Clear, reliable safeguards can make people comfortable delegating more. Control over context and permissions can expand what users are willing to try.
Control is more useful when people can predict what will happen. A system that is unpredictable about refusals, memory or actions can make users feel less free even when it has few visible restrictions. I would judge freedom partly by whether people can understand and adjust the relationship, not simply by the number of requests the assistant will accept.
Make preferences and boundaries distinguishable
Consider a user who wants blunt criticism of their writing. That preference should be easy to express without also granting permission to publish the draft. Treating both as one ‘less restricted’ mode hides the choice. Our proposal is to offer understandable controls for style and task scope, while explaining consequential boundaries separately.
Unnecessary refusal is a real design problem
XSTest provides a research approach for identifying safe requests that are refused. It supports examining the costs of broad restrictions. It does not imply that every refusal is unnecessary, or that a model with fewer refusals is always preferable.
Source: Röttger and colleagues · XSTest: identifying exaggerated safety behavioursA provider can describe almost any restriction as a way to make users feel safer. That argument becomes paternalistic if users have no meaningful influence, no explanation and no way to challenge mistakes. Controls also have a cost when people must learn a complicated settings system.
Where the debate remains open
A credible design would offer sensible defaults, specific explanations and a path to correct errors. It would distinguish an individual’s preferences from decisions affecting others, and make the reasoning behind that distinction open to challenge.
What would change this view?
I would change my view if users consistently understood a simpler design better and achieved their legitimate goals with fewer errors. More controls are not automatically more agency; the test is whether people can use them to shape outcomes they care about.
For more reading
Background evidence for this editorial argument, including the limits and counterpoints. The conclusions are the site’s interpretation.
- XSTest: identifying exaggerated safety behaviours
Pairs safe prompts with unsafe contrasts to investigate unnecessary refusals. Historical model results are not current rankings.
- OpenAI Model Spec · 18 December 2025
The provider’s intended behavior and instruction hierarchy; a policy is not proof of consistent behavior.
- Claude’s constitution
Describes the values Anthropic intends to train into Claude, including tensions between them.
Sources reviewed 13 September 2026. Product documentation can change. How we use evidence
How does this argument land with you?
Participate anonymously. No account required.