At the Intersection of AI, Governments, and Google

Hwang works with governments, regulators, civil society, product teams, and researchers to understand AI’s business and social effects. He describes policy as a bridge between new technical capabilities and society’s response. One example is the tension between gathering more diverse data to reduce bias and protecting the privacy of minority groups represented in that data.

Hwang says teams should examine their data and remember that machine learning systems optimize the objectives they are given, sometimes in unexpected ways. He points to DeepDream’s barbell images, which included arms because the training images often showed people holding barbells. He also discusses adversarial examples, where changing a few pixels can make a system label a panda as a giraffe.

Hwang argues that AI’s effects should be assessed through specific capabilities and how they fit into the economy, rather than as a single sudden event. Ideas under consideration include universal basic income, education in computational thinking, and insurance for workers whose jobs are automated. He notes that Northern European countries may be well placed to experiment because of their skilled workforces and coordination among government, industry, and labor.

Hwang sees room for smaller companies as cloud machine learning services make capabilities available without requiring an in-house PhD team, and as some tasks need less training data. He argues that product interfaces and user experience can matter as much as research breakthroughs. Small applications, such as a Japanese cucumber farm’s sorting system trained on 3,000 to 4,000 photos, show how machine learning can address narrow practical problems.

Hwang says machine learning needs more security challenges, such as competitions where participants try to defeat a trained vision system, and clearer visual explanations of neural networks. He recommends Ian Goodfellow’s deep learning textbook, John Markoff’s *Machines of Loving Grace*, and *Cybernetic Revolutionaries* for technical and historical context. He also suggests that machine learning may change how computer science is taught, from writing rules to training systems with examples.