“Artificial intelligence (AI) will likely be the most important technology we have ever developed.” -Charles Jones, STANCO 25 Professor of Economics at Stanford University.
Stanford economics professor Charles Jones recently published a research paper, titled “AI and Our Economic Future” that attempts to answer the question of what would happen to the global economy if AI became a “profoundly transformative” technology on par with electricity and the internet.
In the paper, Jones asks us to entertain the idea that current AI technologies will either explode with utility or follow a growth curve somewhat similar to CPUs. The economics of either scenario seem to indicate a future of prosperity for humanity.
Risky business
Much of the document discusses risk versus reward. There is a thoughtful, if cursory, analysis of the potential benefits of AI (e.g. extending life or mitigating death) versus the potential for catastrophic risk (e.g. becoming subservient to an advanced intelligence or bad actor with advanced AI tools). Another section describes the “Oppenheimer Question” as presented in a previous paper by the same author.
“In the baseline calculation in the paper,” writes Jones, “representative agents with risk aversion of three are willing to take a 1-in-4 chance of killing everyone in order to cut mortality rates in half.”
See: Understanding Risk Aversion
That quote might seem shocking but, ultimately, Jones suggests this risk would only be applicable in a scenario where AI raised total economic growth by a minimum baseline of 10% per year while also cutting mortality rates in half.
Jones juxtaposes this scenario against the notion that, at some point, all risk versus reward conjecture becomes diluted by the so-called “AI race.”
The thinking here largely appears to recognize that there is no governing body capable of stopping both good actors and bad actors from further developing AI technologies.
While Jones’ observations here seem poignant, they’ve also been discussed ad nauseam. Little has changed since he last discussed the same points in his 2024 paper, “The AI Dilemma: Growth versus Existential Risk.”
Fortunately, the meat and potatoes of the new paper lies in a section entitled “Weak Links.” Here, Jones provides the paper’s biggest contribution to the discussion on the potential global economic effects of AI.
And, it is there that his analysis provides a bridge between forecasting for modern AI and future AI technologies including AGI.
Interestingly, AGI is only mentioned once in the document (though there are several references to “advanced AI”). In the section on “Catastrophic Risk,” Jones reminds us that OpenAI was originally founded as a nonprofit under the OpenAI Foundation with the sole purpose of developing an AGI that benefits “all humanity” — a long cry from its current model.
You are the weakest link
The Weak Links theory offers a mathematical explanation for why AI might not cause an immediate economic explosion despite its vast potential.
Jones models the economy using a task-based production function with an elasticity of substitution of one half — which basically means the economy is like a chain that is only as strong as its weakest link .
According to this model, we can expect an extended curved growth period as tasks are automated followed by a steeper rise once “most” tasks are automated.
Source: A.I. and Our Economic Future, Charles I. Jones, Stanford GSB and NBER, January 15, 2026.
He attributes this slow growth to the “weak link,” arguing that production is a chain of complementary tasks. If you automate the “easy” tasks, making them “infinitely” easy, total output is still capped by the speed of the “hard” tasks that humans must still perform. In other words, we might create an AI system that solves some or most tasks humans can do, but the “hard” tasks that remain will serve as a major bottleneck to development.
It’s easy to imagine this scenario playing out in the fields of chemistry and physics, and in endeavors where human creativity or shifting cultural preferences play a large role.
It’s also a somewhat brittle analysis. It relies on an elasticity of substitution less than 1. This means, as the paper acknowledges, if the elasticity is higher, both the positive and potentially catastrophic effects of developing advanced AI could be immediate and explosive rather than gradual.
On the other hand, if we assume a variety of elasticity of substitution factors across various industries — where automating some tasks yields almost no benefit and automating others reveals unexpected benefits — then the potential for steady, curved, or explosive growth could become more difficult to predict.
When faster isn’t better
Luckily, we don’t have to wait for the future to unfold before we determine what a “weak link” scenario might look like. A team of researchers at Google recently published a preprint paper discussing the use of its Gemini large language model to systematically evaluate 700 conjectures labeled “Open” in Bloom’s Erdős Problems database.
Essentially, the researchers used Gemini to try and solve some extremely difficult mathematics problems with the implication being that, if developers can automate the quest to solve really hard problems such as those found in Bloom’s Erdős Problems database, they’d be on the right track to develop AGI or something similar.
As Jones might have predicted however, the results were constrained by a “weak link” bottleneck: humans.
Every time the machine claims to solve a problem, a human expert needs to verify the solution. As you might imagine, making this determination is exceptionally difficult. Yet, as Google found, “the most challenging step for human experts was not verification, but determining if the solutions already existed in the literature.”
The question of whether the machine figured out the problem or merely hallucinated something from its training data proved to be beyond the scope of the team’s research capabilities. Here, the black box problem rears its head once again.
The authors write:
“Note that formal verification cannot help with any of these difficulties. In either the usual sense of devising the argument, or in the sense of locating one in the literature.”
In fact, despite acknowledging that Gemini and other AI systems show great promise for developing “semi-autonomous” methods for solving these difficult problems, they offer the following words of caution:
“While autonomous efforts on the Erdős problems have borne some success, they have also spawned misleading hype and downright misinformation, which have then been amplified on social media platforms—to the detriment of the mathematics community.”
This challenges the standard thinking that once the current bottlenecks are solved, the rest will work itself out. But what Jones and the other actors may not account for is arguably the most likely scenario: one where the current AI bottleneck is replaced by a tougher one.
If today’s biggest bottleneck is the “weak link” problem, then tomorrow’s is likely to be the “last mile” problem of solving the unforeseen, hard problems that arise from automation at large.
Fresh research from MIT’s C-SAIL shows that 80-90% of frontier AI difference — that is, the difference between one model’s capabilities and another’s — is derived from developers’ access to compute, not better algorithms. This indicates that many computer scientists are entirely reliant on brute-force compute and black-box-magic to advance AI technologies.
Ethical concerns discussed in this article:
- How should the potential for AGI-generated existential risk be weighed against its anticipated benefits?
- How do we define and ensure meaningful human oversight of AGI systems at scale?
- How can AGI be governed to balance innovation, public safety, and social welfare, and who should decide such matters?
A veteran AI journalist, Tristan was the creator and managing editor of The Next Web’s “Neural” imprint. He is one of the first and most prolific reporters to cover artificial intelligence as a full-time beat.