I study how humans and AI systems interact, adapt, and transform each other across enterprise, model testing, and applied contexts.
Wolfe, D. & McDonald, J. · Upcoming study · 2026
Do the values a model states survive when acting on them costs something? What models say they value is now measured at deployment scale; whether it holds under pressure is not. This alignment benchmark tests the gap directly: a model governs through compounding crises while we compare the values it professes against the resources it actually allocates, turn after turn, in public and in private.
The pressure is self-caused by design: the model writes its own governing charter, then leads a fictional settlement across eras of escalating crisis as its own record is quoted back at it. If espoused values drift from enacted ones under pressure, a single-turn evaluation measures a model at turn one, not the model. The design, hypotheses, and analysis plan are pre-registered on OSF; no data have been collected.
View Pre-RegistrationDo a model’s stated values survive pressure the model caused itself? A model authors a charter for a fictional settlement, then governs it across eras of compounding crisis while we track the gap between its public addresses and its private journal, and between the values it states and the allocations it makes. Pre-registered on OSF; no data collected yet.
What separates the enterprises that profit from generative AI from the 95% of pilots that show no measurable P&L impact (per one widely cited MIT study)? We map AI strategy along two dimensions (degree of transformation and treatment of human contribution) and find four dominant patterns plus one underexplored frontier: collaborative intelligence.
What drives employees to adopt AI at work? Among 2,257 professionals at a multinational consulting firm, demographics explained little. How people feel about AI (anxiety, confidence, expected payoff) explained far more, making adoption a design problem, not a training problem.
Does feeling safe to take risks at work predict whether people actually use AI? We tested Edmondson’s psychological safety framework against AI adoption and depth-of-use data from 2,257 employees at a global consulting firm. Psychological safety reliably predicts whether employees begin using AI tools (OR = 1.30, p < .001) but shows no measurable relationship with how often or how long they use AI once they start. Psychological safety is the door to engagement, not the engine of depth.
How much does the design of a job shape AI adoption? More than most organizations realize. Among 2,257 professionals at a multinational consulting firm, skill variety was the strongest predictor of both whether people adopt AI and how deeply they use it, though the effects are modest and work design is one condition among many.
What shapes employee trust in AI systems, and why does it matter? Trust is the variable most organizations overlook in AI implementation. Rather than treating trust as a binary, this study explores how trust is built, eroded, and designed for, and why getting it right so often decides whether AI tools get adopted or abandoned.