One success is generalised into broad capability

What it is

Capability overgeneralisation occurs when success in one controlled case is treated as evidence of reliability across similar objects, language, places, or tasks. A demonstration strongly shows that the robot can act, while hiding dependence on lighting, placement, training coverage, and human preparation.

Why it happens

Users cannot inspect model coverage or control margin and infer category rules from sparse behaviour. Fluent success inflates perceived generality when conditions remain unstated. A high prior then reduces monitoring and contingency preparation in a changed situation, converting inference error into over-reliance.

Studying it

Initial demonstration diversity and fluency can be manipulated before near- and far-transfer tasks. Capability prediction, delegation, monitoring, and recovery reveal generalisation. Objective success must remain controlled. Technical knowledge, prior systems, and exposure to marketing are relevant covariates.

Where it stops holding

Generalisation is useful for mature, stable product categories. It becomes hazardous with narrow training coverage, severe tails, or similar-looking tasks requiring different perception. Failure can likewise be overgeneralised into total distrust; the objective is calibration, not suppression.

Applying it

• Pair successful demonstrations with their enabling conditions and at least one recoverable boundary case.
• Ask users to predict changed objects, lighting, or wording before revealing outcomes.
• Track expansion of delegation and reduced monitoring after early success; surface the relevant difference before an unsupported transfer.

Related

• Same group: X3.02.1 What a robot cannot do matters as much as what it can do · X3.02.3 Boundaries must be conveyed through behaviour, not only manuals
• Adjacent: X1.02 Robot types and expectations · X3.03 Truthful expression
• Search terms: capability overgeneralization · trust calibration · transfer of learning