Modern conversational AI agents can handle complex, multi-turn tasks, asking clarifying questions and proactively assisting users, yet they often falter in long conversations, losing track of interruptions or drifting into irrelevant responses. Improving them requires continuous training and feedback, but the “gold standard” of live human testing is expensive, slow, and hard to scale. As a more scalable alternative, the research community increasingly relies on user simulators: large language model (LLM) agents instructed to role-play as human users. The difficulty is that these simulators can exhibit a realism gap, showing unrealistic patience or encyclopedic domain knowledge that real users rarely have. The analogy is a flight simulator: the most useful ones reproduce messy reality, including unpredictable weather and rare hazards, rather than an idealised version of flight.
Measuring the realism gap
To close that gap, it first has to be quantified. A study from Google Research, published as ConvApparel (a benchmark dataset and validation framework for user simulators in conversational recommenders), sets out to do exactly that. ConvApparel is a dataset of human-AI conversations designed to expose where today’s simulators diverge from real human behaviour and to provide a path toward simulated testers that researchers can trust.
A dual-agent data-collection protocol
To capture the full range of human reactions, from satisfaction to clear irritation, the work uses a dual-agent collection protocol: participants were randomly paired with either a helpful “good” agent or an intentionally unhelpful “bad” agent, and their conversations were enriched with first-person annotations of satisfaction. Deliberately eliciting negative as well as positive experiences is what later enables counterfactual testing, checking whether a simulator behaves realistically when conditions change rather than merely echoing typical interactions.
A three-pillar validation framework
ConvApparel evaluates simulators along three complementary dimensions instead of simple surface-level imitation. Population-level statistical alignment checks whether a simulator reproduces the aggregate patterns seen in real users. A human-likeness score assesses how convincingly individual simulated turns resemble human ones. Counterfactual validation tests generalisation by examining how a simulator responds to situations it was not trained on, such as an unhelpful agent. Together, these pillars distinguish simulators that genuinely model human behaviour from those that only look right on average.
What the study found
According to the paper, a significant realism gap is present across all the simulators tested, underscoring that current LLM-based simulators are not yet a full substitute for human testing. Within that picture, data-driven simulators, those learned from real conversational data, outperformed a prompted baseline, and the difference was most pronounced in counterfactual validation, where the data-driven approaches adapted more realistically to unfamiliar behaviour. This suggests that grounding simulators in real interaction data, rather than relying on instructions alone, is a promising direction for more trustworthy automated testing.
Limitations and what to watch
The findings come from a single dataset focused on conversational recommenders, so they may not transfer directly to other domains such as customer support, coding assistants, or open-ended chat. A persistent realism gap across all simulators means automated testing should complement, not replace, human evaluation, particularly for high-stakes deployments. Benchmarks can also be gamed: a simulator tuned to score well on one framework may not generalise, which is precisely why counterfactual testing matters. As with most LLM research, results depend on the specific models, prompts, and data used and may shift as those change. The full methodology and results are described in the Google Research write-up and the accompanying paper.
Why it matters
Reliable user simulation could make the training and evaluation of conversational agents far cheaper and faster, but only if the simulators behave like real people, including when those people are impatient or dissatisfied. Work like ConvApparel reframes the problem from “can a model imitate a user” to “can it be measured against human behaviour rigorously enough to be trusted,” which is a prerequisite for using simulators in place of costly human testing. For a related applied perspective on analysing real human-AI conversations, see the guide to customer call sentiment analysis.
