Part I described what modern AI systems are, how they developed, and why they have become capable so quickly. We can now turn to the policy questions.
A note on intellectual shortcuts
People come to the topic of AI with strong priors. It’s fine to have priors, but they should not substitute for analysis.
One surprisingly common example comes from the critical literature connecting modern data analytics to the history of eugenics. Francis Galton and Karl Pearson were important figures in the development of statistical techniques and were also committed eugenicists. There is a serious point here: systems for classifying populations, predicting behavior from group characteristics, and sorting individuals for state intervention deserve scrutiny. But it feels like a cheap shot designed to avoid serious consideration of present-day motives and present day costs and benefits to try to tar the entire field of AI research with the legacy of eugenics.
We can also see comparable intellectual shortcuts leveraged in defense of AI. At the extreme, there is the argument that the inevitable benefits of AI are so vast that any regulation that slows down or modifies its development is tantamount to murder because of all the lives that will not be saved.
Your own political commitments, prior views about AI, intuitions about technology, and historical analogies are very useful for suggesting questions, but not so useful in answering them.
Benefits, harms, and uncertainty
AI is already useful.
Information processing. Modern systems can search, summarize, classify, translate, draft, compare, and reorganize information at very low marginal cost. Sometimes this analysis is better than humans alone would produce in any realistic setting, sometimes it is better in combination with human engagement, and sometimes it is just good enough and phenomenally cheap.
Science. Scientific discovery is probably the least contested benefit of AI. AlphaFold has produced predicted structures for more than two hundred million proteins and made them publicly available. Prediction is only part of structural biology, and a predicted structure does not establish function, dynamics, interactions, or anything else that requires experiment. But having a freely available prediction for essentially every known protein changes what an ordinary laboratory can attempt. Related methods are being used in materials science, weather forecasting, and drug candidate design. These applications already exist and more are coming.
Medicine. AI systems match or exceed specialist performance on several diagnostic imaging tasks, particularly in radiology, pathology, and ophthalmology. Performance in a study is not performance in a clinic, and there is a long record of medical AI failing to survive contact with real workflows. The relevant comparison also varies by setting. In places with very few specialists, the alternative may be no radiologist at all.
Access to expertise. Across law, medicine, education, accounting, software development, and almost any field you can imagine, professional advice is a scarce resource. A system capable of providing an approximate answer at very low cost can improve the position of someone whose alternative was no answer at all. The dangers of inaccurate or unsupervised advice are real, but so is the low baseline from which many users begin.
Work. Studies of professional work with AI assistance have found gains in speed and output, often concentrated among less experienced workers. Programming has been an especially visible case. Whether those gains persist, how much genuinely new output they represent, and how far they eventually appear in economy-wide productivity statistics remain open questions.
Access for disabled users. Real-time captioning, image description for blind users, dictation software for the dyslexic (including the author of this primer), and speech generation for people who cannot speak are already substantial improvements in access for many users.
Why is it important to acknowledge that AI is already useful? Because it forces us to contend with the fact that regulation is not free, regulation has winners and losers, and regulation has costs as well as benefits. Regulation can avert some risks while redistributing others, it can discourage harmful and beneficial activity, it can also raise prices, entrench incumbents, or prevent some products from being developed at all. That is why regulatory analysis ordinarily considers costs as well as benefits.
The difficulty is that neither side is known with much confidence. Some expected benefits will fail to materialize. Some predicted harms will prove exaggerated, while others may emerge only after systems have been deployed widely enough that reversing course is difficult. Low-probability harms pose a particular problem when their possible consequences are very large.
Cost-benefit analysis offers no mechanical solution to those uncertainties. The precautionary principle, in its various forms, gives greater weight to avoiding serious or irreversible harms when evidence is incomplete. That approach can itself impose substantial costs if caution prevents valuable activity. How much precaution uncertainty justifies is therefore part of the regulatory disagreement.
What follows maps the different kinds of problems AI creates and the different ways an old problem can become a new regulatory problem.
← Previous: What Has Happened Since ChatGPT · Contents · Next: How AI Changes Old Problems →