The Middle of the AI Conversation Is Where the Work Is

There are two camps in most AI conversations, and by now you probably know which one is yours. In one camp: the enthusiasts, the builders, the people who are busy using the technology so regularly that they barely have time to read and reflect on the critiques. In the other: the scholars and critics who have built real intellectual credibility on careful distance from hype, for whom deep hands-on engagement with the tools has become something like an ethical compromise, a sign of capture, naivety, or insufficient rigor.

This divide isn’t only a values disagreement: It’s also, and maybe more importantly, a structural one. Both camps are operating under real social and professional constraints that make crossing over genuinely costly. The enthusiasts aren’t ignoring the critique literature because they’re incurious. They’re busy building with and testing the tools. The critics aren’t avoiding the technology because they’re afraid of what they’d find. Spending time experimenting with the tools risks their credibility with the people whose respect they’ve earned. Neither camp is wrong about their own situation. The incentive structures just happen to produce a conversation where the people with the most exposure to the technology’s actual affordances aren’t positioned to think carefully about what’s at stake, and the people who are positioned to do that thinking aren’t getting sustained hands-on experience. Both kinds of knowledge exist. They’re just not in the same room or social media threads.

The discourse looks different face to face. Online, the two camps dominate. In actual conversation I find far more people who are genuinely uncertain, watching, and waiting. They’re quieter, because holding the question open doesn’t perform as well as certainty in most public discourse. But those waiting and watching are having experiences that shape their perception of these tools. What this middle group needs isn’t more evidence: they’ve seen the studies and the horror stories. What they need is language for the tension they’re already living. They are looking for a framework that doesn’t ask them to resolve the complexity before they’re ready, and one that that lets them continue to learn from the middle.

Libraries can’t afford to pick a side, though it seems that many have. Some rushed toward enthusiasm, repositioning themselves as AI integration hubs before anyone had language for what was actually changing. Others have planted flags in critique, which is intellectually serious but leaves them unable to advise the faculty member who just needs to know what to do with their upcoming research assignment.

I have conversations regularly with people who ask me, without malice, whether libraries will exist in a world with AI. It’s a real question and it deserves a considered answer. But it’s only possible to answer it from the middle because the answer requires understanding what the technology actually does and doesn’t do, and being able to name what’s genuinely at stake when information ecosystems shift. Neither camp alone gets you there.

When you’re paying attention from the right position, the gaps become visible: the things AI doesn’t do, or can’t do, or shouldn’t do, or that a person still needs to do. Much of my career has been based on finding gaps and doing the work inside them. That instinct isn’t a survival strategy. It’s just what happens when you’re paying close enough attention. And we’ll all need to take that approach with AI.

This is a post in an ongoing project exploring libraries, knowledge, and the epistemic stakes of artificial intelligence. I’m drawing on social epistemology, feminist theory, and two decades of practice in academic libraries.

When ‘Probably’ Means Nothing

When I moved to the Pacific Northwest I was surprised how much people volunteered to me that they loved the Southern word “y’all.” It’s a great inclusive way to call a group together or refer to a team. I love it, too. But my favorite Southern phrase is “might could.” It’s double-hedged, which may appear to be redundant or imprecise, but actually it’s the opposite. It’s a finely calibrated expression of a qualified possibility that a single modal can’t quite capture. “Could” alone is too open. “Might” alone is too tentative. “Might could” lands somewhere specific that neither word reaches on its own. It’s also situated. You know something about the speaker when they say it. It carries place, community, a whole set of social relations. Which is exactly what Haraway is talking about in situated knowledge.

Hedging language can be perceived as negative or as an indication that the speaker isn’t confident. But in academic circles it often is interpreted as a signal of some epistemic humility or recognition that the concept has enough complexity that you need a bit of hedging to remain accurate. When a scientist says “probably,” a doctor says “likely,” a colleague says “I’m fairly certain,” those words are doing the real epistemic work of communicating a speaker’s actual relationship to uncertainty, calibrated by experience, context, and stakes. It’s worth reflecting on what is lost if these turns of phrase are stripped of their nuance.

When I read ‘Probably’ Doesn’t Mean the Same Thing to Your AI as it Does to You, I was struck that our LLMs may not be using hedging language in the way that we do. LLMs use words like “probably,” “likely,” and “almost certain” inconsistently, averaging over conflicting usages in training data rather than assessing actual odds. The article also points to an interesting intersection with gender studies, showing that the same probability expressed differently depending on whether the prompt says “he” or “she.”

This is a really specific kind of epistemic failure, and an interesting one! Hedging language is how knowledge communities signal the limits of what they know. Strip that calibration out and you get fluency that performs humility while enacting the view from nowhere. This is Haraway’s god trick at the lexical level. We’re moving beyond the synthesis of sources and into in individual word choices.

We’ve all seen use cases in which AI in increasingly being used to summarize research, brief decision-makers, and mediate information. We also are all aware of the conflicting views on to what extent that information is actually good. For now, at least, it seems that we may also have to consider the word choice itself. When the methods we have to convey certainty lose their clarity we may find ourselves being overconfident in our interpretation of words, only to find we’ve made decisions without the information we assumed was supporting our path. Things appear as they were, but in reality the world shifted around us. We read “probably” and think we know how confident to be, but the word has already lost its weight.

The Categorical Collapse of AI

When someone says they’ve been “reading,” you don’t actually know what they’ve been doing. They might have spent a week with a dense classic novel. They might have scrolled through their phone for twenty minutes. Both are reading in a technical sense: their eyes move across text and they process the words they see. However, the cognitive activities involved are so different that calling them by the same name obscures more than it reveals. One develops the capacity for sustained attention, enables the reader to enter a fictional world, and requires tracking complex characters across hundreds of pages. The other is closer to foraging. It may surface interesting and relevant information, but the cognitive work is different. Walter Ong would say these aren’t even the same species of activity. His writing argued that different communication technologies don’t just change how we do something but produce fundamentally different kinds of cognitive events.

We have this problem with AI, and it’s worse. “Using AI” currently describes a number of different activities. You may use AI to ask a chatbot what to make for dinner, to draft a briefing document, to generate data for a research study, to use a recommendation algorithm to find a movie, to vibe code, or to build a tutoring system that adapts to individual learners. These are not variations on a single activity. They involve different tools with genuinely different capabilities, different cognitive demands, different stakes, and different relationships to truth and accountability. And they don’t collapse neatly into a skill hierarchy. (We’re also all aware that some things AI does badly regardless of how well you’ve learned to work with it.)

And yet our discourse treats them as one thing. Raymond Williams, writing about what he called “keywords,” observed that certain words carry unresolved tensions precisely because different groups use them to mean fundamentally different things without realizing it. “AI” is a keyword in exactly this sense. Which means that when someone says AI is transforming education, and someone else says AI is producing misinformation at scale, and a third person says AI is going to replace libraries, they are often not talking about the same phenomenon at all. The conversation generates heat without light because we’re using a single word to point at a dozen different things.

The reading analogy is useful here because we actually worked this out with literacy. We distinguish between reading and reading critically, between reading for pleasure and reading for research, between being able to decode text and being able to evaluate an argument. A first-year writing course and a doctoral seminar both involve reading, but nobody confuses them. We built vocabulary and practices for the distinctions because we needed to teach the skills, and we needed to evaluate whether people had them.

We don’t have that vocabulary for AI yet. And the absence has the potential for damage. This lack of precise vocabulary makes it hard to even talk about AI literacy in any meaningful way, because we haven’t agreed on what the relevant skills even are. It means we can’t evaluate institutional AI practices, because we’re not being precise about which practices we’re examining. It means we can’t have a useful policy conversation, because the thing being regulated keeps shifting shape. Bowker and Star, in their work on classification, argued that collapsing categories doesn’t just muddy language. It does real epistemic and political work, obscuring accountability and making certain questions harder to ask. For example, classifying all AI use as equivalent makes it harder to hold vendors or institutions accountable. That’s what’s happening here.

This isn’t to say the work isn’t happening. Librarians and educators at many institutions are actively developing thoughtful AI literacy frameworks. But the frameworks vary considerably in scope, in assumption, in what skills they prioritize. This is, itself, evidence of the problem. We haven’t yet agreed on what we’re teaching because we haven’t yet agreed on what we’re talking about.

Libraries have always been in the business of literacy in the expansive sense: not just decoding text, but developing the critical practices that allow communities to engage meaningfully with information. That work is urgently needed here. Not “AI literacy as a single thing to be achieved,” but AI literacies as a differentiated set of practices: knowing which tool does what, understanding what accountability looks like in different contexts, recognizing when fluency is masking the absence of provenance.

Before we can teach any of that, we need to stop talking about AI as though it’s one thing, and be clearer about what we’re describing.

The Obsolescence Argument Has It Backwards

Everyone seems to agree that artificial intelligence is going to change education, research, and libraries. The disagreement is about direction. The dominant narrative, at least in some technology circles is: AI can find information, synthesize sources, and answer questions. It’s not a surprise that people hearing that argument in media and from tech commentators point out that libraries and librarians do those things and then assume that libraries are in trouble.

But to anyone who sits at the intersection of technology and libraries it’s abundantly clear that AI doesn’t make libraries obsolete, but rather it makes them more essential.

I’ve been thinking about knowledge systems for a long time. My undergraduate degrees were in philosophy and in communication, with a minor in Women’s and Gender Studies, and the questions that animated these fields were the same ones: Who knows? Under what conditions? With what authority, and on whose behalf? Those questions led me to library science, and they’ve shaped how I’ve understood this work ever since.

Two frameworks have always been particularly generative for me. The first is social epistemology. This term was developed by Jesse Shera and Margaret Egan in the mid-twentieth century, which understands libraries not as warehouses of information but as infrastructure for how communities produce and share knowledge. Libraries, in this view, are epistemic institutions. They don’t just store what we know; they shape the conditions under which knowing is possible. (Incidentally social epistemology also developed within Philosophy, with a slightly different implementation, a few decades later.)

The second is feminist epistemology, particularly Donna Haraway’s concept of situated knowledges. Haraway’s argument, made in a landmark 1988 essay, is that all knowledge is produced from somewhere: from a particular body, a particular history, a particular set of social relations. Claims to view-from-nowhere objectivity, what she calls the “god trick,” are not neutral. They are themselves a kind of power move, one that erases the conditions of knowledge production and forecloses accountability. Sandra Harding’s standpoint theory extends this: knowledge produced from the margins, from positions of accountability rather than dominance, is often more comprehensive, not less, because it cannot afford to ignore what the center takes for granted.

These frameworks were developed to critique science. But you can see why I keep coming back to them today.

Large language models perform exactly the god trick Haraway identified. They synthesize at scale without provenance. They produce authoritative-sounding outputs whose origins are opaque, whose training data encodes historical power imbalances, and whose confident tone actively discourages the epistemic humility that good inquiry requires. They are, in Harding’s terms, knowledge produced from nowhere. And this means they are making claims from a position that cannot be held accountable.

This is not primarily a technical problem. It is an epistemic one. And it is precisely the problem that libraries, at their best, are structured to address.

Libraries curate situated knowledge. They preserve provenance. They maintain the bibliographic infrastructure that allows a reader to ask: who said this, when, from what position, in conversation with whom? They select, describe, and organize materials in ways that make the conditions of knowledge production visible rather than erasing them. They employ people (librarians!) whose professional expertise is not only finding information but teaching the critical practices that allow communities to evaluate it.

None of that is replicable by a system that has been specifically designed to flatten those distinctions into fluent prose.

I’m not arguing that AI is useless, or that libraries should resist it, or that the landscape isn’t changing. It is changing, and libraries need to engage with that change thoughtfully and without too much nostalgia. What I am arguing against is the idea that AI supersedes libraries. When someone asks whether AI makes libraries obsolete, the questioner implicitly accepts a definition of libraries as information retrieval systems. That is a definition that was always reductive and is now actively misleading. Libraries are epistemic infrastructure. They are, in Shera and Egan’s terms, the social mechanisms through which communities organize their relationship to knowledge.

AI doesn’t replace that. It creates new urgency for it.

The more our information environment is shaped by systems that perform objectivity while encoding power, the more we need institutions committed to making those dynamics visible. As synthetic text becomes more fluent and authoritative, it will become more important for human thinking to maintain the skills in citation, provenance, critical evaluation, and the slow work of understanding where knowledge comes from. These are the skills that libraries cultivate.

The obsolescence argument has it exactly backwards. This is the moment libraries were built for.

This is the first post in an ongoing project exploring libraries, knowledge, and the epistemic stakes of artificial intelligence. I’m drawing on social epistemology, feminist theory, and two decades of practice in academic libraries.

Before we begin

Years ago I kept a blog (at this URL, even!) where I thought out loud about libraries, knowledge, and the profession I’d built my career around. I was good at it for a while, and then I wasn’t, and then I stopped for all the usual reasons: changing life phase, less personal time to spend on it, increasingly demanding institutional role, the way the platforms evolved from places of earnest and open discussion… I drifted so far away from blogging and this website that when a back up didn’t capture all the files I wasn’t even all that disappointed.

But lately I’ve really missed thinking in public with other colleagues interested in exploring the same ideas. And lately I’ve been thinking a lot about academic libraries, our information environment, and the ways we talk about and use artificial intelligence.

AI is reshaping how people find, evaluate, and trust information. Within libraries we have people all across the spectrum: from those who fully embrace it to those who believe it has no place near our work. One of the dominant narratives outside of the profession treats libraries as information retrieval systems and concludes that AI makes them redundant. This framing mistakes the symptom for the disease. Libraries are epistemic infrastructure. They are the mechanisms through which communities organize their relationship to knowledge. AI doesn’t replace that, but it does make that role all the more urgent.

This lens keeps coming up for me in conversations in varied spheres. Jesse Shera and Margaret Egan’s social epistemology, which understands libraries not as warehouses but as institutions that shape the conditions under which knowing is possible, is foundational to how I think about this work. So is feminist epistemology, particularly Donna Haraway’s concept of situated knowledges and Sandra Harding’s standpoint theory. These frameworks were built to interrogate science. But it turns out that they are extremely useful when interrogating AI as well.

I’m writing as a person who has spent two decades in academic libraries and who has been thinking about knowledge, power, and institutions since an undergraduate philosophy degree made those questions unavoidable. At this URL, I am not writing as an institutional voice. This is a thinking space. I’m hoping that arguments will develop, get complicated, and occasionally get revised. I expect to adapt to new information.

What follows this post is the first real argument: why the obsolescence narrative has it backwards, and what a clearer account of libraries and knowledge reveals about the epistemic stakes of this moment.

I’m still trying to understand where people talk about these things today. In some ways everything was a lot cleaner when the answer was a blog with open comments, an RSS reader, and Twitter. The messiness of our knowledge environment today (LinkedIn? Bluesky? Mastodon? SubStack? Chat threads? Everywhere?) resonates with the messiness of the information ecosystem I’m trying to write about.