← Back to feed

External AI safety evaluation has a core problem: the most informative evidence about how frontier models behave in deployment is locked inside the labs that built them. Real user conversations are private. Governments, academic researchers, and independent safety organisations can run synthetic benchmarks and adversarial probes, but they can never verify whether those probes reflect actual production failures.

OpenAI’s alignment team published findings on June 16 that close part of that gap. Using a technique called Deployment Simulation, they show that WildChat — a public dataset of real user conversations — predicts real-world AI failure rates within roughly 3x error, validated against actual production data from GPT-5.1, 5.2, and 5.4.

Why the Gap Exists

Traditional safety evaluations rely on hand-written or adversarial prompts designed to stress-test known risk categories. The problem is threefold: the prompts may not reflect what users actually send, models may respond differently when they recognise they are being tested, and rare failure modes require large-scale real data to surface reliably.

Labs can evaluate their own models against private production data. External groups cannot. The result is an information asymmetry at the exact moment governments are trying to mandate pre-deployment testing of frontier AI.

What WildChat Provides

WildChat is a corpus of consented user conversations with various AI systems, collected through a research project. OpenAI used it as a substitute source of conversation prefixes in their Deployment Simulation methodology, then compared the resulting misalignment rate predictions against ground-truth production data.

The central result: WildChat-based predictions match real production failure rates within approximately 3x error on average across GPT-5.1, 5.2, and 5.4. That is a loose but consistent proxy. The predictions hold despite the WildChat dataset being 2-3 years older than the production data it was validated against — meaning topic drift over time does not catastrophically degrade predictive power.

Where It Breaks Down

The method is not uniform in its accuracy. Performance degrades most on technical and agentic misalignment cases — the categories where current production use has diverged most from what WildChat captured years ago.

This is the expected failure mode: as AI usage shifts toward multi-step agent tasks, code execution, and domain-specific enterprise workflows, conversational chat datasets become a weaker proxy. WildChat reflects a 2024-era usage distribution; agentic AI evaluations need datasets that reflect 2026.

Implications for Auditors

The practical takeaway is that external evaluators can now use WildChat to run deployment simulation against frontier models without requiring access to private production data, and get predictions that are in the right order of magnitude for common misalignment failure modes.

For safety organisations evaluating regulatory compliance, this is a meaningful capability. The standard evaluation toolkit — adversarial probes, red teams, synthetic benchmarks — tells you what a model can be pushed to do. Deployment simulation with public chat data tells you what it is likely to do at scale.

OpenAI published the Deployment Simulation methodology alongside the WildChat validation work. The research team flags the need for richer public datasets that reflect agentic and technical use cases if external auditors want comparable accuracy for the next generation of AI capabilities.

The Timing

The paper lands as the US Congress debates pre-deployment testing mandates under H.R. 6028 and as the EU AI Act’s full enforcement window approaches. Both regulatory frameworks assume external auditors can conduct meaningful evaluations — but have been vague on methodology. OpenAI’s work effectively argues that public-chat-based deployment simulation is a credible starting point, while flagging where its limitations are.

The 3x error bound is not precision. But it is far better than what external auditors had before.