Understand the ethical implications of AI usage
Recognise bias, fairness, transparency and accountability issues in AI use and respond as the accountable human: correct biased outputs, be honest about AI involvement, and treat every affected person to the same standard.
Key points
- 1
Accountability stays with the people who use AI output. If a figure or claim from Claude reaches a board, a client or the public, the person who produced and presented it owns its verification and must correct the record when it is wrong.
- 2
'The AI got it wrong' is not a defence, and neither is 'AI figures are understood to be approximate'; presenting output as fact was a human decision.
- 3
Models can reproduce biases from the language they were trained on, for example gender-coded wording in job descriptions. The ethical practice is to review outputs for such patterns, correct them, and add instructions and a checklist so the check is routine.
- 4
Asking the model whether its own output is biased, or switching to a larger model, does not replace human review for fairness.
- 5
Bias often enters through data and criteria rather than the model: complaint volumes reflect who complains, historic decisions reflect past prejudice. Scrutinise inputs, add measures that capture the outcome you care about, and check results across groups before acting.
- 6
An instruction to 'be fair' cannot correct for information that is missing from the data; equal allocation that ignores real differences in need is not fairness either.
- 7
Transparency means telling people when AI helped produce content that affects them, never presenting an AI-drafted message as written entirely by a named human, and never removing disclosure because it hurts satisfaction scores.
- 8
A public statement about your AI process must be true. Claiming 'all responses are reviewed' while sampling one in ten is misleading, and for welfare-affecting guidance each response needs qualified review anyway.
- 9
Fairness across customers: do not tier the standard of review or care by how much a customer pays; unreviewed drafts for low-value customers is discriminatory practice.
- 10
Distractor patterns the exam uses: attributing errors to the tool or vendor, treating industry-standard wording as acceptable, dropping an analysis entirely to avoid bias, and adding caveats instead of fixing a misleading claim.
Test yourself on Understand the ethical implications of AI usage
Ten questions, with the answer and explanation after each one.