ModelRefs / AI Personalization — AI Glossary

AI Personalization — AI Glossary

Adapting AI model outputs to individual user preferences, history, and context to improve relevance and satisfaction.

Overview

LLM personalization approaches: user profile injection in system prompt, retrieval of user history (RAG over past interactions), few-shot examples from user preferences, and fine-tuning on user data (with privacy constraints). Applications: personalized email drafting, recommendation explanations, tutoring systems adapting to student level.

Reference details

Topic	applications
Last reviewed	2026-06-24

How it works

A representation of the user is assembled — stated preferences, derived traits, interaction history, current context — and the design question is which parts reach the model on which request.
Profile injection places a compact summary in the system prompt, so it applies to every request. It is simple and it is paid for on every call whether or not it is relevant.
History retrieval instead searches past interactions for items relevant to the current request and injects only those, so cost follows relevance rather than being constant.
Preference examples work as few-shot demonstrations of how this user likes things done, which steers style more reliably than describing the style in words.
Fine-tuning on a user's data gives the strongest adaptation and the worst operational properties: a model artefact per user to store, evaluate and retrain, plus a data-deletion obligation that weights make hard to honour.
The failure mode is a feedback loop. Personalising on observed behaviour narrows what is shown, which narrows what behaviour is observed, so without deliberate exploration the system converges on a caricature of the user.

Example: Why injection and retrieval differ in cost, not in capability

A 500-token profile injected on every request across 1,000,000 requests a month costs 1,000,000 x 500 = 500,000,000 input tokens. At 3 units per million that is 500 x 3 = 1,500 units a month, spent before the user's actual question is processed. Now suppose only 20% of requests are ones where the profile changes the answer. Then 1,000,000 x 0.8 = 800,000 requests carry it for nothing, which is 800,000 x 500 = 400,000,000 tokens, or 400 x 3 = 1,200 units of the 1,500 — 80% waste, exactly tracking the relevance rate. Retrieval instead pays on the 200,000 relevant requests: fetching three history items at 150 tokens each is 200,000 x 450 = 90,000,000 tokens, or 90 x 3 = 270 units, plus the retrieval infrastructure. That is 1,500 against 270, a 5.6x difference arising entirely from paying for relevance rather than for presence. The crossover is the relevance rate: injection wins when nearly every request needs the profile, because then retrieval adds a lookup for something you would have sent anyway.

Commonly confused with

Personalisation is not memory, although memory is one way to implement it. Memory is a store of what happened — facts, past turns, stated preferences — and its job is accurate recall. Personalisation is an adaptation of behaviour, which may use memory and may instead use a profile, a segment or a fine-tune. The distinction matters for the failure modes: memory fails by forgetting or by recalling something the user wanted forgotten, while personalisation fails by over-fitting to past behaviour and narrowing what the user is offered.

When to use it

Reach for it when:

Repeated interaction with the same user, where there is enough history for adaptation to be grounded.
Style and format preferences, where few-shot examples of the user's own material work better than describing the preference.
Retrieval over history when only some requests need it, which is the usual case and where the cost difference is largest.
With deliberate exploration budgeted in, so the feedback loop does not narrow the experience.

Reach for something else when:

Cold users, where there is nothing to personalise on and a sensible default is better than an inferred one.
Profile injection on every request when relevance is sparse, which is where the cost is almost entirely waste.
Fine-tuning per user without a deletion story, since a right-to-erasure request against weights is a retraining obligation.
Where consistency matters more than fit — compliance, safety, pricing — and output varying by user is itself the defect.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Personalization — AI Glossary.

Frequently asked questions

What is AI Personalization?