FairRec steers AI movie recommendations toward greater demographic stability

The question: what changes when only gender changes?

Large language models can recommend items directly from a user’s interaction history without recommendation-specific training. But changing a protected attribute, while keeping that history fixed, can change the content of their suggestions.

In a preprint submitted on November 6, 2026, Ziwei Wu and Jingrui He of the University of Illinois Urbana-Champaign introduce FairRec to address this problem under restricted model access. Their method needs neither the recommender’s weights nor its full set of next-token scores. It does, however, require candidate log-probabilities: scores indicating how likely the model considers a proposed continuation. A service providing only generated samples would need a different scoring approach.

The fairness target is narrow and explicit: semantic stability under a minimal change to an input attribute. The paper distinguishes this from causal fairness and warns that similar recommendations do not themselves establish equitable exposure or user satisfaction.

Defining a disparity threshold

FairRec represents recommendation lists using embeddings, numerical representations used to compare semantic content. Its control score compares a candidate list with reference recommendations for users who have similar contexts but a different protected attribute. It takes the largest embedding distance across those references.

An offline calibration stage sets a disparity threshold from an upper quantile of reference scores. That threshold remains fixed during decoding, separating the definition of a violation from the strength of the intervention.

The conformal calibration interpretation has conditions. Its usual marginal coverage guarantee requires exchangeable scores and a fixed, independent reference bank. Exchangeability means the calibration and new scores can be treated symmetrically in their joint distribution. Using the same examples both as references and as scored calibration points does not automatically satisfy the guarantee.

Moreover, that coverage concerns uncontrolled outputs from the calibration distribution, not outputs selected adaptively by FairRec. If no suitable cross-group references exist, the score is undefined and deployment requires a documented fallback, such as broadening the neighborhood or abstaining.

Two ways to select the next item

FairRec constructs recommendations item by item. Each selected item becomes part of the context for proposing the next one.

The hard variant removes candidate extensions whose disparity exceeds the threshold and retains high-likelihood feasible partial lists. Successfully returned lists must pass a final threshold check. If no candidate survives, it abstains. Filtering partial lists is a heuristic: it can discard a beginning that might have led to an acceptable complete list.

The soft variant instead subtracts a disparity penalty from each candidate’s log-probability. After completing a recommendation, it checks whether the final list violated the threshold. Violations increase the penalty weight for subsequent queries; nonviolations decrease it, subject to a nonnegative weight. The threshold itself does not change.

This is selection among proposed items, not a way to create an acceptable item when none is proposed. The paper also notes that penalizing raw disparity is a heuristic surrogate: lowering disparity does not necessarily lower violation frequency under every candidate distribution.

What the experiments found

The authors evaluated MovieLens-1M and Amazon Movies & TV with Qwen3-8B and Llama3.1-8B-Instruct. They used a 70/30 calibration/test split and gender as the protected attribute. Their embedding model was fine-tuned on recommendation-specific data.

The evaluation’s CFR metric measures the mean embedding shift when the protected attribute changes while context stays fixed. This differs from the cross-group reference score used for control. Lower CFR indicates greater measured semantic stability.

Both FairRec variants recorded lower CFR than direct zero-shot recommendation, UP5 and FACTER in all four dataset–model settings. Soft FairRec had the lowest CFR in three settings. For MovieLens with Llama3.1-8B-Instruct, its CFR was 0.243, compared with 0.701 for zero-shot recommendation and 0.440 for FACTER.

Recommendation quality was assessed using HR@10, a hit-rate measure, and NDCG@10, a ranking measure that accounts for positions within the first ten results. Soft FairRec exceeded FACTER on both metrics in all four settings, but its NDCG@10 remained below zero-shot recommendation throughout. Hard FairRec achieved the lowest CFR on MovieLens with Qwen3-8B, alongside lower ranking scores than the soft variant.

A conditional guarantee, not demonstrated rate control

The theoretical result bounds the soft decoder’s empirical violation rate above its target by a term proportional to 1/T, where T is the number of queries. It requires uniform enforceability: a finite penalty level must exist above which every query produces no violation.

The authors call this a strong assumption and do not verify it experimentally. The experiments also do not measure the realized online violation rate. Thus, the reported results support reduced semantic disparity on these benchmarks, not an experimentally established violation-rate guarantee or fairness across all relevant dimensions.