Facade on the Federal Reserve Building in Washington DC
getty
On March 27, 2026, Scott Cunningham, an economist at Baylor University, presented to the Federal Reserve Board of Governors and did something unusual. Rather than lecturing about AI, he used it, live, as they watched, to do real economic research. An AI agent crawled a replication archive, downloaded 305,000 congressional speeches, classified each one by immigration sentiment and returned results that broadly replicated a landmark study. Total cost: eleven dollars.
The talk was called “AI Agents for Research Workers,” and the framework Cunningham built around his demos leads to a genuinely uncomfortable conclusion. The same technology that makes researchers more productive can, if they respond to it the wrong way, make their work worse. If he is right, there is no doubt that this applies to knowledge work of all kinds.
I spoke with Cunningham at length about the talk. What emerged was something more candid than a slide deck can capture, a portrait of a researcher who is simultaneously evangelizing these tools and worrying about what they are doing to him to him, and to us.
What an AI agent actually is
Most people think of AI as a chatbot. A person types a question, the system types back an answer, the person copy-pastes the result into whatever they were working on. An AI agent operates differently. It reads files, writes code, runs it, hits an error, diagnoses the problem, fixes it and continues, all inside a project directory, without the human typing a line of code. This is huge because the vast majority of knowledge work outputs are fundamentally encoded in digital files and the operations that are the work itself are just modifications to those files. With AI agents, the human role shifts from operator to supervisor. In Cunningham’s framing, the human becomes the principal; the agent is, well, the agent.
This matters because agentic technology changes what work is feasible to automate. A chatbot can help someone write a script. An agent can run an entire analytical pipeline in three programming languages simultaneously and flag where the results diverge.
Cunningham discovered agents last November while deep in a high-stakes project with an immovable deadline. “I had painted myself into a corner,” he told me. The empirical problems were layered and tangled, the kind that ordinarily eat weeks. He fed them to the agent. “I’d never felt anything like it,” he said. In December he wrote a Substack post telling his readers, mostly quantitative researchers in the social sciences, that they hadn’t heard of this tool yet but soon it would be all they heard about. The post was the start of a series on using Claude for research work. His Substack added 6,000 subscribers.
The productivity paradox
Cunningham frames the cognitive output of science (research, analysis, code) as what economists call a production function, a formal model of how inputs combine to produce output. In manufacturing, the inputs might be labor and capital. Here they are human time and machine time. Before AI, these were complements. Producing cognitive work always required some of each, the way a kitchen needs both a cook and an oven. As machine capability rises, the relationship changes. The inputs become substitutes, and the economics push toward a corner solution: all machine, no human.
Here is the problem, according to Cunningham. Human time is more than labor. It is the mechanism through which attention accumulates into human capital, the knowledge and judgment that make a researcher’s work worth reading. Time produces attention, attention produces knowledge, knowledge produces judgment. AI allows that chain to be bypassed entirely, skipping straight to output. If researchers maintain their own investment of time while using the agent, output unambiguously increases. But if they let the machine replace their engagement, what Cunningham calls the “danger zone,” they can end up producing work that is worse than what they would have managed without AI at all.
Cunningham’s productivity curve. The Post-AI curve (red) lies above the Pre-AI curve (blue), meaning AI raises potential output at every level of human engagement. But if researchers reduce their own time investment too sharply — below the threshold a “danger zone” where output (Qd) actually falls below what they would have produced without AI at all (Qpre). The paradox: a productivity-enhancing technology that makes you worse off. Adapted from Cunningham’s presentation to the Federal Reserve Board of Governors, March 2026.
John M. Drake
I find this argument persuasive, and it resonates with what I see in natural science research. A good research paper represents somewhere between 1,000 and 3,000 person-hours of work, sometimes more. Much of that time feels like drudgery in the moment, but it is the process through which a scientist develops the judgment to know what matters in the data and what does not, or is an artifact, or an error. AI can compress the hours. Whether it can substitute for what the hours produce is a different question.
A recent study from Harvard Business School suggests it cannot. Researchers led by Iavor Bojinov gave 78 workers at a financial services firm the task of writing investing articles, with and without AI assistance. AI cut writing time by roughly three-quarters. But workers whose expertise was distant from the task, technology specialists writing about investing, still scored about 13% below the domain experts, even with full AI access. The tool compressed time dramatically. It did not close the knowledge gap.
In our conversation, Cunningham described his own version of this tension. He now reads academic papers by having an agent break them into chunks, summarize each one and generate interactive slide decks. He prefers it to reading the paper itself. But he worries about whether this preference is eroding something. “I just don’t know what I’m doing to myself,” he said. “Am I de-skilling myself?” The question was genuine. At present, he does not have an answer.
What the agents actually did
The Fed talk was built on demonstrations. In one, Cunningham replicated a 2022 PNAS study by Card and colleagues that classified 140 years of congressional immigration speeches. The original paper fine-tuned a RoBERTa language model on 7,626 human-annotated examples, weeks of work by multiple research assistants, if not months. Using an AI agent to orchestrate OpenAI’s batch API, Cunningham’s virtual team classified all 305,000 speeches with zero-shot prompts in 2.6 hours. The big trends held. Partisan polarization since the 1970s, Democrats increasingly pro-immigration, the gap widening over time. Where the labels disagreed, the language model tilted toward “neutral,” a compression bias rather than random noise.
That compression matters, and a working paper by Ludwig, Mullainathan and Rambachan explains why. Language model classification errors behave nothing like classical measurement error. When the errors correlate with the outcomes being studied, and with LLM-generated labels they easily can, downstream statistical estimates can be severely biased. Even 69% agreement between human and machine labels (the accuracy Cunningham achieved in his replication of the congressional speech analysis) is no guarantee of valid inference. A small human-coded validation sample can diagnose and correct the problem, but applying the fix requires understanding the theory underneath the convenience.
In a second demonstration, Cunningham audited a published paper on Brazil’s psychiatric reform using what he calls cross-language replication. The underlying logic is simple and clever. Coding errors in R, Stata and Python are language-specific, so they are plausibly independent. Replicating a pipeline in all three languages and comparing at each step provides a rigorous check: when results match to several decimal places, confidence is high; when they diverge, something is wrong. Cunningham’s agent wrote 96 scripts across three languages in a single session and found real problems: a suspicious variable that mechanically declined over time, large differences between treatment and control groups on nearly half the measured background characteristics, and undocumented duplicates.
Cunningham has built custom tools that, in his words, “aggressively send these agents after my code all the time,” writing referee reports of his own work, stress-testing results, replicating analyses across languages. “The cost of producing high quality code has basically fallen to zero,” he told me. The cross-language audit is a compelling approach to a real problem, a way to generate independent evidence that results are not artifacts of a particular language’s defaults or edge cases. What Cunningham described was not merely faster research but what he called “moving through failure faster,” reaching the point where a fatally flawed project reveals itself in days rather than months.
The new equilibrium
The implications for both academic and industrial research stark. Cunningham reckons that economics has roughly 87 journals, 3,800 publication slots and 39,000 submissions per year from about 12,000 research-active economists. If the average researcher’s output goes from three papers a year to ten, and at roughly $3,200 per year for submission fees and AI tools the cost is less than a single conference trip, submissions could swell to 150,000 or more. The slots stay fixed. The referee pool, voluntary and unpaid and already overstretched, certainly doesn’t expand. Desk rejection rates would need to rise from roughly 50% to 89% just to hold the line. As I have written about previously, many scientists think the peer review system is already melting down.
Cunningham put the equilibrium in blunt terms. Right now, an academic economist goes up for tenure with six papers in six years. “If you had to have six papers, now you’re going to have to have 12,” he said. “Everything will adjust, and who knows what that equilibrium is going to be like.” This semester he is teaching two new courses at Harvard, a PhD probability class and an undergraduate statistics course, while pushing four papers toward submission, consulting for Amazon and writing a weekly Substack. “I really think it would be impossible just on the class alone,” he said. The tool makes this workload feasible. The new equilibrium will make it mandatory.
Cunningham’s numbers are for economics, but the dynamic applies broadly. Any field that relies on voluntary peer review and fixed journal capacity faces the same arithmetic. And the challenge is not just volume. When AI raises the floor on writing quality, formatting and statistical presentation, the superficial cues that editors have long used to triage manuscripts will stop working. Distinguishing serious science from polished but hollow output will require deeper and more costly evaluation at exactly the moment the system can least afford it.
Cunningham’s talk ended with a question the Fed should take seriously, and not only for economics: How do we keep humans on the right part of the curve? The partner model, where the researcher directs and the agent executes, preserves human capital accumulation. The automation model, where the agent does everything, is individually rational and collectively corrosive. The binding constraint in research is shifting from production to evaluation. Our institutions, journals and tenure committees and funding agencies, are built for a world where producing research was the hard part. They will need to be rebuilt for a world where judging it is.