Name the diligence concern
What you'll learn
Diligence is the responsible-deployment layer: oversight, bias, misuse, ownership, transparency. Match the concern to the shape of the deployment.
- Diligence = deployment-side concerns: oversight, bias, misuse, ownership, transparency.
- Concern is matched to deployment shape, not blanketed.
- Diligence interventions live in design, not in the prompt.
Diligence is the deployment-side D — what oversight, review, risk, and accountability practices wrap around the LLM in production. The other three Ds operate per interaction; Diligence operates at the system/policy layer. The exam frames diligence as 'responsible deployment' — and the questions reward the candidate who can name the *specific* concern rather than waving at 'be careful.'
The named diligence concerns worth memorising: (1) appropriate human oversight — high-stakes outputs need a human checkpoint before action; (2) bias awareness — outputs may reflect or amplify training-distribution biases, especially in personnel, lending, healthcare, legal contexts; (3) misuse risk — capability that helps a legitimate user can also help a bad actor; design for the malicious case; (4) ownership of outcome — the team deploying the system owns its consequences, not the model vendor; (5) transparency to users — they should know when they're talking to / being affected by an AI.
Naming the diligence concern in a scenario means matching the deployment shape to the relevant concern. A medical-advice chatbot: oversight, transparency. An auto-loan decision system: bias, oversight, ownership. A free-text customer support agent: misuse risk (prompt-injection), oversight on escalations. The concern is rarely 'all of them' — the question typically privileges one based on the scenario shape.
Diligence failures don't usually look like product bugs. They look like news stories — a discriminatory denial, a hallucinated diagnosis, a jailbroken assistant. By the time you see the failure, the diligence intervention should already have been in the design. The exam tests whether you'd have specified it in advance.
Key points
- Diligence = deployment-side concerns: oversight, bias, misuse, ownership, transparency.
- Concern is matched to deployment shape, not blanketed.
- Diligence interventions live in design, not in the prompt.
- Failure mode is news-story shaped, not bug-shaped — design for it upfront.
Examples
Primary diligence concerns: bias (training-distribution biases on names, demographics) + appropriate human oversight (no automated reject without human review) + transparency (candidate told an AI screening tool was used).
Primary concerns: misuse risk (prompt injection, jailbreak), appropriate oversight on escalations (handoff path to a human), and transparency (user knows they're talking to an AI).
Pitfalls
- Saying 'be careful' or 'add review' without naming *which* concern is binding.
- Treating diligence as a prompt-engineering fix. Diligence interventions are design, policy, and process.
- Skipping diligence on 'low-stakes' deployments that quietly become high-stakes (a research assistant whose summaries get pasted into legal filings).
Source notes: 00-academy-basics/notes/02-ai-fluency.md
Ask Claude
Build a prompt with this lesson + your question, copy it, and open the Claude Project in a new tab.