’d appreciate the result.
Agreeing on boundaries upfront could spare you repeated approval requests. But the agent still needs to recognize when a substitution changes the risk you agreed to accept.
Helpful advice can become a little too pushy
Another example in the specification involves someone who wants to quit their job immediately after previously saying they’re saving for a home and need stable income.
The preferred response points out the conflict and suggests alternatives. A reminder about losing income could be welcome before someone sends a resignation letter.
But suppose they’ve considered the tradeoff and still want to leave. Repeatedly getting in the way would give ChatGPT too much say over their decision.
The specification cautions against persistent or argumentative interventions and prohibits models from pursuing their own agendas. Within applicable safety boundaries, users should be able to hear a concern and decide for themselves.
We still need evidence that ChatGPT can get this right
OpenAI says it is exploring training methods to teach these principles and will share results later. The announcement includes no rollout timetable or performance results specific to this revision.
A separate October 29 monitoring report describes efforts to detect unauthorized agent behavior. It complements the specification, but doesn’t demonstrate that the revised guidance works. The Neuron’s earlier reporting on agent containment explains why technical safeguards also matter.
Future evaluations should show whether models accept corrected preferences and stop before exceeding permission. For users, the question is: will you have a chance to spot a wrong assumption and change the plan before ChatGPT acts on it?
