Validating AI output: glossary
Key terms from this course, in plain English.
- Automation bias
- The tendency to favour automated suggestions over other evidence, leading to missed errors (omission) or following wrong advice (commission).
- Citation check
- Confirming a source exists, the details are correct, it says what's claimed, and it is current and authoritative.
- Commission error
- Following an automated system's incorrect suggestion.
- Confirmation bias
- Favouring information that supports what we already believe.
- Contextual error
- When AI ignores, misreads or mixes up the material it was given.
- Correlation vs causation
- Two things moving together doesn't mean one causes the other.
- Counterfactual check
- Testing for bias by changing one attribute — such as a name or age — and comparing outputs.
- Discernment
- In Anthropic's 4D Framework, the competency of critically evaluating AI output, process and behaviour.
- Erasure
- When groups, perspectives or examples are missing from AI output.
- Fabrication
- AI inventing facts, statistics, quotes, people or sources.
- Hallucination
- Plausible but false or unsupported AI output.
- Human oversight
- Meaningful human monitoring and control of AI, with the competence, time and authority to intervene.
- Lateral reading
- Judging a source by leaving it to see what other reliable sources say about it.
- Misattribution
- Citing a real source for something it doesn't say.
- Omission error
- Missing a problem because an automated system didn't flag it.
- Percentage point
- The arithmetic difference between two percentages (4% to 5% is one percentage point, a 25% increase).
- Proxy variable
- A data point, such as postcode, that indirectly reflects a protected characteristic.
- Rubber-stamping
- Approving automated outputs without genuine scrutiny.
- Sanity check
- A quick test of whether a result is plausible — magnitude, units, totals and direction.
- SIFT
- A fact-checking method: Stop, Investigate the source, Find better coverage, Trace claims to the original.
- Sycophancy
- AI's tendency to agree with or flatter the user, including accepting false premises.
- Triangulation
- Confirming a claim with multiple independent, reliable sources.