From Indicator to Claim

Several standards processes are open at once, including the EU's harmonized standards, the revision of NIST AI RMF 1.0 and Australia's draft national AI safety standards. What counts as sufficient evidence is still being settled.

The paper puts forward a hypothesis, not a finding: standards written quickly may favor what is easy to verify, such as reporting deadlines and log retention, over what is hard to verify, such as whether a behavioral claim holds under real conditions. When a measure becomes a target, its relationship to the property it stands for can weaken.

The paper adapts the measurement framework that Wallach and colleagues propose for evaluating generative AI. That framework distinguishes the background concept, the systematized concept, the instrument and the measurement, and it offers seven lenses for interrogating validity. This paper applies the same logic to two further objects, and the extension is the author's own and hasn't been endorsed by the original authors.

Reading a requirement. For any behavioral requirement, three questions apply: which construct does it name, what evidence would count and what inference would that evidence support? The paper works through provisions of NIST AI RMF and the EU AI Act, showing where each leaves a question open for behavioral constructs.

Reading the evidence. The reviewer traces the chain from claim to conclusion and asks what each indicator can establish and what it cannot. Transparency is the worked example, because one word covers several constructs. A comprehension quiz used as evidence of understanding becomes an instrument that needs its own validation, and every step closer to behavior moves the problem without removing it.

Not every requirement that mentions people needs behavioral measurement. Behavioral constructs, such as trust, reliance and comprehension, are psychological states or behaviors. Assurance constructs with behavioral components, such as transparency, explainability and meaningful oversight, are defined in design or process terms but depend on how people respond. Safety, robustness, security and fairness fall outside the paper's scope unless a requirement specifies a behavioral component.

The failure to watch for is construct drift: a requirement defined as a process, such as a disclosure existing, gets evidenced as a human response, such as users understanding it, without anyone arguing the two are the same construct.

Behavioral measurement has a long record of documenting how evidence from people goes wrong. The paper sets out seven threats most likely to affect assurance evidence, each with the research that documents it and a design response.

Reactivity. People who infer a study's purpose change their behavior, so observed vigilance can exceed ordinary vigilance.
Prevalence effects. Rare targets are missed more often, so a test with many planted errors can overstate detection of rare real ones.
Stated versus revealed behavior. Self-reports are poor guides to actual processes.
Proxy-task validity. Preference for an explanation can be read as better decisions when it isn't.
Ecological validity. Participants, tasks and tools may differ from deployment.
Complacency over time. Vigilance degrades with sustained exposure to reliable automation.
Metric gaming. Compliance with the indicator replaces the property it stands for.

Evidence for several of these comes from other domains, such as visual search and human factors, so the paper treats their transfer to AI oversight as a hypothesis.