AI systems that are generally smarter than humans benefit all of humanity
Amid the excitement surrounding the development of generative artificial intelligence technology there’s a pervasive sense of dread. The emerging consensus seems to be that AI will replace a lot of jobs. Already, many generative AI tools have been built with the goal of replacing entire jobs. See Devin, an agent built by Cognition Labs marketed as “the first AI software engineer,” Julius, your AI data scientist, or OpenAI’s DALL-E and Sora image and video generation models, which are poised to disrupt the creative industries.
Young adults stepping into their careers may be unsure how to prepare for the future. It requires answering the question, “In 10 years, what will I be able to do better than AI?” As one of those young adults, I try to offer some points of optimism.
- Our jobs haven’t been automated away yet!
This is not the first time in history the labor market has encountered a paradigm-shifting new technology. Neither the printing press, the first industrial revolution, nor the agricultural revolution caused widespread unemployment. Many jobs were replaced by technology, but more were created as a result. Employment rates remained steady, wages went up, and poverty rates went down. In some cases, disparity increased, but throughout history, the average person has become wealthier as a result of technological innovation.
The modern era is the result of a long series of such technological revolutions, and today, most people who want a job have one. Unless AI is categorically different, we should expect the same pattern to follow. Of course, the sorts of jobs that are demanded may shift, and the nature of jobs in the age of artificial intelligence may feel different, but there is reason to believe both changes will be positive, too.
- The boring stuff will be automated first.
Modern jobs are mediated by layers upon layers of tools, with humans operating at the highest level of abstraction. Many of these tools have become so commonplace that we forget we are not directly interfacing with the object of our work, but something else that has made that interaction more fluid, efficient, and usually, more pleasant and rewarding.
Consider, for example, the calculator. Before the widespread use of the calculator, much knowledge work consisted of repetitive mathematical calculations done by hand. Today, office workers, engineers, and researchers work with software performing unfathomably large numbers of these calculations per second. Their time is spent manipulating spreadsheets, tweaking formulae, and other sorts of tasks only possible when arithmetic calculations are made trivial.
The automation of more repetitive jobs throughout history has enabled us to work on the sort of tasks that engage our faculties for creativity, multi-step reasoning, and long-term planning. And these sorts of tasks are difficult in a fundamental way for modern AI systems, which have powerful long-term memories and working memories but lack episodic memory — the sort of memory that allows us to learn from experience and build contextual knowledge over time. This makes it hard for AI to engage in the workflow that typifies our jobs — planning out a schedule for the day and following through, working on a report that requires iterative research, writing and editing, etc. For the time being, AI is a tool, and like other tools, should change the nature of our jobs in a desirable way, by automating away the most boring tasks.
- Comparative advantage in the age of artificial general intelligence.
It is possible, of course, that AI is different from tools of the past in a critical way. It is important, now, to distinguish between narrow artificial intelligence, and artificial general intelligence. Narrow AI is very good — superhumanly good — at some tasks, but struggles with many others, and the scope of its abilities is small.
Today’s AI systems are narrow. They can write code around human level, comprehend large amounts of text at absurd speeds, and translate between languages at the level of experts. But they struggle to do math, produce content that is factually accurate, and model cause and effect relationships in the physical world. Narrow AI is not a categorically new innovation. Throughout history, we have built things that are superhuman at some subset of human tasks — we call these tools! Current AI systems are simply another tool to augment human ability.
Throughout history, we have built things that are superhuman at some subset of human tasks — we call these tools! Current AI systems are simply another tool to augment human ability.
Artificial general intelligence (AGI), by contrast, refers to AI that is generally smarter than humans. It has superhuman ability in all domains. In the age of AGI, the answer to the question, “What can I do better than AI” is approximately NOTHING.
Today’s frontier AI labs aim to develop artificial general intelligence. The development of AGI is OpenAI’s stated mission: “to ensure that artificial general intelligence — AI systems that are generally smarter than humans benefit all of humanity.”¹ There is reason to doubt that the AGI project will succeed, but if it does, we cannot draw on historical precedent to extrapolate our role in the new labor economy. There will be no step in the labor abstraction ladder that an artificial intelligence should not be able to fill, higher performing than a human. As daunting as this future sounds, even in this case, there will be demand for human labor.
Economist Noah Smith argues that even in the age of AGI dominance, humans will not struggle to find work. He writes,
“I accept that AI may someday get better than humans at _every conceivable task_. That’s the future I’m imagining. And in that future, I think it’s possible — perhaps even likely — that the vast majority of humans will have good-paying jobs, and that many of those jobs will look pretty similar to the jobs of 2024.”²
He attributes this to comparative advantage, and I find his argument convincing. To illustrate, consider a more familiar case: a smart, but busy, business executive. She communicates clearly and could write her own emails better than anyone else but hires a personal assistant to manage her correspondence anyway. Why does this make sense? The opportunity cost of the executive is very high — responding to an email takes time that could instead be spent negotiating billion-dollar deals. The personal assistant’s comparative advantage is that responding to emails is the way she can contribute the most value, relative to all the other things she can do.
A similar dynamic may play out in the age of AGI. Superhuman intelligences will have superhuman abilities to create value, so their opportunity costs for working on human-level tasks will be greater. If an AI can make $1,000 an hour doing data entry but $10,000 an hour doing scientific research, then it is going to work on scientific research, leaving the data entry job to you or me. Even though the AI may be better at the data entry than you, you still add value by doing it. Put another way, the existence of something better than you at everything does not imply zero demand for your skills.
Final Thoughts
I’ve been thinking about artificial intelligence for some years now, and a lot of that thinking has been dominated by career anxieties. I’ve spent years training to develop a certain skillset and would be devastated if in a few years’ time my knowledge were rendered useless. I now think those anxieties were misplaced. I expect AI to make us more productive and increase the value of human skill sets, regardless of how intelligent AI systems become.
My hope is that by approaching the labor question with cautious optimism, we can open space for more pressing concerns. The most important questions of the AI era will be related to ownership, risk mitigation, and value alignment. We may see enormous value created by AI in the near future. It is critical that we consider new paradigms of ownership that distribute that value equitably and protect those who otherwise would not have a stake in the future of AI. It is these questions, rather than the nature of the labor market, that will determine the impact of AI on human well-being.
Footnotes
^1 https://openai.com/blog/planning-for-agi-and-beyond
^2 https://www.noahpinion.blog/p/plentiful-high-paying-jobs-in-the