The arrival of any new technology tends to stir widespread fears about automation and the replacement of human workers. The Luddite movement during the early Industrial Revolution is perhaps the most commonly cited example of this technophobia, but similar fears have cropped up periodically since.
But the threat of artificial intelligence seems greater than any previous technological breakthrough: People worry not just about being replaced, but about the possibility of the human race becoming extinct.
Lately, technologists have been asking each other, “What's your p(doom)?” – that is, how likely is it that an AI will destroy humanity? The answer may change from day to day, but the exact number chosen on a scale of 0 to 100 is less important than the fundamental question: “How pessimistic are you about AI?”
In March 2023, more than 1,000 technology researchers and executives signed what became known as the “pause” letter. They called on all AI labs to suspend “the training of AI systems more powerful than GPT-4” for at least six months.
The signatories range from AI pioneer Yoshua Bengio to economist Daron Acemoglu to tech billionaire Elon Musk. This is a group that should not be ignored, but as far as I can tell, they have been. There has been no notable pause, but it is worth noting that as of this writing, no AI systems much more powerful than ChatGPT-4 have been released.
The experts who signed the pause letter were adhering to the precautionary principle, which calls for slowing down and learning more when an activity is likely to cause irreparable harm. But while some experts stress caution, others argue for the rapid development of AI technology. After all, for more than a decade, many economists, including myself, have watched startups bring incredible advances in AI to market. Breakthroughs have included cancer detection, safety systems, tools to assist people with disabilities, private tutors to expand educational opportunities, and self-driving cars to increase mobility and independence for seniors.
A moratorium would only delay such benefits. So while a moratorium may indeed be wise, we need to be sure that our concerns about potential harms are based on more than speculation.
To gauge the validity of concerns about AI, it helps to understand their origins.
The first is historical experience: most new technologies have both benefits and costs, and sometimes the latter outweigh the former. Thomas Midley Jr. invented tetraethyl lead, a gasoline additive that dramatically improved vehicle reliability by reducing engine knock, and freon, a key chemical in refrigeration. But his inventions caused lead poisoning and released chlorofluorocarbons that depleted the ozone layer, so both were eventually banned.
In the 2023 book Power and ProgressAcemoglu and Simon Johnson argued that this result is usually the norm. Economists generally believe that the benefits of most new technologies outweigh their costs, but some believe AI is the latest in a long history of potentially dangerous technologies that were introduced before we fully understood what their costs would be.
The second concern about AI is that it can cause a variety of identifiable harms. There is a conflict over how to train AI systems: The New York Times' It has filed a lawsuit against OpenAI and Microsoft for alleged copyright infringement. Some innovators have made it clear that their goal is to replace workers with AI, while students are already using large-scale language models to complete exams and assignments.
In addition to disrupting education and the workplace, AI is also being used to surveil citizens, mimic individuals through deepfakes, and strengthen military weapon systems. There are also legitimate concerns that incumbents that already hold vast amounts of data will gain market power. These identifiable harms can be weighed against the benefits of the technology and addressed by regulation where necessary, or outright bans, as in the case of leaded oil and chlorofluorocarbons.
This brings us back to the third cause for concern: many of the consequences of new technologies are unintended and unpredictable, and this is likely to be true of AI as well. For example, there are currently widespread concerns that social media is damaging children's mental health. But the point is that, given the ubiquity of these platforms, there are no easy regulatory solutions.
The fact that unintended consequences are often difficult to reverse presents a problem in itself, seemingly forcing a binary choice between using AI or not using it. But this is a false dilemma. In fact, it means we need to do more small-scale experiments with AI to identify potential harms, when we might be able to limit them.
The potential for irreversible unanticipated consequences of new technologies suggests a more nuanced and multifaceted role for AI regulation. We need to be committed to ongoing research to identify consequences as soon as they emerge and conduct appropriate cost-benefit analyses.
Moreover, we need to find ways to make the adoption of AI more reversible, potentially including through institutional and infrastructural reforms, so that society can enjoy the benefits of AI while providing a greater level of insurance against actual damages if AI produces undesirable side effects.
Joshua Gans is Professor of Strategic Management at the Rotman School of Management, University of Toronto. © Project Syndicate 2024www.project-syndicate.org
