AI and problems of scale — Benedict Evans

🌈 Abstract

The article discusses the challenges and implications of scaling up various technologies, particularly in the context of law enforcement and data privacy. It explores how the increased scale and automation of certain capabilities, such as surveillance, data analysis, and content generation, can fundamentally change the nature of these technologies and raise new ethical and legal concerns.

🙋 Q&A

[01] AI and Problems of Scale

1. What is the key point made about the story from Georges Simenon's detective novel? The story illustrates how the police can listen in on a single phone call with a warrant, but the ability to listen to millions of calls at scale raises different concerns.

2. What examples are given of how increased scale and automation can change the nature of certain technologies?

• Facial recognition and license plate scanning at scale, beyond individual "Wanted" posters
• Automated generation of fake images and voices, beyond individual Photoshop edits
• Automated cheating on homework and exams, beyond individual instances

3. How does the article suggest our reactions to these changes in scale may be influenced by factors beyond just the technology itself? The article suggests our reactions are influenced by perception, culture, and politics, rather than just the technology. It gives the example of different attitudes towards national ID cards in the UK, France, and the US.

4. What point is made about how these technologies can create new ways for things to "break" or go wrong at scale? The article notes that increased automation and scale can create new ways for systems to malfunction or produce biased or harmful outputs, similar to how databases have caused problems at scale in the past.

[02] Generative AI and Scaling

1. What example is given of how generative AI can enable new forms of abuse at scale? The article discusses how high school students can use AI image generators to create fake nude images of large numbers of their classmates.

2. How does the article suggest our reactions to these new capabilities may evolve over time? The article suggests our reactions may change as we become more familiar with the new capabilities, and we may decide they are a new and unacceptable development, or just a new expression of an old issue we don't worry about as much.

3. What point is made about the role of regulation in addressing these challenges? The article notes that you can't regulate away all bugs or mistakes, and that part of the solution is ensuring people understand the limitations and potential errors of these automated systems.