Biomathematicus

Science, Technology, Engineering, Art, Mathematics

Nov 10, 2026 · @Juan B. Gutiérrez

The bottleneck is no longer production

Artificial intelligence (AI) expands the volume of what we can produce. Accountability and verification still run on a fixed and limited budget of human attention. That mismatch is the idea I brought to a panel led by Micaela Parker from ADSA at the AIxED conference, alongside Linda Clark of Brown and Mine Dogucu of Harvard, on how data science and AI programs keep up with practice.

The mismatch is getting worse. Organizations are moving toward putting AI in the lead, and that thins the human layer that does the checking. We get more output and fewer people able to stand behind it.

If that is the problem, then the job of education is clear. We have to build the human capacity to verify, and we have to build it before we hand students a machine that produces on their behalf.

Where I teach

I teach at the University of Texas at San Antonio. It is one of only 21 universities in the country that hold both Carnegie’s top research classification and its Opportunity designation for higher access and higher earnings. Many of our students are the first in their families to attend college, and many are multilingual.

I come to this question from two directions: data science, and the sociocultural foundations of learning. Both shaped ALICE, the adaptive learning system I build. It gives each student a personalized, AI-supported environment in multiple languages, and we are running pilots with support from the National Science Foundation and the National Institutes of Health.

Those pilots show a measurable positive impact on learning.

A tutor is not a tool

The same technology can play two opposite roles in a classroom. As a tutor, AI helps a student build a skill: it asks, adapts and explains, and the student does the work. As a tool, AI does the work, and the student supervises.

Supervising requires the very skills the tool makes it easy to skip. So the order matters: foundational skills come before AI as a tool. Technology should not replace handwriting, and it should not replace the slow work of learning to reason.

We already accept this logic elsewhere. We want our kids to read, but we might not want a ten-year-old reading A Game of Thrones. Reading is released in levels, and the conversation about AI in education needs to pass through release levels too.

From kindergarten through twelfth grade, my default is no AI as a tool. AI can help develop basic skills before we teach students AI techniques, but as Pam Eddinger put it, doing that well takes money, will and policy.

Before students are released to AI as a tool, three ideas need to be in place:

  • Attribution. Who or what produced this result, and can I say so honestly?
  • Trust. When is relying on the machine warranted, and when is it not?
  • Veracity. Is the result true, and how would I check?

Four quick answers

The panel included a lightning round. These were my answers.

  • Should introductory courses ban AI for the first few weeks? Yes as a tool, no as a tutor.
  • Is mathematical and statistical thinking more important now? Yes, because verification is now the scarce resource, and mathematics and statistics are how we verify.
  • Is prompting a skill worth teaching? Not in general. Yes in context, inside a discipline where the student can judge what comes back.
  • Is the entry-level data analyst job disappearing? Yes, and we must fight to keep it. The first rung is where people learn to be accountable for a result. Remove it, and nobody grows into the senior person who can verify.

Who gains

Does AI widen or narrow the gap between students who arrive well prepared and those who do not? It depends on the order in which we release it.

Students who arrive well prepared tend to come from affluent school districts. We already have a system that lifts the top. A well-prepared student can check what the machine produces. An underprepared student cannot, and receives the same confident output with no means to question it. Released as a tool without stages, AI widens the gap. Released first as a tutor, in the student’s own language and at the student’s own level, it can narrow it.

The cautionary tale is the pandemic, which forced digital learning on everyone at once. National test scores had been falling since about 2012, and the pandemic deepened the slide. The latest federal results show nine-year-olds back to pre-pandemic reading levels, while thirteen-year-olds remain below them with no significant improvement.

Those thirteen-year-olds lost in-person teaching in second and third grade, the years when foundations are laid. We see the effects in college now. A push for digital learning too early can erode the basic skills everything else depends on.

What “responsible use” leaves out

Everyone now speaks of the responsible use of AI. In practice the phrase means one of three things: just use it, use it as a study buddy or tutor, or create with it. All three are answers to the question of how to use it.

The questions that go undiscussed are different. What does it cost? To whom? And what can the people affected contest?

There is also a problem of scale. Local problems benefit from local decisions, but AI is a global issue and needs a global discussion. Organizations are moving toward AI in the lead, and no global discussion is under way about whether they should.

The one change I would make

Asked for one change to undergraduate education, I would choose a gated release. First, AI as a tutor. Second, AI as a tool. The gate between them is foundational skill: the student shows they can do the work, and check the work, before the machine does it for them.

AI will keep expanding the volume of production. The human attention available for accountability and verification will stay fixed and limited. Education is the one place where we can make that attention count for more.

Sources

biomathematicus