Monthly Archives: September 2026

two futures for AI, work, and learning

Posted on by
Reply

AI tools can be designed to accomplish tasks that have been done so far by human workers. Engineers use rubrics to assess and refine the outputs of AI. Their rubrics are often based on human performance. For example, how well does as an LLM perform at writing a contract or allocating investments, compared to a lawyer or an investment advisor? The next version will do it better.

On the other hand, AI tools can (or could) be designed to accomplish tasks that human beings cannot do or truly don’t want to do.

This is a choice. The consequences could be profound.

If AI is designed to replace human work, then we risk a scenario in which many people may get laid off, aggregate growth suffers because those people have less money to spend, the job losses are concentrated among knowledge workers, the benefits of education (broadly defined) shrink, and people lose commitment to developing their own minds.

We could see not only an economic recession but a mental one. For example, a college degree would be worth less on the market, and students would not see the point of challenging themselves to learn. Intelligence is a good, and the market value of human intelligence could slip if there is also a growing supply of artificial intelligence.

This scenario would be bad for business, not only for workers and learners. A consumer-oriented company that replaced all its salespeople, managers, accountants, and so on with machines would also contribute to reducing its own consumer base. That would be a classic Tragedy of the Commons. The only companies worse off than those that cut their own employees would be those that didn’t. Labor’s share of income would shrink, but capital’s slice would also be smaller than it is today.

We are familiar with the story that new technologies replace some workers but also boost productivity, which creates new jobs. The technologies of the Industrial Revolution tended to replace higher-skilled artisans, forcing many workers down the skill ladder. Those people suffered, yet today’s workforce is much better educated and better paid. If AI turns out like steam engines and machine looms, we’ll be OK in the longer run.

Then again, we have never had a technology so relentlessly tunable to replace any worker, even someone in a brand-new occupation. Besides, the tools of the Industrial Revolution displaced skilled artisans but not people who had formal educations, whereas AI seems to threaten the most advanced occupations, even pure mathematicians. That means that there may be no way to escape by learning higher skills.

Reading and math scores fell rapidly all across the world between 2018 and 2022, and then further in 2025:

As reported by Elizabeth Grenier in DW.

The PISA report shows that this decline is correlated with AI use. In 2025, it says, “AI chatbots (such as ChatGPT) were used for schoolwork by a majority of students in most participating countries and economies.” But there was variation in how much it was used, and where. “PISA shows that students who do not use AI chatbots for schoolwork generally score higher than students who do, and are more likely to report higher levels of help-seeking behaviours.”(OECD 2026 p. 236)

Specifically, students who never used AI to summarize an assigned text or to draft text for writing assignments scored much higher than those who ever used it for those purposes, whether or not their scores were adjusted for socioeconomic status (fig 1.4.13).

This evidence does not prove causality, but it’s consistent with the concern that AI tools can replace rewarding tasks that human beings have done (such as reading) and even cause a global mental recession. Inventions like railroads and typewriters never lowered reading skills.

It is also worth exploring positive scenarios, in which AI tools are trained and refined by rubrics that ask whether they are accomplishing something valued by humans that humans cannot do or do not want to do. Then we could see economic growth, profits for companies that solve new problems, and positive externalities, such as better health and more learning. The PISA report adds, “among students who use AI for schoolwork for a general purpose (‘to help me learn’), those who report moderate use score slightly higher in PISA than students who use it either rarely or intensively.” I have no doubt that AI tools can help us learn instead of replacing our learning, but this is a matter of design.

I worked with Tufts engineering students to develop a rubric to assess tools like the Civic Helpdesk, asking whether these tools provided useful advice for voluntary groups. To some extent, we imagined experienced community organizers as our models. Thus our tool could replace people, although paid organizers have always been critically scarce. But the Civic Helpdesk can also draft boring documents so that voluntary groups can concentrate on meatier topics. That is an example of freeing people to do more valuable work..

Zeke Emanuel and Vinod Khosla argue that AI tools can perform the algorithmic aspects of medicine (applying “detailed guidelines from professional societies”). In such cases, AI tools would be designed to replace some work by doctors. However, presumably, the same physicians could then concentrate on aspects of medical care that are not resolved by guidelines.

Earlier, I described a race to replace workers with AI as a Tragedy of the Commons. Companies would hurt themselves as well as others by reducing employment, thus shrinking their consumer base.

Problems of collective action require collective solutions–not always governmental regulation (although that is worth considering) but sometimes voluntary collaboration or pressure from organized people.

The current resistance to new data centers may not be the ideal way to counter to Silicon Valley. When residents block a data center in their own community, it will probably be built somewhere else. But such resistance introduces friction that could lead to genuine negotiation.

At a high level of generality, the goal of workers, consumers, and their elected representative should be to press AI companies to refine their tools to perform tasks that people cannot do or truly don’t want to do.

We are very early in the history of widely available AI. It is too soon to tell which path we’re on. Dean Baker sees little evidence of jobs lost to AI yet. He notes that jobs in the insurance industry should be at risk, yet the number of insurance positions shrank by just 2.5 percent in a year, versus an 8.6% decline in manufacturing, which should be less exposed to AI.

On the other hand, the share of GDP that workers capture has fallen since 2000–and sharply since 2020. Brent Neiman writes, “While we can debate about [the causes] over the past five decades, when it comes to the last five quarters, technological change has to be a leading contender in explaining its rapid decline.” He implies that AI already “substitutes for labor”:

The Financial Times’ John Burn-Murdoch shows that the college wage premium (the extra amount of salary that comes with a college degree) has suddenly fallen since 2020 in both the USA and the UK.

This trend could be good news if it meant that people without college degrees were being paid better than in the past. In an equitable world, the college premium would be small. AI could even help make the world more equitable by, for example, helping employers to find fully qualified prospective workers who don’t have expensive college degrees. (See “How AI can advance a skills-first labor market” by Opportunity at Work.)

However, since labor–in the aggregate–has been capturing a smaller share of GDP since 2020, it seems more likely that college graduates are being squeezed by technology. Meanwhile, hundreds of millions of younger children are simply learning less.

Source: OECD (2026), PISA 2025 Results (Volume I): Future-Ready Students, PISA, OECD Publishing, Paris,See also: AI as the road to socialism?; attitudes about AI by age; why the humanities could never be automated; and other posts on AI.

Wendell Berry, September 2

Posted on by
Reply

On September 2, I posted Wendell Berry’s poem with that title on social media, moved that he had just missed the anniversary of the day when, according to the poem, he accepted “decline” and recognized that he would dissolve into “sleep.” The original appeared in Poetry magazine, May 1970. Berry died on August 31 of this year.

I have since spend a bit more time with the poem, and I will offer a few notes in the hope that they may encourage you to give it additional attention. (Berry always invited us to attend.)

It is written in iambic pentameter. There are rhymes: “hawks” and “stalks” get the poem going, and later, “knew” rhymes with “grew”; “as in” with “garden”; and–more loosely–“darkening” with “fading.” I hear the last word of the poem, “flight,” as a rhyme of “light,” although that word came at a caesura in line 10.

Neither the rhythm nor the rhyme scheme is perfectly regular. Nor are rows of sunflowers. But the regularity is sufficient to connect this poem to the tradition of rhymed English iambic pentameter that Chaucer began in the 14th Century. Running to twelve lines, it can be read as variant of a sonnet.

Early September would likely be hot in a place like Kentucky. The poem is set in the evening, when nighthawks hunt for insects, and the “landscapes of the sky” are fading at dusk. These features connect “September 2” to a long tradition of rustic elegy. As one of countless examples, Virgil’s second “Eclogue”:

The shadows lengthen as the sun goes low;
Cool breezes now the raging heats remove;
Ah! cruel heaven, that made no cure for love!
I wish for balmy sleep, but wish in vain ...
           (translated by John Dryden, 1697)

The nighthawks, sunflowers, and garden pull powerfully upward, as suggested by the words “rose,” “tall,” and “straining to the light.” The nighthawks are even “summoned” up. But the final two lines work like the “turn” or concluding couplet in a Shakespearean sonnet. They suddenly describe the narrator’s downward motion.

Berry was, famously, an actual farmer. But the garden in this poem is in his mind. Acceptance grows in his garden, not planned or planted there but naturally arising. He has worked in his mental garden; now he is ready to sleep.

Like most poems, this one should be read as fiction. But it is hard to ignore the fact that Berry had 56 years ahead of him to describe restraint and acceptance after he wrote these lines.

The last phrase is hard: “my leaves all dissolved in flight.” It echoes the sentence that likens the nighthawks to rags tossed in the sky. The birds rise into sight, while the narrator’s leaves dissolve, flying, out of view. But what are they–him? Are they about to fly from their stems when autumn comes?

If his leaves are his writings, I do not think they will soon dissolve.

See also: read slowly, read aloud; notes on Auden’s September 1, 1939; and the humanities for specialists and for everyone (on attention).

navigating the social sciences

Posted on by
Reply

Here is a tentative effort to map the social sciences for a novice, such as a student who’s thinking of taking some social science courses. (It is a counterpart to my recent post on the humanities.) I will risk over-generalizations to offer some pointers.

I would start with three familiar cases:

  1. Researchers want to understand and, if possible, predict a situation in which many individuals face a similar choice that is concrete and observable. For example, people can buy a specific good at a given price, apply for a job, pursue a particular kind of education, marry, or vote. Although individuals’ motives and beliefs vary, there is a typical reason that one would do each of these things (e.g., one buys a good to possess it) as well as costs, disadvantages, and competitors. Thus it should be possible to model the whole situation with an equation that shows the outcome on one side of the equal sign. To test the model, we want to collect as much data as possible, and the data tend to be concrete and unambiguous. It may not be necessary to work with samples if data can represent all the relevant behavior, as a price reflects all the recent purchases of a given good or an election tallies all the votes. Economics relies heavily (but not exclusively) or large-scale data that markets generate or that bureaucracies collect for planning purposes. The mathematics that’s most useful includes statistics (especially statistical tools for inferring causality) and calculus, which reveals maximum values subject to constraints.
  2. Researchers want to understand a mental characteristic that is not measurable with a single datapoint. For example, they want to understand anxiety, hand-eye coordination, prejudice, or knowledge of the American Civil War. They will probably study specific human subjects, rather than use data to describe populations, and they will devote attention and ingenuity to deriving measures of unobserved, latent, and/or unconscious characteristics. This is the point of a laboratory test of perception, a multi-item survey that measures an attitude, an in-depth interview, or an exam, which might include graded essay questions along with multiple-choice items. Again, statistics is a useful tool, but now often the main statistical question is whether an individual differs from the mean of a relatively small sample that shows a lot of variation. Qualitative research becomes more prominent. And a different kind of math emerges as useful: factor analysis and related methods that turn many datapoints into scales.
  3. Researchers want to interpret a whole human situation: a cockfight in Bali, adolescence in Samoa, a ceremonial feast in the Pacific Northwest, or indeed, a US high school or a physics lab or a Manhattan apartment building with a doorman. The key questions are what happens and what that means. Often the important happenings (cockfights, gifts, rites of passage, holiday tips) recur regularly and may not seem optional or determined by the individuals’ characteristics, such as their personal desires. Quantitative data become less relevant. Instead, the researcher’s main method is to spend a considerable amount of time in the community, observing intensively and listening to many informants.

The bastions of these approaches are, respectively, economics, psychology, and anthropology. Each discipline arose to study a restricted range of topics and then expanded by applying its favored methods to other topics.

Economists began with markets but have studied topics like voting since at least 1942 (Schumpeter) and probably earlier. Some of the pioneering anthropologists sought to understand what they then called “Man” by conducting ethnographies of people who were presumed to live as early humans had lived. Several assumptions of that approach are now rejected, and ethnography is now applied to all kinds of communities. And psychology emerged from early studies of perception and cognition by people like Wilhelm Wundt (1832-1920) and William James (1842-1910) plus efforts to define and treat mental illnesses by doctors like Sigmund Freud (1856-1939). The discipline now extends to all areas of human and animal life.

As a first approximation, I would say that sociology and political science are not defined by their favored methods; they are methodologically pluralistic. Instead, they are defined by their topics: respectively, society and politics. That said, both make more use of representative surveys of populations than other disciplines do. (Economists would rather work with revealed preferences, and psychologists often intensively study smaller samples.)

Also, economics and psychology are often (not always) methodologically individualist, meaning that they seek to analyze group phenomena as aggregates of individuals’ behavior. But sociology has contributed powerful models and methods that treat other things–such as classes or institutions–as the causes of outcomes. Similarly, political science contributes a focus on power, which can be missing in models that presume that people have psychological traits and choose accordingly.

Unfortunately, I do not read much research in geography, but its unique contribution is spatial analysis. A characteristic tool of geographers is a map. Public health employs many of the methods favored in the other social sciences (including maps), but it addresses diseases and syndromes that affect human beings. “Epidemiology” is derived from the Greek words for “upon the people.” Focusing on non-human factors that are “upon us” requires studying phenomena like contagion and mutation that apply to bacteria and viruses in relation to people.

Disciplines like public policy, communications, and education are methodologically pluralistic and defined by their topics. In fact, some professors of these disciplines hold PhDs in the liberal arts and apply methods that they learned there. However, experts in these fields might rightly emphasize their distinctive approaches to research. For instance, public policy research has a distinctive emphasis on feedback loops (how policies affect society, which then affects policy). And education thinks developmentally.

Finally, history is often classified among both the humanities and the social sciences. A cultural historian who seeks to enrich our perception of human artifacts is very much a humanist. Likewise, a narrative historian who weaves events and intentional action into a nonfiction narrative is humanistic. But it is also possible to employ methods like econometrics and epidemiology to the past, in which case history operates as a social science.

See also: navigating the disciplines (from 2020, with a video); the humanities for specialists and for everyone; against methodological individualism; three cores of contemporary social science (which is worryingly similar to the current post, but written in 2016); is social science too anthropocentric? etc.