ConvApparel: Why AI User Simulators Still Suck and How to Fix Them

Google Research just dropped ConvApparel, and honestly, it’s about time someone took a hard look at how we’re testing conversational AI.

We’ve all been there: you’re building a chatbot, you test it against an LLM-powered user simulator, everything looks great. Then you put it in front of real humans and it falls apart. The simulators are too nice, too patient, too knowledgeable. They don’t get bored, they don’t get frustrated, they don’t suddenly decide they want a pizza instead of sushi halfway through the conversation.

That’s the “realism gap” ConvApparel is trying to measure and bridge.

The problem with user simulators

LLMs are trained to be helpful assistants. That’s their whole deal. So when you tell one to roleplay as a confused, impatient, or inconsistent human user, it’s like asking a professional ballerina to play a clumsy oaf. They can try, but it’s not natural.

The result? Simulators that are:

• Too verbose (real users don’t write paragraphs)
• Too patient (real users abandon conversations)
• Too knowledgeable (real users don’t know your API docs)
• Too consistent (real users change their minds)

If you train your agent against these perfect simulators, it’ll fail in the real world. Every time.

Enter ConvApparel

ConvApparel is a new dataset of human-AI conversations, but it’s more than just a collection of chat logs. The team at Google designed a clever dual-agent collection protocol: participants were randomly routed to either a helpful “Good” agent or a deliberately unhelpful “Bad” agent. This captures the full spectrum of human behavior, from satisfaction to profound annoyance.

The validation strategy has three pillars:

1. Population-level statistics (do simulators match human distributions?)
2. Human-likeness scoring (do individual turns feel human?)
3. Counterfactual validation (can simulators adapt to new, unseen agent behaviors?)

The third one is the real innovation here.

Counterfactual validation

Here’s the thing: a simulator that just memorizes training data is useless. You need it to react plausibly to situations it’s never seen before, especially frustrating or confusing ones. That’s what counterfactual validation tests.

The idea is simple: throw a simulator into a conversation with an agent that behaves completely differently from anything in its training data. If the simulator just repeats patterns from training, it’ll fail. If it can genuinely adapt and roleplay a realistic human response, you’ve got something worth using.

This is harder than it sounds. Most current simulators struggle because they’ve been trained on conversations with helpful agents. When they encounter rudeness, confusion, or incompetence, they don’t know how to react.

Why this matters for conversational AI

Conversational Recommender Systems (CRSs) are a big focus here. These are AI agents that help you decide what to watch, eat, or buy. They need to handle complex multi-turn interactions, ask clarifying questions, and deal with users who don’t know what they want.

If your simulator can’t realistically test these scenarios, your agent will ship with blind spots. ConvApparel gives researchers a way to quantify those blind spots and fix them.

The bottom line

ConvApparel isn’t flashy. There’s no new model architecture, no breakthrough in reasoning. But it’s the kind of foundational work that makes everything else possible. Without good evaluation, you’re just guessing.

I’ve been working in this space long enough to know that most evaluation frameworks are either too narrow or too expensive. ConvApparel strikes a good balance: it’s grounded in real human behavior, but it’s designed to scale.

The paper is worth a read if you’re building conversational agents. And if you’re just using simulators without thinking about the realism gap, well, now you know better.