Practice
A persona is an audience description, not a user model
Personas communicate who a product is for, and do it well. The thing they are routinely asked to do instead — stand in for a specification of human capability — is a job they were never built for.
A persona names an audience and gives it a face, which is a real communication problem solved. The difficulty begins when it is asked to carry weight it cannot bear: the proportion of a population matching a persona falls as attributes are added to it, and no procedure exists for showing a persona to be wrong. A model of the user specifies capability ranges and can fail a test. The two artefacts answer different questions.
Chapman and Milham, writing from inside Microsoft where they could count more than two hundred personas in use across development teams, made the first sustained methodological critique of the method. Two of their arguments matter for product work.
The first concerns coverage. A persona accumulates attributes because attributes are what make it vivid: a role, a level of experience, a device, a working pattern, a motivation. Each addition narrows the set of real people matching on every dimension at once, and the narrowing is multiplicative rather than additive. A persona specific enough to be memorable may correspond to a vanishingly small fraction of the population it came from, and nothing in the method requires that fraction to be calculated.
The second concerns falsification. A persona cannot be adequately verified or refuted, and so has no demonstrable validity. It is a narrative about somebody who does not exist, assembled by a process that does not preserve the link back to the data. A disagreement about what the persona would do has no empirical resolution; it is settled by whoever argues more persuasively or outranks the others in the room.
Human factors reserves the word model for an artefact with a different obligation: one that states what people can be relied upon to do, within what range, under what conditions, in terms checkable against a measurement. Reach envelopes come from anthropometric distributions with percentiles attached, legibility thresholds from acuity data, predicted task times from timing models with published parameters, error likelihood from a taxonomy applied to a task structure. Each is wrong in a way somebody can demonstrate, which is what makes it useful in an argument.
The distinction is not between rigorous and unrigorous artefacts. A team needs a shared and memorable answer to who the product serves, one that survives contact with people who will never read a research report, and a persona is an efficient instrument for exactly that. The failure mode is substitution: the persona is adopted and then consulted on questions it contains no information about — whether a sequence can be held in working memory, how long a step takes under interruption, what proportion of the intended population can perform the action at all. It answers in the only register it has, which is plausible narrative, and in a meeting that is indistinguishable from an answer grounded in data.
Both artefacts can be held at once. The persona states the audience. A task decomposition states what performing the work requires. Capability data states what the population brings to it. Where a persona is doing all three jobs, the symptom shows early: design discussions turn on what the named character would prefer, and resolve by appeal to a disposition somebody on the team wrote. The argument has gone circular without anyone making an error of reasoning.
Sources
Chapman, C. N., and Milham, R. P. (2006). The personas' new clothes: Methodological and practical arguments against a popular method. Proceedings of the Human Factors and Ergonomics Society 50th Annual Meeting, 634–636. Source of the verification and falsification arguments and the observation that representativeness falls as dimensionality rises.
Chapman, C. N., Love, E., Milham, R. P., ElRif, P., and Alford, J. L. (2008). Quantitative evaluation of personas as information. Proceedings of the Human Factors and Ergonomics Society 52nd Annual Meeting.
Annett, J., and Duncan, K. D. (1967). Task analysis and training design. Occupational Psychology, 41, 211–221.
All sources are paraphrased; nothing above is quoted. The characterisation of capability data is general to the human factors literature rather than drawn from a single cited source.