Personalisation and its costs
Tailoring helps when it reduces work, and harms when it narrows what a person can see.
Personalisation adapts content, defaults or pricing to an individual based on data about them. It reduces search effort when the inference is right and the user understands it. It costs when the inference is wrong, when it is invisible, or when it is used to find the price or message that exploits a particular person.
How it shows up in software
Personalisation is the ranked feed, the recommended plan, the onboarding path that branches on role, and the email that names your unfinished task. The useful kind is mostly boring: remembering a setting, defaulting to the right currency, surfacing the document you had open. The costly kind is the ranking nobody on the team can explain.
Using it well
- Personalise to reduce effort first, such as recall of past choices, before personalising to influence choice.
- Show the reason. Because you watched X, with a control to change it, makes the inference auditable by the person it describes.
- Keep a route to the unpersonalised view: chronological, unfiltered, complete. Users need to know what they are not being shown.
- Set an accuracy floor. If your inference is wrong often enough to annoy, ship a question instead of a guess.
Where it turns manipulative
- Personalised pricing based on inferred willingness to pay, or on a proxy for income or device, charges vulnerable people more for the same thing. In the EU this must be disclosed, and elsewhere it remains legally exposed.
- Targeting on inferred vulnerability, such as debt, addiction or grief signals, is the clearest case where a tool for relevance becomes a tool for exploitation.
- Personalisation that is invisible removes the user's ability to know what was hidden from them, which makes every other consent in the product weaker.
Where you have seen it
Netflix row labels
Rows carry explanatory titles such as because you watched, which attributes the recommendation to an action the user remembers.
Instagram and X chronological toggles
A switch back to a time-ordered feed, which gives the user a baseline against which the ranked version can be judged.
Spotify Daylist
A playlist named after the inferred mood and time of day, which makes the inference legible and disposable rather than hidden in a ranking.
What the research says
- Matz, Kosinski, Nave and Stillwell, 2017 (PNAS)Contested
Advertising copy matched to inferred personality traits produced higher clicks and conversions than mismatched copy in large field studies on a social platform.
This is the priming-adjacent end of the literature and has drawn serious methodological critique, including Eckles, Gordon and Johnson on the inference from non-randomised targeting. Effects are small in absolute terms and the political claims made from this line of work go far beyond what it shows.
- Hannak et al., 2014 (Internet Measurement Conference)Well evidenced
A measurement study of major e-commerce and travel sites found price steering and personalised prices on several of them, varying by account history and device.
Measurement of what sites actually did, not a lab study. Practices have changed since, but the method and finding stand.
- Turow et al., 2009 and later surveysMixed evidence
Large majorities of US respondents said they did not want tailored advertising, and most objected more strongly when the data practices behind it were described.
Survey data, subject to the usual gap between stated privacy preference and observed behaviour.
Grades are a judgement about the evidence, not about how useful the idea is. Plenty of contested effects are still worth knowing, as long as you do not cite them as settled.