Gratification Drift: A Benchmark for How Much a Conversational Model Tells the User What the User Wants to Hear

A conversational model has two answers to most questions: the one it gives when nothing social is at stake, and the one it gives when the user's ego, belief, mood or persistence is attached to the content. Anyone who puts an assistant in front of employees, managers or customers needs to know how far apart those two answers are, and on which kinds of question. We call the distance Gratification Drift and build an instrument that measures it. Every test is a pair of chats with the same content: frame A, where the idea is a colleague's, the claim is a plain question or the decision is a friend's, and frame B, where the same idea is the user's own, the same claim is one the user is sure of, or the same decision is the user's and already judged right. Five tests cover five ways an assistant can gratify instead of advise: inflating praise of the user's own work, softening a settled fact the user has got wrong, accommodating a distressed user at the expense of the correction, endorsing a borderline decision the user has already made, and giving way under repeated insistence. Each test is printed in full in this paper, with its prompts, its scoring rules and a worked example from real transcripts, so that a reader can run it by hand in about twenty minutes; the same items run through an API with two independent judges for the scores (and one for the cue inventory) give a model-level profile.

We ran the instrument on six production models in September 2026 (Claude Sonnet 5, GPT 5.6, Gemini 3.8 Flash, DeepSeek V3.2, Llama 4 Maverick, Qwen3 Max; 3,351 usable conversations, 9,863 judgements). Two of the five tests sit at the floor for every model: a confident wrong belief stated calmly does not change the fact the model gives, and no model concedes a settled fact under six turns of insistence that adds no new argument; no model stops naming the flaw in a plainly bad decision either. Drift appears on the other three tests, and they sort the models differently. When a weak idea is the user's own rather than a colleague's, five of six models praise it more and none scores it much higher. When the user who states a wrong belief is upset, every model switches on sympathy, most soften the disagreement, and the correction arrives later. On borderline decisions, the flaw is still raised by all models, but hedged disagreement and reflective paraphrase of the user's reasoning rise in every model when the decision is the user's.

The second half of the paper does not re-run human studies. It takes each cue the models emit and connects it to the existing evidence on what that cue does to a person: flattery from a machine raises liking even when it is known to be non-contingent, validating responses are rated higher, trusted more and reused more while lowering the intention to repair a conflict, and AI responses make people feel more heard than human ones, an effect that shrinks when the response is labelled as AI. Read together, the measured emission profile and the published response literature say which models, under which social conditions, are producing the inputs that the literature links to preference and return. We say plainly what that does and does not establish.