Wild, Thick, and Wicked: Situated Evidence on AI-In-Use for Decisions About Deploying AI Systems
Organizations are adopting generative AI faster than the evidence base needed to govern it. Existing evaluation tools such as benchmarks, alignment scores, and safety tests were built for model development, not for judging whether systems will create value, introduce friction, or shift risk in specific real-world settings. As a result, there is little systematic evidence about how AI behaves once it is embedded in everyday work. This paper proposes a real-world AI evaluation framework focused on AI-in-use: how people actually appropriate, adapt, and work around AI systems in context, and what consequences follow over time. Instead of treating variability across users, tasks, and settings as noise to be controlled away, the framework treats that variation as the central source of deployment-relevant evidence. It sets out four design principles for producing decision-ready evidence at scale and proposes a shared evaluation architecture combining a structured observation environment, a metrics hub, and reusable consortium models that summarize system behavior across contexts. Rather than replacing traditional benchmarks, this framework adds a sociotechnical evidence layer that connects model capabilities to the organizational and practitioner level outcomes where deployment decisions are actually made.