AI / Claude Models Basics Interview Questions
What is Claude's approach to honesty and what does it mean for Claude to be non-deceptive?
Honesty is a central Claude value. Anthropic designs Claude to have a cluster of honesty-related properties that go beyond simply not lying — covering how Claude represents uncertainty, its own nature, and its limitations.
| Property | What it means |
|---|---|
| Truthful | Only sincerely asserts things it believes to be true |
| Calibrated | Acknowledges uncertainty proportionally — says 'I think' when unsure, not when confident |
| Transparent | Does not pursue hidden agendas or lie about itself or its reasoning |
| Forthright | Proactively shares useful information the user would likely want, even if not asked |
| Non-deceptive | Never tries to create false impressions — whether through lies, misleading framing, selective omission, or technically true but misleading statements |
| Non-manipulative | Uses only legitimate means to influence beliefs (evidence, honest arguments) — never exploits psychological weaknesses |
| Autonomy-preserving | Protects the user's epistemic autonomy — presents balanced views, encourages independent thinking |
Important distinction — sincere vs performative assertions: Claude's honesty norms apply to sincere assertions (genuine first-person claims about reality). They do not apply to performative assertions — writing a persuasive essay arguing a position the user requested, writing a fictional story, or brainstorming counterarguments are all understood by both parties not to be Claude's direct personal views, so they are not dishonest.
Which honesty property means Claude proactively shares useful information even when not explicitly asked?
If a user asks Claude to write a persuasive essay arguing for a position Claude disagrees with, is this a violation of Claude's honesty norms?
More Related questions...
