AI & Responsible Technology · 04 Nov 2026
AI can help draft a proposal, summarise a document or suggest code. The risk is treating fluent output as proof of understanding. Someone who knows little about a topic may also struggle to spot a convincing error. Useful adoption therefore needs both tool skills and subject knowledge.
Understand the failure before designing the check
The NIST Generative AI Profile discusses risks including confabulation, privacy and problematic human reliance. A generated answer may contain plausible but incorrect claims. This makes verification a workflow decision: decide what must be checked and who is qualified to check it.
Use a draft, check, decide routine
Draft: define the task, audience and boundaries. Ask the tool to distinguish supplied facts from assumptions. Treat any requested citations as leads to inspect, not evidence by themselves.
Check: compare factual claims with original sources. Open referenced documents and confirm that they support the actual statement. For calculations, reproduce the numbers independently. For code, review permissions, dependencies and failure paths, then test in an appropriate environment.
Decide: assign a person responsibility for what is sent, published or deployed. Greater consequences call for a stronger review. A social caption and a decision affecting a person’s employment should not pass through the same approval process.
Set boundaries before sharing information
Do not paste customer records, credentials or confidential business material into a tool simply because the interface is convenient. Confirm which tools your organisation permits, what information may be used and how it is handled. Where possible, work with a minimal, anonymised example rather than a complete record.
Practise spotting an error
In an internal training exercise, give participants a sample AI-written response and ask them to identify every claim that needs evidence. Include an unsupported number, an ambiguous recommendation and a citation that does not match the sentence. Label the exercise as a simulation and use no real customer data.
Discuss how each issue would be detected in daily work. A checklist only helps if someone has time, access to sources and the knowledge needed to use it.
Takeaway: start with a low-consequence task, document the checks and compare the reviewed result with your normal process. AI literacy means knowing where the tool helps and recognising when human expertise must lead.
Topics: AI literacy, Human oversight
Sign in to comment or rate this insight.
Back to insights