Stanford Graduate School of BusinessContext of AI Model Deployment Matters: Joshua Blumenstock, U.C. Berkeley School of Information
Stanford Graduate School of Business
Start with the problem 0:02
Model performance is only one part of whether an AI system works in government. The first step is to start with the policy problem, not the technology. In Togo, that meant helping Minister Lawson allocate cash to about 150,000 people out of 6 million, using a mix of models built for that one task.
Costs change the answer 1:30
The best model is not always the best policy. Once you count the real costs of training and data collection, the recommendation can change. A proxy means test can be accurate, but at about $4 per household it may lose to a cheaper model that uses mobile phone or satellite data.
Context shapes impact 3:00
Results in the lab can differ from results in the field. A mobile and satellite model may find poor households best, yet fail to cover people without phones. The same is true for generative AI in education, where tutoring can hurt without guard rails but not when basic structure is added.
Build for deployment 4:31
The evidence base in low-income countries is still thin, and studies point in different directions on who benefits most from AI. For many low-income countries, the main question is not slowing frontier AI. It is getting useful AI into the hands of people and institutions through better electrification, connectivity, local language models, incentives, and training.
Evaluate the full stack 8:00
AI should not be treated like a fixed intervention. It is a stack of choices that includes the model, the way it is tuned, and how it is used. That is why end-to-end evaluations matter, with downstream surveys and outcomes, not just engagement or subjective ratings.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
