Delegating versus offloading

“Are you ‘offloading critical thinking skills’ when you delegate tasks to an intern?”

This is a (bad faith) question about generative AI posed to New York Times columnist Jamelle Bouie on Bluesky, but despite its bad faith beginning, it led me to think about the difference between hiring a human assistant and asking generative AI to perform the same task.

I got to work as a research assistant a few times when I was in grad school. I loved it. My task was actually not unlike what people ask large language models (LLMs) to do. I had to create annotated bibliographies for other researchers to assist them in their projects. This is a role similar to that of an intern and part of my job was in fact doing critical thinking on someone else’s behalf. Were the professors offloading their critical thinking onto me? I would argue that they were not.

When I got these jobs, I was always selected for specific skills I had already demonstrated. For one project, I was chosen because I had been a TA for a class about a related topic. In another, the professor had taught me and worked with me multiple times in the past. In other words, I wasn't a random selection. There was critical thinking involved in choosing me to perform some element of research.

Before each project, I sat down with the professor and we went over the project they were working on. They talked about their goals, what kinds of articles and sources would serve them best, and why they needed an assistant for the project. Then I went to work finding sources.

I wasn’t drawing on a random and broad database, either. I used library databases I have already used and some specialized databases related to their topics. Then I skimmed articles that I found. If the article had enough points in common with the professor's goals and projects, I added it to the bibliography and wrote a paragraph or three summarizing the article. I didn’t copy and paste the abstract but wrote about how I thought the specific articles related to the professor's goals and why I'd added them to the list. In some cases, I thought the article was unlikely to be very useful but it had enough in common with the professor’s goals that I included in anyway for them to judge, and I could explain that in my bibliography.

All of that requires a relationship. I needed to know what the professor was looking for and why. If I had questions, I could ask them. If they didn't like my choices, they could talk to me.

While I think people absolutely can offload critical thinking onto other people, mentoring and training someone as a researcher isn't the same thing as offloading your brain. Academic work that seems menial is part of how we get to academic work that is essential and you have to learn how to perform that rather thankless, often frustrating work in order to get to the exciting products that change how we think. Researchers who hire an assistant to help need to be on top of that assistant’s output in order to make sure it’s the kind of output they need, and in working with the assistant, they can serve as one aspect of that researcher’s training.

A human assistant is better at providing what a scholar needs because of this relationship even before we get to hallucinations. I'm slower than an LLM, but I'm far better at accurately reporting real existing articles that I found. I’ve seen many outputs from LLMs that report fake articles, fake quotes, and false evidence confidently stated as fact. If you have an assistant you trust (because you’ve paid attention to their work and helped them with it along the way), you can generally assume that what they’re providing you is, at the very least, real.

While I was working as a research assistant, I was also working on my own research for my dissertation. And, importantly, in my own research, what was useful for my project was not always obvious. I cannot tell you how many times I read a book or an article and thought it was interesting but not relevant to my question, only to find that it actually solved a problem that emerged when I was writing. I hadn’t seen the problem in advance because writing is a different kind of thinking than reading, note-taking, or oral discussion. If I just asked an algorithm to find articles about my topic, I wasn't going to find everything I needed to find. Research builds on research. You find something; it sends you down a new path. You find something that doesn’t seem relevant, but in reading the article you still learned the material and now it’s in your brain when you do need it.

Human relationships are a huge part of how and why we learn. When we talk about academic research as a "conversation," yes, that's often a metaphor, but it is also just literally how academic work happens. We talk to each other in person and in writing. Again, I think it's possible to offload critical thinking tasks onto another person, but assigning a task to a researcher is not inherently that.

However, an LLM is not a person. It cannot reason. It does not have a relationship with you, even if you feel like it does. It does not “understand” anything. That means it doesn't understand your task or what you aim to do. It doesn't understand which sources are already involved in the conversation you're entering. It doesn't understand which of those sources are important to your project. If it goes off base, you can enter new inputs to try to steer it back, but it does. not. understand. Which is why, when you tell an LLM that the last source it gave you does not exist, it will often apologize and then offer another fake source. It's more than predictive text, but not a lot more.

When we say they're not search engines, someone always replies that some are connected to search engines now, which is true, but they don't use them correctly, so honestly, how does that matter? If it can search but instead predicts a string of text that seems potentially useful based on a prompt, that’s less useful than a human entering two key words into a search engine and using their human brain to evaluate the results. Moreover, a recent study suggests that LLMs that are capable of internet searches are actually less accurate than those attached to a more limited database. And summarizing a source (partially and often incorrectly) is not the same thing as providing a person with that source.

The word “intern” in the original post is unintentionally telling. An intern is someone learning to do a job. The purpose of the task isn't just to "offload" something but to teach a new person how to do that task. When I first worked out this train of thought on Bluesky, Joshua Nudell, a historian, replied “Your experience also highlights something missing from this mechanistic formulation: research assistantships and internships are *also* about bringing up the next generation” and “beyond even that the assistant is tasked with applying their critical faculties about relevance, etc.”

In other words, while people can absolutely take advantage of that relationship (asking an intern to get coffee isn't generally teaching them what they're there to learn), the benefits of an internship are at least somewhat shared. The intern gets money and/or experience and/or class credit and/or relationships with the people they meet on the job. The profession gets a future. The researcher gets help from someone with the relevant skills to actually assist them. Because we live in a society, those benefits are not usually equally shared, but they are not entirely one-sided, either. In a world where you have to have a job to have money and you have to have money to survive, there is a real value in jobs being given to humans.

One of the best ways to learn is to teach someone else. Teaching someone else isn’t offloading your critical thinking. It’s spreading it. And one of the great benefits of teaching is that, in connecting with human minds unlike your own, you will learn, too. An annotated bibliography made by a thinking person is a tool that is targeted at a particular set of goals. The person using it still has to think actively to decide how to implement what the other researcher found. They still will need to reject many of the sources, and they still will need to do their own research to find more sources.

An LLM isn't actually learning from their experience, even though we call the process "learning." "Hallucinations" (they aren't hallucinations) are increasing in many models.

You cannot compare working with a person to working with an LLM. They are fundamentally not the same thing. But for many people, because the output they receive looks superficially similar to the work a person would do, the difference isn't obvious.

If you order a gemstone mug online and you get a plastic painted mug instead, you know the difference immediately. However, if you ask an LLM to create an annotated bibliography it will produce something that looks and reads a lot like an annotated bibliography. Some of the sources may even be real. Some of the summaries may be good (because they were stolen from people who know things). It won’t be as quickly clear that you’ve been cheated.

But the methods an LLM uses are not even close to the same process a human researcher uses and it doesn't offer close to the same benefits for any party. There isn't an intern now. No one is being trained for the future of the profession. The results you're given suck, but you may not recognize it right away. When you do recognize it, you can't ask why they did it that way.

Most intellectual work requires frustrating, repetitive work and skipping any part of that process means you will miss out on the actual rewards. When I read five books to get a single partial quote, I still have the benefits of having read those five books. When ChatGPT scans its stolen archives to find the next likely word in an iterative sequence, it does not make judgment calls and you don't learn a damn thing.

And assuming it works exactly as planned, it is still looking for a median. Academic work prioritizes new findings and angles, but an LLM can only reiterate what already exists. It will get you the easiest to find articles because they'll be the most cited. It will prioritize dead white guys because academia has historically prioritized white guys, more of whom are dead than alive.

And it often doesn't work! But even working as planned, an LLM is a bad tool for most researchers.

Post cover image: Hand holding keys on orange background by Marco Verch under Creative Commons 2.0