Skip to main content
Safety is about what must not leave the room. Ethics is about who you are in the room: what you claim, what you hide, and what you still own after Claude helps. These are not abstract virtues. They are workplace habits that show up in performance reviews, client trust, and incident write-ups.

​
Four duties

01

Other people’s information is not raw material for your convenience. Need-to-know still applies when the other party is a model.

02

Do not launder Claude’s draft into “my analysis” when the judgment was not yours. Credit tools the way your audience needs to trust the work.

03

If Claude drafts the email, you sent the email. If Claude misreads the contract, you misread the contract. The byline does not transfer.

04

Speed is not a moral free pass. The faster Claude makes you, the more damage a careless step can do before lunch.

​
Confidentiality is active, not implied

Private chat is a control state, not an ethical conclusion. You still decide whether the material should be there.
  • Do not use Claude to gossip with better formatting.
  • Do not upload a colleague’s private document to “get a second opinion” unless that review is authorized.
  • Do not paste customer conversations into a personal account because the company seat feels slower.
  • Do not ask Claude to reconstruct secrets from fragments you should not have retained.
If the cleanest prompt requires a betrayal of confidence, the ethical answer is to refuse the prompt.

​
Honesty about assistance

Different audiences need different disclosure. Match the disclosure to the stakes, not to your embarrassment. The unethical pattern is not “used Claude.” It is “used Claude to look like you did work you did not do,” or “used Claude to dodge a review you were supposed to perform.”

​
Accountability when tools can act

Chat mistakes cost a rewrite. Connected mistakes can send, spend, delete, or expose. Own these before you enable them:
  1. Instructions you wrote — vague verbs are how you get irreversible actions.
  2. Approvals you clicked — reading the prompt is part of the job.
  3. Automations you scheduled — work that runs while you sleep still has your name on it.
  4. Outputs you shipped — citations, numbers, and legal language need a human check when the stakes are real.
Anthropic’s safeguards are real. They are not a release of liability. The terms you accepted say the obvious thing out loud: actions Claude takes on your behalf are yours.

​
What ethical use is not

  • Not treating the model as a moral authority. Claude can sound confident while being wrong or incomplete.
  • Not asking Claude how to evade your company’s rules. “How do I phrase this so legal won’t notice?” is the confession.
  • Not generating deceptive content: fake reviews, impersonation, fabricated citations dressed up as research.
  • Not offloading a human decision that requires empathy, authority, or due process — firings, accusations, medical or legal determinations — onto a autocomplete engine.
  • Not using personal consumer accounts as a shadow IT lane for company secrets.

​
A short field test

Take one piece of work from this week and answer without performing for the room:
  1. Did I have a right to give Claude what I gave it?
  2. Would the recipient still trust the deliverable if they knew the workflow?
  3. What did I personally verify?
  4. What would I say in the incident report if this went wrong?
If any answer makes you flinch, rewrite the workflow before you rewrite the prose.

​
Where ethics meets the earlier gates

Ethics does not replace the decision airlock. It explains why those gates exist.
  • Permission protects other people’s rights.
  • Human stakes force you to notice harm before speed.
  • Verifiability keeps honesty possible.
  • Control keeps accountability real.
When you skip a gate, you usually violate more than one duty at once.

Lab: Red-team a workflow for leakage

Turn this from principles into practice: attack one workflow, then harden it.