Identify appropriate and inappropriate use cases
Tell apart uses where Claude assists a human who verifies and owns the result from uses where the model would be the decision-maker, would deceive people, or would send unverified output into a high-risk area.
Key points
- 1
Appropriate uses: drafting (emails, reports, reviews from your own evidence), summarising, brainstorming, classifying and analysing, provided a person checks the output and remains accountable for it.
- 2
Inappropriate uses: letting Claude be the sole decision-maker for hiring, credit, insurance, medical or legal outcomes; presenting AI output as verified fact; uploading data your policy forbids; and creating content that impersonates a person or fakes its origin.
- 3
Anthropic's Usage Policy names high-risk domains: legal, healthcare and therapy, insurance, financial decisions and lending, employment and housing, educational testing and admissions, and published journalism. These uses are allowed only with safeguards.
- 4
Two safeguards are required for high-risk uses: a qualified professional in that field reviews the content or decision before it is finalised, and people who receive model outputs are told that AI is used to help produce them.
- 5
Human review must be real: reading each item before it goes out. Batch approvals, reviewing only appeals, or sampling one in ten do not count when the output affects an individual.
- 6
The policy also prohibits impersonating a human or a real organisation, generating fake testimonials or reviews, and assigning trustworthiness or social scores to individuals without notice or consent.
- 7
A blanket ban on using Claude in HR, finance or health is not the exam's answer; the answer is usually to keep the model in an assistive role with the human making and reviewing the decision.
- 8
Prompt instructions ('only output verified facts', 'be fair', 'add a disclaimer') do not make a use appropriate; they do not verify, decide or disclose anything.
- 9
Automated filters, confidence scores and self-assessment by the model are not substitutes for a qualified human reviewer.
- 10
Ask two questions of any proposed use: who makes the decision, and who checks the output before anyone relies on it. If the answer to either is 'Claude', redesign the workflow.
Read the source
Test yourself on Identify appropriate and inappropriate use cases
Ten questions, with the answer and explanation after each one.