Welcome again to The State of AI, a fresh partnership between the Financial Times and MIT Technology Review. Throughout the following two Mondays, writers from both entities will engage in discussions on different facets of the generative AI revolution that is reshaping global influence.
This week, Richard Waters, FT columnist and prior West Coast editor, converses with MIT Technology Review’s editor at large David Rotman regarding the genuine effects of AI on the employment sector.
Bonus: If you hold a subscription to MIT Technology Review, you can engage with David and Richard, along with MIT Technology Review’s editor in chief, Mat Honan, for an exclusive live discussion on Tuesday, December 9 at 1pm ET concerning this subject. Register to join here.
Richard Waters states:
Any groundbreaking new technology typically experiences uneven uptake, but few have been as inconsistent as generative AI. This unevenness complicates the evaluation of its expected influence on separate enterprises, much less on productivity throughout the broader economy.
On one side, AI coding tools have transformed the tasks of software engineers. Mark Zuckerberg recently anticipated that half of Meta’s coding would be generated by AI within a year. Conversely, most organizations are experiencing minimal or no advantages from their initial investments. A widely referenced study by MIT revealed that up until now, 95% of gen AI initiatives yield no return.
This has fueled the skeptics who argue that—due to its inherently probabilistic nature and tendency to hallucinate—generative AI will never profoundly impact business.
However, for many analysts of technological history, the absence of immediate effects is merely the usual delay related to transformative new technologies. Erik Brynjolfsson, then an MIT assistant professor, initially articulated what he termed the “productivity paradox of IT” in the early 1990s. Even with significant anecdotal proof that technology was altering work methods, it failed to appear in the cumulative data as increased productivity growth. Brynjolfsson concluded that it simply takes time for companies to adjust.
Substantial investment in IT eventually emerged with a remarkable increase in US productivity growth that began in the mid-1990s. However, that surge slowed a decade later, followed by a second period of stagnation.
Regarding AI, organizations need to establish new infrastructures (especially data platforms), redesign fundamental business workflows, and retrain employees before they can anticipate observing outcomes. If a delay effect accounts for the slow-results, there may still be reasons for positivity: Much of the cloud infrastructure necessary to introduce generative AI to a broader business audience is already operational.
The prospects and hurdles are both significant. An executive from a Fortune 500 firm mentions that their organization has conducted an extensive analysis of its analytics usage, concluding that its workforce generally adds little or no value. Eliminating outdated software and substituting that ineffective human labor with AI could yield substantial results. However, as this individual notes, such a transformation would necessitate major modifications to existing procedures and could take years to implement.
There are already some positive early signals. US productivity growth, which had been steady at 1% to 1.5% for over fifteen years, surged to more than 2% last year. It likely reached a similar figure in the initial nine months of this year, although the absence of official data stemming from the recent US government shutdown makes this confirmation impossible.
However, discerning the durability of this resurgence or the extent to which it is linked to AI remains difficult. The impacts of new technologies are rarely experienced in isolation. Instead, the advantages tend to accumulate. AI is benefiting from previous investments in cloud and mobile computing. Similarly, the current AI surge might only serve as a precursor to advancements in sectors that could more broadly affect the economy, such as robotics. While ChatGPT may have captivated public interest, OpenAI’s chatbot is unlikely to have the final say.
David Rotman answers:
This is my favorite topic of discussion these days regarding artificial intelligence. How will AI influence overall economic productivity? Forget the captivating videos, the allure of companionship, and the promise of agents to handle monotonous everyday tasks—the key factor will be whether AI can enhance the economy, which translates to increasing productivity.
However, as you pointed out, it’s challenging to precisely determine how AI is influencing such growth or how it will do so moving forward. Erik Brynjolfsson suggests that, much like other so-called general-purpose technologies, AI will experience a J curve where initially, there is a slow or even negative influence on productivity as companies heavily invest in the technology before ultimately enjoying the benefits. And then comes the boom.
Yet, there is a counterexample challenging the patience-required argument. Productivity growth linked to IT increased in the mid-1990s, but has been lackluster since the mid-2000s. Despite advances in smartphones, social media, and applications like Slack and Uber, digital technologies have contributed minimally to significant economic growth. A notable productivity boost has yet to materialize.
Daron Acemoglu, an economist at MIT and a 2024 Nobel laureate, contends that the productivity benefits from generative AI will be significantly lesser and will take much longer than proponents of AI expect. This is because, despite the remarkable nature of the technology, the domain is too narrowly concentrated on products that are of limited significance to the largest business sectors.
The figure you mention stating that 95% of AI initiatives yield no business advantages is revealing.
Consider manufacturing. It’s undeniable that some variant of AI could provide assistance; envision a factory worker taking a photo of an issue and seeking advice from an AI agent. The challenge is that the major tech corporations creating AI aren’t genuinely motivated to address such routine tasks, and their extensive foundational models, predominantly trained on the internet, are not particularly useful.
It’s easy to attribute the absence of productivity effects from AI thus far to business practices and inadequately trained personnel. Your example of the Fortune 500 executive seems all too common. Yet, it is more beneficial to explore how AI can be trained and optimized to give employees, like nurses, teachers, and those on the production floor, enhanced abilities and make them more effective in their roles.
This distinction is crucial. Some companies that have recently announced significant layoffs have cited AI as the cause. The concern, however, is that it could merely be a temporary cost-cutting measure. As economists like Brynjolfsson and Acemoglu concur, the productivity enhancement from AI will arise when it is utilized to create new job categories and augment worker capabilities, not solely when it is used to eliminate jobs for cost reduction.
Richard Waters replies:
It appears we are both exercising caution, David, so I’ll attempt to conclude on a hopeful note.
Some assessments suggest that a significantly larger proportion of current labor is within the scope of today’s AI. McKinsey estimates 60% (compared to 20% for Acemoglu) and projects annual productivity increases across the economy of up to 3.4%. Furthermore, calculations like these assume automation of current tasks; any novel applications of AI that enhance existing roles would, as you point out, be an added benefit (and not merely from an economic perspective).
Cost reduction often seems to be the first priority with any new technology. However, we are still in the initial phases, and AI is evolving rapidly, so there remains hope.
Additional reading
FT chief economics commentator Martin Wolf has expressed skepticism about whether technological investment boosts productivity but suggests that AI may prove him incorrect. The downside: Job losses and wealth concentration could lead to “techno-feudalism.”
The FT‘s Robert Armstrong contends that the surge in data center investment need not result in a crash. The greatest risk is that debt financing will play too significant a role in the expansion.
Last year, David Rotman wrote for MIT Technology Review about ensuring AI benefits us in enhancing productivity and what adjustments will be necessary.
David also authored this piece on how to effectively measure the influence of basic R&D funding on economic progress, and why it can often exceed expectations.
