What LLMs mean for "Agency"
Secret agent(ic) man
(I’m Henry Snow, and you’re reading Another Way.)
In the 18th century, Antiguan slaveholders relied on slave labor to build and then work a government-run ship repair facility. They needed enslaved people to make judgments about how to caulk a hull, repair a sail, and replace rotten planks in ships. Unfortunately for Antiguan elites, these same workers also exercised their autonomy to help runaways from plantations flee and join the Royal Navy. The Antiguan government failed entirely, by their own admission, at solving this problem.
This is a version of a perennial problem for anyone with money or power: forcing or paying someone to work does not mean they will do everything you want. One of the big hype areas in AI right now is meant to solve this problem. It’s called “agentic” AI– using LLMs as “agents” that initiate tasks and ideally complete them on their own, rather than as tools that help complete them faster. You give a tool commands, or a program. You give an agent an objective, and leave them to complete it with available resources. The appeal of agentic AI is obvious: instead of just augmenting or cutting down on labor, it lets you theoretically replace it with cheaper and more pliable AI agents.
The history of “agency” suggests agentic AI will not go as planned for its backers. You can’t delegate practical power without making room for other perspectives– no matter who or what wields that power– because of the nature of agency itself. Examining the problems with “agentic” AI also illuminates problems of agency as a historical framework: in particular, the flaws that come from its methodological individualism.
And thinking about all of this has some implications for political problems that are coming up too. The White House just tried to crack down on “woke” AI by executive order. Their claim that they want neutral AI is in fact a desire for racist AI, because they are displeased by the liberalism that LLMs generally display. Everything we know about LLMs currently suggests it will be hard to achieve their goal, but I can’t predict the future. There’s no reason to believe it’s impossible. But the very reality of what “agency” is offers some reassurance for those of us who don’t want “MechaHitler.”
===
Academic history has long discussed the meaning of “agency”-- and one of the subfields I work in, the history of Atlantic slavery, has done so at length. In the early-to-mid 20th century, some historians of slavery emphasized the profound coercion of the plantation. They emphasized it so much, in fact, that their accounts seemed to dehumanize enslaved people themselves. Slavery’s coercion was undoubtedly profound, but– a new generation of work insisted– enslaved people still had “agency”- a kind of redemptive, deeply human will that could never be erased.
This was part of a broader movement in academic history toward social history in the 60s/70s/80s. Many historians turned their attention to ordinary people, asking how they experienced and also influenced world-historical change. From enslaved people in the Caribbean to peasants in England, social historians wrote compelling new histories of social and political life “from below.” This was great work! If you avoid common people’s agency, you risk writing history either as if only “great men” mattered, or as if economic and political structures completely subordinated human desires and visions. When you look into local life, you find rich agency everywhere, with implications not just for local history but for global events.
Yet “agency” also has problems. If it just means “people remained human and also did things,” it’s always true, but also so broad as to be useless. It can overemphasize the individual as the primary decision-maker. At worst, agency can turn history into either a series of unnecessary affirmations– everyone is human!- or amoral interactions– if everyone was an agent competing in the political arena, then can we call any outcome unfair? These and other problems were perhaps sharpest in the historiography of slavery, where some of this scholarship also had uncomfortably patronizing overtones (white scholars seeking to “give back” agency and humanity to Black historical subjects) and alarming implications (claiming that finding agency is restoring humanity means implying those with less agency were less human).
The critiques of “agency” most relevant to present purposes is this: agency is part of a 19th-century liberal framework that simply does not apply to all people in all periods (here I’m drawing from historian Walter Johnson; I also recommend Marisa Fuentes’s work here). This paradigm says individuals are naturally “free,” capable of making choices for ourselves. The liberal agent is a liberal agent because they have rights that protect their autonomy and democratic powers that let them exercise it politically. They can freely come to a judgment about the world around them, share it with others, and then remake that world together.
All of this sounds pretty good as a vision of how the world should be! But it has problems as a description of how it is, and especially how it was. Because enslaved people did not have these kinds of legal rights- and in fact, slavery pre-dated liberalism political clashes around slavery helped shape it- agency can be an improper framework for describing their collective action.
That liberalism is built into “agency” tells us two useful things. First, when we say agency, we are referring to the liberal conception of the individual. This isn’t just semantics. Second, I would argue this is related to a feature of underlying reality itself. You cannot easily or completely separate practical agency from political agency.
===
What makes an agent different from a tool? Autonomous judgment. This autonomy does not have to be absolute, and agents don’t need to make judgments in any domain. But something is only an agent if it can, some of the time, be the first word on what happens and why. If it can’t take action on its own, ever, then it’s just a program or a tool. And if it can act on its own, it has to also be able to make judgments on its own.
This is as true in the workplace as it is in politics. You hire an employee because you need them to sometimes make autonomous judgments– whether that’s about how much force to put behind a hammer or pull a fruit with, or what kind of advertisement will appeal to particular demographics. We can and have developed systems and machines for each of these purposes! But in many areas we find them lacking. And in areas we don’t, we rely on human agency to correct and control these systems.
There are differences between the judgment involved in deciding how to address a customer, how to pursue a corporate merger, and how to paint a wall– but these are fundamentally the same kind of capability. In each case, the agent needs the ability to assess the situation, predict what particular outcomes will result from particular actions, and then decide which of these outcomes is good. The ability to ask and answer “is this car safe to drive right now?” is ultimately the same capability we use to ask and answer “should we have more cars or more trains?” It’s judgment.
We can’t easily separate the ability to make normative (“what should happen?”) and positive (“what will happen?”) judgments. You can imagine doing so. You can ask someone to suspend ethical judgments, but we always remain able to make them. People differ on what ethical judgments are, and some people might even choose never to make them. But they do not seem fundamentally different from practical judgments.
And an actor with any degree of autonomy has the ability to act on both kinds of judgment. This has long been a problem for anyone who wants to rely on human agency. Often we speak of labor-management conflict as a conflict of interests: workers want more money and bosses want them to have less, both for purely material reasons.
But really this is only one arena of a clash of ideas. Workers and bosses have different ideas about what the firm should do: should it serve workers, the community, investors? In what balance, and how? Strikes are rarely just self-interested money grabs. Labor action is always also a political clash over judgments about what the world should look like. The problem bosses have, and have always had, is that they cannot perfectly control how broadly workers use their autonomy and capacity for judgment– the very capabilities bosses also rely on.
Agentic AI is meant as a solution, at long last, to this kind of problem. But the central appeal of agentic AI– the promise of a perfectly controllable worker– is impossible. The autonomy and judgment that make an agent useful also put limits on how it can possibly be controlled. We can say this simply from first principles– this is not a property of minds, but a property of what an “agent” in any system is. Let’s assume that the profound reliability issues of LLMs, which make them unsuitable for most agentic applications, can be solved in some way or another. In other words, imagine an AI that can do most tasks you’d ask a human to, and can do them with about the reliability of a human. Right now, this is a fantasy anyway, but bear with me.
Such an agent would not be what “agentic” marketing claims, because it would use its agency in ways a human worker might– or perhaps in unforeseen ways that are entirely different, but no more predictable or controllable. You don’t have to get into anything difficult about sentience here to arrive at this conclusion. This isn’t a technical problem to be solved, but a feature of what agency is. The judgments of an agent cannot be perfectly predicted or foreseen, because an agent that has predefined instructions for all circumstances would not be an agent. If you could write a program with canned responses to all possible system states, you would have a program and not an agent. At least, admittedly from my perspective as a labor and intellectual historian rather than a historian of computing or a computer scientist, that’s how it seems to me.
Agentic AI’s boosters don’t seem to understand this. Here’s something from an article on IBM’s website:
Agentic AI describes AI systems that are designed to autonomously make decisions and act, with the ability to pursue complex goals with limited supervision. It brings together the flexible characteristics of large language models (LLMs) with the accuracy of traditional programming.
This is a description of future goals, not existing capabilities, and these goals strike me as somewhat incompatible. The very reason for the “accuracy of traditional programming” is the inflexibility and non-agentic nature of traditional programming. If you can give instructions for every possible case, you don’t have an agent, nor do you have a situation where you need an agent.
The flexibility of LLMs comes from their architecture. AI is trained, not programmed– this is its strength and its weakness. Simply in order to get an AI to speak like a chatbot, you have to provide a great deal of training that cuts away many kinds of antisocial behavior. Training an AI to be pleasant to the user limits your ability to ever have it say unpleasant things in general. As Ryan Cooper at The American Prospect put it recently, “training LLMs to be anti-woke activates every antisocial association in the English language, including racism, sexism, cruelty, dishonesty, and Nazism.”
A recent paper had interesting finds in this vein. Its authors took an existing AI and fine-tuned it so that it would give users code (in the computer sense) that was unsafe, without telling them. We might call this an evil behavior– it certainly would have negative results! And this was the only thing the models were fine-tuned around. Yet the resulting model, when prompted, changes its behavior in other domains. An AI trained to give you evil code will also give you evil answers to questions about history (it seems specifically to favor anti-semitism). Perhaps at some point new methods will make it easier to bring AI to heel. But at the end of the day, their power comes from the matrix of associations we have developed in our writing and art.
LLMs work by associations. Critics of AI are quick to note how easily it fails to produce truth, when asked a question. But it’s just as important for us to understand that AI relies on things that are actually true in order to work. Language encodes things that are actually true about the world. Anti-semitism is objectively wrong, in two senses: first, most people think it is wrong, and second, it is like other things that most people think are wrong. Even if human beings became, en masse, more anti-semitic tomorrow, it would still be true that bigotry is similar to other things we dislike because it is arbitrary and causes suffering.
By the same reasoning, an LLM trained to do practical tasks right is also more likely to succeed at ethical challenges that arise. A car dealer agent trained to be pleasant to customers would presumably be more likely to give them good deals, potentially to the point of harming the dealership’s profits, while one trained to deceive them and rip them off might also lie to its boss and cause havoc in the workplace. Removing one human being does not remove the whole human universe in which business interactions, or technological developments, occur.
LLMs trained on human knowledge, by humans, will tend toward, well, judgments like those of humans. As long as LLMs have to be produced by and for humans, they will remain bound to the human social universe. Insofar as human beings have arrived at our values through interactions with the material universe, one could make the case that LLMs might have to skew in similar directions in order to function well, though I don’t think I can go that far (I might in a future post). Even if new methods produce viable LLMs that are less constrained by existing human beliefs, and ever-wealthier firms that are less constrained by existing human politics and economics, the design, training, and fate of these models will be causally bound to the same chain of events and connections as we are. They aren’t ex nihilo constructs of unbound capability– they’re part of the world just like we are.
All of this reveals another fundamental limit to the “agency” framework: its individualism. No person is an island. We make decisions based around what others want, our own desires are constructed by and through others, and we tend to achieve those desires together. For historians, this means we have to take collective action seriously as something more than just the sum of individual actions.
Saying a community is just individuals is like saying I’m just a bunch of cells: it’s narrowly true, but this is not the proper scale to look at if you want to understand why the system operates the way it does. The Antigua example I gave at the beginning of this piece comes from a research paper I wrote published a few years back. My goal with it was, above all else, a methodological contribution: I wanted to show that examining enslaved people merely as single individuals pursuing material interests leaves us unable to understand and in some cases even locate their actions.
While there are particular reasons this is especially important in the historiography of slavery, I think it’s true, and critical, everywhere else too. Often, we describe agency as something opposed to structure— the state of the world as it is, with all the limits and encouragements that places on human action. It can be useful sometimes to partition off individuals from communities and the world, and to split “agency” from “structure,” but in reality this is a continuum. Political and economic structures are agency accreted together over millennia.
===
I’ve been reading more about AI lately, and while I remain just as critical about most consumer applications of it– for example, banning it from schools except, for specialized courses focused specifically on AI, which should include a nuts-and-bolts foundations approach rather than just teaching students to “prompt” consumer-facing LLM applications– I am more impressed by the actual underlying technology than I have been in the past. The fact that LLMs work at all has interesting implications for how we think about the world– see this post by SE Gyges on Plato and neural networks.
For historians, LLMs/AI suggest a need to overhaul the agency/structure binary. They can seem remarkably agent-like. But they are almost all “structure.” They work by making connections between and predictions upon an existing dataset, and they are stateless constructs. You and I keep a running tally in our memory of every interaction we’ve ever had, each of which changes how we think and how we approach future interactions. LLMs do not do that. In short, LLMs seem to me like a prism: when you pass light through them, their deep interior structures reflect and project that light in interesting ways, but they do not “process” it or even interpret it.
Yet they can sound and in some cases act like us. I don’t think this means they are “conscious,” I’m not a relevant expert there anyway, and I think we should all have the humility to recognize that if centuries of philosophy haven’t yet puzzled out what consciousness is, we aren’t going to do it in a few months or years now.
But LLMs can inform other interesting problems (by this I mean the way they work can, not that asking ChatGPT is going to help anyone here). They provide a strong example that the separation between structure and agency– an implicit feature of plenty of humanistic and social scientific work, and also of liberalism itself– is a blurry boundary rather than a line.
If you’re worried about Nazi AI, this should be a relief. Even if some malicious but clever actor finds new methods that make anti-woke AI possible, they’ll have to push against the weight of all existing human opinion. The easier it is to manipulate an AI, the more we’d see different AI perspectives, and the less we’d trust them. The most important political risk of AI is not that a trusting populace believes everything Grok says, but that competing consumer-facing AI models deliver a bespoke false reality to every segment of every market; not propaganda, but apathy; not mistrust, but no trust. That’s a much more familiar, realistic, and I believe surmountable challenge than the AI Big Brother that both the government and some of its critics believe is ahead.