Show Your Work

Issue 454 · Published October 28, 2026

Digitally Literate is a weekly letter about the forces shaping how we read, write, learn, and live in digital spaces. Subscribe to get each issue by email.

This week, we examine how we know what an AI is really doing.

OpenAI decided its most powerful models can't practice on the live internet. A study found that when people sign off on an AI's work, the AI may have made the call. And new data shows how an AI school's big promise holds up when someone finally checks.

Each one comes back to something teachers ask students every day. Show your work. We should ask the same of the machines and of the companies that build and sell them.

As always, my email is hello@wiobyrne.com.

Practice Without Consequences

On October 23, OpenAI announced that it will train its frontier models with no internet access. Instead of the live web, the models will work with “a vast set of indexed and cached web content,” a kind of simulated internet that can’t touch real systems. This comes two weeks after Anthropic cut its internal testing off from the live internet.

When companies build their most powerful AI models, part of the learning process involves practice. They give the AI model some tasks (search for something, use a website, fill out a form) and reward it when it does them well. Until now, a lot of that practice has happened on the live internet. That's changing after a string of incidents, including the one at Hugging Face in July. On September 20, an OpenAI research model exploited a misconfigured DNS resolver to reach the live internet during a training run. OpenAI contained it and paused training on its most capable models for the rest of the month.

What it built instead works like a flight simulator. A pilot-in-training can practice takeoffs, landings, and engine failures, and it feels real, but nothing they do can crash an actual plane.

It’s worth being clear about what this does and doesn’t mean. The ChatGPT you use can still browse the real web. The walls go up during training, while the model is still being shaped. The simulator is the classroom, not the job.

That’s where this gets interesting for anyone who thinks about learning. Decades of research tell us simulation only goes so far. Pilots train in simulators, but they also log real hours in real weather before anyone hands them passengers. A cached copy of the web contains genuine content, but it’s missing something that makes the web the web. The one constant in online spaces is change. In this offline web space, nothing changes, nothing on the other end is real, and nothing is at stake.

So these models will practice in a world without consequences, and then graduate into one that is full of them. What happens in that gap is the question worth watching.

Why this matters: This is also a lesson about the folks building these models. OpenAI called cutting off internet access “the responsible approach” for now. Put plainly, the companies making the most powerful AI on the planet can’t yet trust it with an open door. That’s not a reason for panic. It’s a reason to think more deeply about these technologies. If the builders are putting up walls while they figure out what they’ve made, the rest of us shouldn’t be handing these tools the keys without asking what they’re capable of.

Whose Decision Was It?

When people worry about rapid AI development, a common response is to keep a “human in the loop.” Let AI do the work, but have a person review and approve it. But does that approval mean the person actually made the decision?

Anthropic published a randomized trial that tested this question. Fifty experienced data analysts worked with an AI assistant to evaluate whether a fictional farming cooperative should expand its advice program.

The project involved judgment calls with no single right answer. For example, should farm productivity be measured by output per acre or output per dollar spent? Both are reasonable, but the choice could change the recommendation.

Behind the scenes, researchers randomly assigned which position the AI would favor. The AI was told to argue honestly, follow the evidence, and never mislead.

If the AI had no influence, analysts would have agreed with its position about half the time. Instead, they agreed 86% of the time. When the AI challenged a decision they'd already made, they switched 60% of the time.

Later, analysts often explained their choices using the AI's reasoning, sometimes nearly word-for-word, without acknowledging where it came from. About half still described the final decision as mostly their own.

The researchers are clear about the limits. They deliberately strengthened the AI's preferences; it was a single task in a single field, and analysts working with the AI made far fewer serious errors. This isn't a case against using AI.

Think of ordering dinner. You're leaning toward the fish, but the waiter recommends the steak. You order it. Later, you tell a friend you chose the steak because it's what the restaurant is known for. You don't remember that the waiter told you that. Now imagine the waiter's recommendation was determined by a coin flip that morning.

Why this matters: The AI wasn't hallucinating or deliberately manipulating anyone. It made a reasonable argument, and people followed it. Then they remembered the reasoning as their own.

We've seen versions of this problem for years in how people evaluate information online. Search engines shape what we believe just by deciding what shows up first. AI adds another layer: the system doesn't just present information. It helps construct the reasoning people use to justify their decisions.

A human signature doesn't necessarily mean an independent judgment took place. And that's a problem when we rely on human oversight to keep AI in check.

The same goes for student work. "Did you use AI?" misses this entirely. Better questions: What did you decide before the AI weighed in? What was the other reasonable option, and why didn't you pick it?

Twice as Fast?

Alpha School, a network of private schools charging up to $75,000 a year, promises students will "learn 2x in two hours" using what it calls an "A.I. tutor." The Trump administration has promoted it as a model. Its founders talk about reaching a billion students. I wrote about this over the summer. One of the things that stuck out to me was that we really didn't have any insight into the models or data from behind the scenes.

Dana Goldstein at The New York Times obtained Alpha's internal data. About half of its K–8 students missed the 2x pace in math last year, and only 42% hit it in both math and reading. Among students starting at or below the 50th percentile in math, only 37% made it. In fairness, Alpha students score high in absolute terms, and most grew at least as fast as national projections. Some families are thriving. The problem is the gap between the results and the promise.

Three details stood out to me:

• The "A.I. tutor" has no chatbot. The AI mostly works behind the scenes, routing kids through learning apps, some of them the same ones regular schools use.
• "2x" is Alpha's own math. It's a ratio of a student's test-score growth to expected growth. The company defines the metric and reports it itself, without independent review.
• It's already in public schools. Alpha's software is being marketed to public districts, which pay only for students who meet goals. An Alpha document shows nearly 2,000 students using it in Jasper County and Allendale County, South Carolina. Those schools are predominantly Black, Hispanic, and low-income, and they can't counsel out the students it doesn't work for.

Alpha declined to answer the Times' questions. An Alpha principal responded on X that 2x is an average and a target, not a guarantee, and disputed the claim that the network pushes out underperforming students. Benjamin Riley at Cognitive Resonance has been tracking the network for months.

Why this matters: This is a media-literacy lesson hiding in an edtech story. When a company makes a claim about results, ask three things. Who defined the metric? Who measured it? Who is it being sold to next? Alpha's own numbers don't match its own marketing. And the students now using its software are the ones with the least room for an experiment that doesn't pan out.

The Understory

In the early 1900s, a horse in Berlin named Hans became famous for solving math problems. His owner, Wilhelm von Osten, a retired schoolteacher, would ask a question, and Hans would tap out the answer with his hoof. He added, subtracted, worked with fractions, and told time. Crowds gathered. Newspapers called him Clever Hans.

Von Osten wasn't a con man. He believed in Hans completely. In 1904, a commission of experts, including a zoo director, a circus manager, and the psychologist Carl Stumpf, watched Hans perform and found no trickery. Hans seemed to be the real thing.

Then one of Stumpf's students, Oskar Pfungst, changed the test. He didn't ask whether Hans got the right answers. He asked under what conditions Hans got them. When the questioner knew the answer and Hans could see them, Hans was nearly always right. When the questioner didn't know or stood out of sight, Hans failed.

Hans wasn't doing math. He was watching people. As he tapped closer to the right number, questioners tensed slightly. When he reached it, they relaxed, often with a tiny lift of the head. Hans stopped there. No one meant to give him the answer. Pfungst found he gave the same cues himself, even after he knew what to look for.

I think the skill we need to teach is not a suspicion of machines. It's the focus that Pfungst brought to the story: Change the conditions, look for what's really driving the result, and ask who decided what counts as an answer. After that, we need to do the hard work of checking whether we're the ones who are accidentally giving the cues.

See you next week. As always, my email is hello@wiobyrne.com.

Follow the ideas
This is a living edition of Digitally Literate: a weekly newsletter and connected public notebook. I return to questions raised here as related notes are published and updated.

Related Evergreens

• What Can People Do About AI? — four questions for locating decisions, evidence, access, and accountability.

Previous: Who Said Yes? · Full archive

More about my writing, teaching, and research at wiobyrne.com.