Artificial Intelligence

Paper Published on Persistent LLM Hallucinations

Joy Buchanan1 Comment

I am pleased to announce the publication of

“LLM Hallucination of Citations in Economics Persists With Web-Enabled Models”

By Joy Buchanan, Stephen Hill, and Olga Shapoval

ABSTRACT: Despite significant advances since 2023 in large language model (LLM) capabilities and the introduction of web-enabled search functionality, citation hallucination remains a persistent problem in AI-generated academic content. We test GPT-4o, DeepSeek V3, web-enabled GPT-4o, and web-enabled GPT-5.1 across prompts derived from Journal of Economic Literature categories. Using a hierarchical citation validation framework, we find that even an advanced web-enabled model produces non-existent citations. While web-enabled models show improvement over their predecessors, the error rates remain high for scholarly applications. Our findings suggest that the fundamental challenge of citation hallucination persists despite technological advances, with implications for academic integrity and the reliable use of AI in research contexts. We provide an online tool for researchers to run LLMs through the benchmark test using our prompts.

Suggested citation:

Buchanan, J., Hill, S. and Shapoval, O. (2026), LLM Hallucination of Citations in Economics Persists With Web-Enabled Models. Southern Economic J. https://doi.org/10.1002/soej.70069

After some time went by since we published our 2024 paper on this topic and GPT became web-enabled, we thought that this silly problem with hallucinating up fake citations might have gone away. So, we went back to do our count again. We found that a significant number of references are still made up.

I believe that this particular problem of fake citations will go away soon. However, I think the starkness of fake references illustrates something more general about the LLM technology and highlights the importance of verification. Feel free to cite our 2026 paper as evidence that hallucinations are still happening, despite improvements along many margins. Since we started down this track, OpenAI released a paper titled “Why language models hallucinate” at https://openai.com/index/why-language-models-hallucinate/ Excerpt:

Hallucinations persist partly because current evaluation methods set the wrong incentives. While evaluations themselves do not directly cause hallucinations, most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty.

You can access our working paper of “LLM Hallucination of Citations in Economics Persists with Web-Enabled Models” on SSRN at https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5318314

We are not in the slow-lane “LLMs are stochastic parrots” group. Stephen Hill does even more with AI in his personal workflow than I do. AI will probably soon qualify for all of the academic prizes we used to award to human writers and scientists. However, as we hand more responsibility over to AI, keep the issue of hallucinations in mind.

Is AI Taking Our Jobs? Signal Boosting Noah Smith

Joy Buchanan3 Comments

Artificial intelligence is supposed to transform the labor market. Depending on whom you ask, AI is either about to eliminate millions of jobs or usher in a new era of productivity and prosperity. I am reviewing the evidence before I teach an “unemployment” chapter…

I want to give a signal boost to economist Noah Smith, who has been following the emerging evidence about AI and employment at his blog, Noahpinion. His answer, so far, is that the evidence for widespread displacement is weak.

In his September 7 post, “AI keeps stubbornly refusing to take our jobs,” Smith points out a contrast. Many technology executives expect AI to make human workers obsolete, and public opinion surveys from Pew Research Center indicate that Americans share these concerns. Yet employment has held up remarkably well, especially among prime-age workers. AI might still take our jobs soon, but we have not seen widespread displacement yet. Economists have been saying for a long time that automating tasks is not the same thing as eliminating jobs.

In “Automation and New Tasks: How Technology Displaces and Reinstates Labor,” Daron Acemoglu and Pascual Restrepo explain that automation can reduce demand for labor performing particular tasks while also increasing productivity and creating new tasks for workers.

Consider software developers. AI can now write substantial amounts of code. One might conclude that fewer programmers will be needed. But programmers also design systems and decide what software should do. Making code less expensive could expand the amount of software that firms want to produce. Greater productivity may increase rather than decrease the demand for skilled programmers. (my old paper on who wants these jobs is “Willingness to be Paid: Who Trains for Tech Jobs?“)

Smith points readers to Guy Berger’s analysis of software developer employment and reporting on firms that have reconsidered replacing programmers with AI. These examples challenge the assumption that better coding tools necessarily mean fewer coding jobs.

One piece of evidence comes from the U.S. Census Bureau’s Business Trends and Outlook Survey, highlighted by Alex Tabarrok at Marginal Revolution and discussed by Smith. Among firms using AI, 44 percent reported that it supplemented or enhanced existing employee tasks. Ten percent reported that AI performed tasks previously done by employees, while 11 percent reported that AI introduced tasks nobody had previously performed. Most firms reported no employment change attributable to AI.

Smith also highlights research by Ara Kharazian, Jonathon Simon, and David Stevens who find that firms adopting AI tend to expand their employment, including entry-level employment. These findings contrast with the popular narrative that businesses are replacing junior employees with chatbots.

What about young workers?

Continue reading →

Prices as a Map for AI Agents

Joy BuchananLeave a comment

New from me at EconLog:

Prices as a Map for Minds That Can’t Know Everything

More at the link above. The summary is:

The striking possibility is that increasingly intelligent AI may not make Hayek obsolete. It may give us a new reason to appreciate him. Intelligence does not abolish the economic problem created by dispersed knowledge. A world filled with capable artificial minds may actually contain more decision-makers, more specialization, more private information, and more discoveries that no central intelligence anticipated.

For those minds, as for ours, prices would be a map of a world too complicated for any one mind to know.

Who owns model risk?

Joy BuchananLeave a comment

Update to readers after doing the CPE (read Internal Audit of AI Agents and Risk for background )

It’s not every day that you get to present to 36 genuine internal auditors for an hour. Not the rowdiest crowd, true to form, but they did answer my poll questions. Many of them work at a regional bank here in Birmingham, AL. Here is the result of an (unscientific) poll from the session:

Opinion Poll: Who should own the risk appetite built into an AI model?

Claude has a better understanding of internal audit procedures than I do and actually helped me come up with the question. The interpretation of the graph is a collaboration between me and Claude, so I will not suppress the em dashes.

The plurality (first line) is the “textbook-correct” instinct — but it’s hollow without capability. Putting ownership on the business unit that deploys the model matches the Three Lines Model cleanly: the first line owns and manages the risks it takes. Good instinct, and worth affirming. The catch is the whole premise of your talk: the deployers usually can’t see the risk preference embedded in their model, let alone measure or set it. So “the business owns it” is right in principle but nominal in practice unless that owner is given the tools to actually recover and govern the appetite. That’s the gap between the poll’s ideal and the room’s reality.

The committee vote (31%) reflects real emerging practice, with a trap. Nearly a third reached for a dedicated AI-governance body — consistent with where NIST’s AI RMF and ISO 42001 point. But a committee can quietly dilute accountability: “everyone owns it” becomes “no one owns it” if it isn’t paired with a clearly accountable first line. Worth naming that risk out loud.

The most important result is the one that isn’t in any single bar: there’s no consensus. If you had asked this room “who owns credit risk?” you’d have gotten a tight, near-unanimous answer. The fact that ownership of a model’s risk appetite scatters across all four choices tells you this accountability is genuinely unsettled in their organizations.

The settled view: credit risk is owned by the first line — the business that originates the exposure. The lending or client-facing unit that decides to extend credit owns the risk of that decision. This is the textbook first-line ownership case, and it’s why credit risk is often the cleanest example used to teach the Three Lines Model.

So, congrats to me and Claude for coming up with a question that split opinion in a room of expert practitioners?

Internal Audit of AI Agents and Risk

Joy Buchanan2 Comments

Members of the Institute of Internal Auditors know that they can “Join the Birmingham IIA on September 23, 2026 for an insightful CPE webinar featuring Dr. Joy Buchanan… “

I am pleased to get a chance to translate Buchanan and Foster (2026) to an industry audience. I have been reading up on audit controls to prepare.

Some help from Claude with the following: Enterprise risk management rests on a simple discipline: an organization decides how much risk it is willing to take in pursuit of its objectives — its risk appetite — sets tolerances around that level, and then works to keep actual decisions inside those limits. Under COSO ERM, the appetite statement and its tolerances are the structure; the ongoing question is one of conformance. It’s part of the IIA’s AI Auditing Framework and the Three Lines Model: management sets and owns the appetite, risk and compliance build guardrails and monitor, and internal audit provides independent assurance that what the organization actually does matches what it said it would tolerate. A systematic gap between the two is a finding.

For human decision-makers, we’ve built machinery to check this — credit policies, delegated authorities, four-eyes review, documented rationale, an auditable paper trail. We know how to reconstruct whether a loan officer’s or portfolio manager’s judgment stayed inside the lines.

Here is where our paper connects. When an LLM makes a risk-and-return decision — approving credit, weighting a portfolio, ranking procurement options — it too has a risk appetite. But almost none of that oversight infrastructure exists for it. We argue that AI agents are already making decisions with economic consequences and an element of risk.

By showing that the softmax mechanism inside an LLM is McFadden’s random utility model, Buchanan and Foster (2026) establish that the model’s choices reveal a genuine utility function — a measurable risk preference. Our portfolio experiment then recovers the parameters: the slope of the indifference curve we report is the model’s risk appetite, quantified.

I know an Internal Auditor. Some of their old functions will probably get automated. But they have new work to do: auditing the AI agents! Our paper is a step toward both measuring and manipulating the risk appetites of AI agents.

Buchanan, J., & Foster, J. (2026). The innate economic preferences of language models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2607.26288

Image by Grok. I don’t sell that mug but EWED does have merch at https://shop.spreadshirt.com/economist-writing-every-day/

We Overrate Books

Joy BuchananLeave a comment

This is per the 2026 discussion of AI “slop” writing.

One of the things I buy at an annual local rummage sale is cheap physical media like books. This year, I picked up a book by a cartoonist who I like and respect. I thought his book would be funny and prescient from the standpoint of the publication date (1995). The book is 250 pages of mostly slop. Humans wrote lots of slop and it got printed by publishers who had a captive audience.

Why did I have such high expectations for a printed book? I think it is because, as of 2026, we are more selective about what we print. A filtering has happened. Many novels printed 100 years ago were junk.

When I think of “books” today, what it really makes me think of is “classics” or the top 0.01% of books.

So, score one point for the slopistas. Human writing was not universally smart or inspiring.

What I hate about slop is seeing it in spaces I used to trust. There was a time when I could log in to LinkedIn and see human writing from people who I had chosen to follow because I like them as people. There was a contract for my attention that is broken with slop.

I sense some push and pull in the algorithm whereby the sites might be suppressing slop, right now, relative to what I was seeing weeks ago. I went to LinkedIn on 7/17/26 to do a slop check and saw none. They might be trying to preserve the lead that James identified earlier this year: The Hot Social Network Is… LinkedIn?

Oddly, one of the worst bot-infested spaces I tread into is Facebook groups about sourdough bread making. I think the space is not important enough for Facebook to police, and the human users are not very sophisticated when it comes to tech. I logged this observation back in January.

Consider this an update to by 2023 post What We Are Learning about Paper Books

Fable on Legibility

Joy Buchanan1 Comment

Claude Fable is Anthropic’s most capable publicly available “Mythos-class” model. It is optimized for long-running autonomous tasks and deep knowledge work. The roll out of this product has been dramatic. Little people like me have access to it for only two weeks, and I doubt I will be able to afford it thereafter. With my window of access, I posed it the following prompt:

“This article indicates that a much smarter model might not be possible because the universe is opaque. Since Tyler Cowen made this statement, AI has helped people make breakthroughs in math and biology. Write this again in 2026 using the latest state of technology. Is the smartest LLM today much smarter than GPT-4? And is the universe legible?” and I copied in my blog post Is the Universe Legible to Intelligence?

Fable replied in 3 pages of text which you can download as a Word doc here.

One line from Fable’s response: “Notice where the wins clustered: mathematics with checkable proofs, protein structures with experimental ground truth, contest problems with known answers. These are the maximally legible domains — places with a fixed target and a way to verify that you hit it.”

The writing is coherent and contains no obvious hallucinations. Is the answer true, and does it tell the whole truth?

Whenever you read something think about who wrote it, and keep in mind that every author/model has a bias and limitations. In my paper with Will Hickman, we found that just reminding people that a paragraph has an author (whether the author is human or AI) increased the demand for fact checking from readers. LLMs will become more persuasive and closer to (but never completely) correct. Keep reading all things with some skepticism whether they are written by scientists, politicians, or AI.

Regardless, there is definitely such a thing as making the known world more legible to AI today. Thus, people are talking about increasing funding for data availability and the possible demise of the “research paper.”

Research papers are more like stories than facts. The demand for stories is not going away, but I definitely cannot predict the future of the write-for-pay scientist.

AI, potentially, could go and get its own new data, instead of waiting for humans to archive it. Thus, the self-improving AI might take us beyond the current models… unless they run up against something that is not legible to intelligence…

The Empire Strikes Back against AI Cheating

Joy Buchanan2 Comments

People are considering whether university evaluations can survive in the AI age. Hollis Robbins wrote on Substack: “How to limit unauthorized AI use in the classroom“

Robbins emphasizes class size and teaching load against the time of an instructor.  An instructor teaching 4 sections with 100 students each is very limited in their ability to monitor and prosecute AI teaching. It’s worse if this instructor is on a temporary contract.

Limited eyes and hands and human attention really are a constraint here, at least for now. Some people see AI tools in the hands of students as the end of education itself.

I have been tweeting my replies to this:

I don’t do remote exams, but I hear about improvements to remote proctoring technology. The arms race is not over.

Technology goes both ways. The phone students were using to cheat are now being marshalled as a “second camera” for remote test proctoring. Instructors are going to largely win this year if they take current technology seriously, for multiple choice and short answer evaluations.

The commercial Respondus program has just added Word extensions. This technology already exists and can run on the students’ laptops.

Right now, a clever student might still be able to shift their carbon-based eyes to a direction where the answer is displayed illicitly. And the instructor’s eyes can only monitor so many eyes. This is all so 2024. This conversation may be over soon. Human students can be placed under the supervision of machine eyes. Right now, we are still dealing with issues of false positives when machines flag students for cheating, but the machines are improving.

I believe that the roads will eventually be dominated by machine drivers and their unblinking eyes. Humans might drive cars for fun in the hinterlands, but it will no longer be considered a serious thing humans to do for work. Monitoring student cheating will become like truck driving. Human eyes are on the way out. We are going to become more cheat-proof than college has ever been before.

As a college professor, that will have implications for my job, although I can imagine a not-completely-negative future. Maybe I could do more fun work with students because the work of proctoring will be handled automatically. I have spent many many hours constructing tests that would be hard to cheat on and watching students take them. I take cheating seriously, and all the faculty at my business school work hard to protect the value of our degree. I predict that this will become a trivial part of teaching within 10 years.

Will students respond with various forms of hacking and deep fakes against such a system? Maybe. So far, in any arms race, Uncle Sam has been winning in the end for a century now.

If there is a will to do so, we could even bring back the research paper by having students work on a monitored computer that does not let them use AI to write. (We could almost do that already, but perhaps the true limiting factor is that, as I like to say, readers are that which is scarce.)

[Credit to my colleagues Art Carden and Anna Leigh Stone who have talked with me about test proctoring this semester.]

Should Practicing Economists Read Tyler’s New Marginalism Book

Joy BuchananLeave a comment

Tyler Cowen’s new (free online) book entitled The Marginal Revolution: Rise and Decline, and the Pending AI Revolution is going to be “interesting,” but should you read it?

Mike Makowsky explained that Academic economists are overcommitted

If you are already struggling to meet your deadlines for referee reports you owe to editors, should you take the time? If you don’t have time to indulge your curiosity about the 18th century and dead thinkers, right in the middle of the semester, should you look at it now or maybe browse it over the summer?

I think it’s worth going straight to the last chapter right now.

“Chapter 4: Why Marginalism Will Dwindle, and What Will Replace It?“

It was written for you and released quickly for this moment. Tyler does not personally have to worry about his job, but you might.

This link will take you straight to an in-browser e-reader https://tylercowen.com/marginal-revolution-generative-book/app/

Or you can download the PDF at https://tylercowen.com/wp-content/uploads/2026/03/TheMarginalRevolution-Tyler_Cowen.pdf

You might face mental resistance to reading this chapter, because you don’t want to hear the message. If that’s you, then it’s especially useful to read this chapter. He’s not correct about everything. Develop your counter argument, to go forth and save marginalism. You can only do that if you understand and name the threats. This is more about methods/professions and less about ideology than you might think from the title.

Here are some quotes that stood out to me

The ties of empirical work in economics to economic theory are evolving, and in particular the explicit ties to intuitive microeconomic reasoning, and marginalist thinking, are being cut. In much of traditional econometrics, the emphasis is on testing pre-existing models…

in machine learning, we let the algorithm build the “theory” for us, noting it may have tens of millions of variables and thus not count as a theory…

So much for prediction, what about hypothesis generation? Well, there is a new approach to that too, using machine learning.

A lot of economists do not regularly describe what they actually do for work. Yes, we are saving the world by writing papers, but what exactly do you do? Do you generate hypotheses? Is that what you are teaching your students to do?

It’s not fun to think of how the econ profession might need to reposition, but we owe it to students. Who better to work on this than tenured professors? 

I think the case for undergraduates students to major in economics is strong. I also think the case for doing 4 years of college is strong for students who want to learn.

Last summer I wrote: Students still need to learn principles

If economics is “more interesting” than hard science, then it might serve to scoop up good thinkers at the undergraduate level and get them doing something more technical than what they would end up doing in a humanities program. When I graduated from college, the fact that most econ student had accidentally learned to code was a benefit to them.

College graduate humans ought to be able to read and pass the Turing Test if they are going to be effective complements to AI.

Economists championing marginalism for students, today, write: For Gen Z, Economics May Be the Key to Success in the New AI World

Let me plug Mike as well for thinking about what research econs do in 2026: The actual AI problem in academic economics “Oh, what shall all the candlemakers do now that the sun has risen?” made me laugh.

What is an AI Skill?

Zachary Bartsch2 Comments

If you’ve been on LinkedIn recently, then you may have seen the chatter about teaching your artificial intelligence to have various skills. I saw one post by a guy who claimed to have created several skills, each representing a tech billionaire.

At first, I thought “I am behind the 8-ball. What is this new thing?”. Obviously I know what the word “skill” is and how people use it, but I had not encountered its use in the context of AI having it. What does it mean for an AI to have a skill? I somewhat dreaded the the work of learning the new skill of teaching my AI skills.

Then I had lunch with a computer scientist and I learned that skills are nothing new.

Continue reading →