Reports are that Anthropic has a dedicated safety & ethics interview round. Here's how to build the first principles to crack it. Start with the distinction most candidates miss. AI safety is stopping the model from causing harm. Mechanical and observable. A guardrail that blocks medical misinformation. A confirmation gate before an agent sends an email. AI ethics is deciding what the model should and shouldn't do. Upstream and often invisible. A policy that the model won't draft termination letters. iTutorGroup's hiring AI worked exactly as designed. The ethics of the design were the failure. Then learn the sizing framework. The candidates who scored highest in our Cohort 3 mocks all used SHIR in the first 90 seconds of every safety question, before they proposed anything. S: Severity. Physical harm sits above discrimination sits above embarrassment. H: Harm scope. 10 users versus 10M users is a different response. I: Immediacy. Active harm or latent risk. Sets response speed. R: Reversibility. Can the action be undone. Determines whether you ship with monitoring or add hard confirmation gates. Pause, write the four letters on a notepad, then come back with structure. That move scored higher than any other in mocks. Then memorize four precedents. Cite by name and month. Air Canada chatbot, Feb 2024. British Columbia tribunal held the airline liable for a chatbot hallucinating a bereavement fare. The defense that the chatbot was a separate legal entity got rejected. Companies own their AI's representations. iTutorGroup, Aug 2023. $365K EEOC settlement. Hiring AI auto-rejected women 55+ and men 60+. About 200 qualified applicants screened out. Bias liability lands on the employer even when the algorithm does the discriminating. Mobley v. Workday, July 2024. First AI vendor held directly liable as an "agent" under Title VII. Collective certification followed under ADEA in May 2025. Vendor liability is no longer theoretical. Gemini image gen, Feb 2024. Alphabet shed roughly $90B in market cap in the days after the pause. Sundar called the outputs "unacceptable." The cost of acting is almost always lower than the cost of being seen as not acting. Then practice the three moves that consistently scored highest in our cohort mocks. Tier the response. Three options side by side with a cost on each. Binary "pull or ship" reads junior. Reframe under pushback. When the VP says "wait until earnings," the highest-scoring answer reframed revenue to headline. End with documentation. If leadership overrides you, write a memo to the manager, the safety lead, and legal. On the record. The full kit: → Video version on YouTube: youtube.com/watch?v=RaBw5S… → Coaching with Ankit Virmani (AI PM, Uber), Prasad Reddy (ex-CPO), and Dr. Bart Jaworski (12,000+ PMs coached) in Cohort 3, which opens tomorrow: landpmjob.com → Mock walkthrough and writeup on the newsletter: news.aakashg.com/p/safety-ethic…
