The Fifty-Year War (or "Reality Bats Last," Part 2 of 2)

The intellectual history of why some of your smartest colleagues are so unsettled by AI and chatbots

For those of you were paying attention a couple of weeks ago, in my last piece on this, I argued that “reality bats last.” Reading that first piece will probably serve you well in understanding this one (and it’s done pretty well with lots of reads, so I can at least somewhat recommend it to you), but this one can also work as a standalone. Also, there’s a Part 3 in the works, probably coming in a couple of weeks, taking on the side that worships only data and computation. That…is also worth a discussion to my mind.

Anyway, in that piece, I showed you a paradox: AI is simultaneously The Simulacrum Fulfilled (the machine that generates meaning with no tether to anything real) and the strongest empirical challenge to strong constructivism in fifty years. The hallucinations aren’t evidence that AI is bullshit. They’re covert admissions that there’s a reality to hallucinate from.

This time I want to explain why that claim makes so many smart people so angry.

But here’s what’s been strange to me about all this: the academy has been arguing about quantification and objectivity for decades — that’s not new. What’s new is the intensity. Scholars who’ve coexisted with statistics departments for fifty years are suddenly treating AI like an existential threat. The critique went from “be careful” to “ban it” almost overnight. That escalation is worth paying attention to. It tells you something deeper is at stake than ChatGPT in the classroom.

The real story starts fifty years ago.

If you liked (or hated) Part 1, you’ll probably feel the same about this one. If it’s making you think, please share it.

The War Against the Positivist Stack

There’s a fifty-year epistemological posture in the academy: a deep, institutionalized commitment that treats the entire apparatus of formalization, quantification, data, technology, measurement, and objectivity as suspect, dehumanizing, expressions of power. This isn’t a recent development. It didn’t start with chatbots. It’s been building since the 1960s, and AI is just the latest and sharpest point of contact.

The lineage is long. It runs from Heidegger through the Frankfurt School, through Dreyfus’s embodiment arguments, through Searle’s skepticism about symbols and meaning, through feminist epistemology’s insistence on standpoint, through the Science Wars of the 1990s, through the entire apparatus of critical theory and STS. Each node added sophistication. Each one also mutated the original caution into something harder and more dogmatic.

Start with Heidegger. He wasn’t a constructivist (his concern was ontological, not epistemological), but the tradition that flowed from him mutated the insight beyond recognition. In “The Question Concerning Technology” (1954), he said technology isn’t neutral. It structures what becomes visible and intelligible. Modern technology reveals the world as “standing reserve,” resources to be optimized and exploited. Smartphones do structure what we attend to. Algorithms do shape what becomes visible. He wasn’t wrong.

The Frankfurt School inherited that worry and weaponized it. In Dialectic of Enlightenment (1947), Horkheimer and Adorno argued instrumental reason subordinates all values to efficiency. Scientific knowledge and technology are “means of exploitation.” Quantification is how domination operates. Capital does optimize relentlessly. Factories are expressions of instrumental logic. This wasn’t crazy either.

Then Dreyfus came along. What Computers Can’t Do (1972) argued that formalization can’t capture embodied understanding. Computers can process symbols but not understand because understanding requires being-in-the-world, practical coping, embodied skill. There was something real here, something about human judgment that formal rules miss. But it metastasized. Formalization became not just limited but inherently inadequate to knowledge. Then: anything claiming formal objectivity became suspect.

Searle’s Chinese Room argument (”Minds, Brains, and Programs,” 1980) got adopted as proof that formal systems are epistemically bankrupt. (Searle himself wasn’t a constructivist. The constructivists weaponized his argument.)

Feminist epistemology made another important move. Harding’s The Science Question in Feminism (1986) argued all knowledge is positioned; Haraway’s “Situated Knowledges” (1988) called claims to neutrality “the god trick,” a fantasy of view-from-nowhere that erases standpoint and reinscribes power. Scientists do have standpoints. Research is shaped by who gets to ask questions. The solution was supposed to be “strong objectivity”: reflexive acknowledgment of position alongside commitment to rigor.

But somewhere between the sensible insight and the institutional reality, something mutated. “Standpoint shapes knowledge” became “there is no knowledge, only perspective.” “Power shapes inquiry” became “claims to truth are just power plays.” “Formalization has limits” became “formalization is dehumanizing.”

The Science Wars of the 1990s cemented it all. Alan Sokal submitted a nonsense paper to Social Text proposing that quantum gravity is “a social and linguistic construct.” They published it. Sokal revealed the hoax. He and Bricmont later published Fashionable Nonsense (1998) documenting the pattern. You’d expect the constructivists to retreat, chastened. Instead the entire constructivist apparatus spread. Bloor’s Strong Programme (Knowledge and Social Imagery, 1976), Latour’s actor-network theory, the demystifying attitude toward science: it all became institutionalized as the correct critical perspective. Not as one view among others, but as enlightened consciousness. To speak of objectivity in most humanities departments after that became epistemically suspect. Not contested. Actually suspect, the way believing in flying saucers is suspect.

The Architecture of Hostility

What constructivists have been defending is a specific epistemology: knowledge requires social construction by situated knowers. The entire project depends on it. Literary analysis constructs meaning through dialogue. Qualitative research depends on reflexivity and positionality. Critical theory privileges lived experience and situated knowledge.

AI doesn’t just challenge this epistemology. AI refutes it while appearing to prove it.

LLMs generate expert-level analysis without lived experience, without community membership, without being positioned anywhere. They do it at scale. This shouldn’t be possible if standpoint epistemology is right. If knowledge requires a subject, a person with stakes, embodiment, social location, then an algorithm without those properties shouldn’t be able to generate knowledge. Yet it does. And it does competently enough that you can’t just wave it away as scientism anymore.

This is the real offense. Sure, job displacement matters. Bias matters. But the thing that actually keeps constructivists up at night is conceptual. The mere concept of AI, that knowledge can be extracted from patterns without social construction, that formalization captures something real, that objectivity isn’t a myth, is an affront to the epistemology that has structured academic identity for half a century.

Constructivists have a name for this offense. They call it “epistemic injustice.” The idea comes from Miranda Fricker’s Epistemic Injustice (2007), and it’s a genuinely useful concept. Epistemic injustice is when a social structure produces injustices in knowledge and understanding. When an algorithm claims objectivity without acknowledging that it’s encoding someone’s standpoint, that’s epistemic injustice. When an AI system erases marginalized perspectives in favor of dominant ones, that’s epistemic injustice.

All true. The problem is what constructivists do with it. They use epistemic injustice as proof that formalization and objectivity are inherently oppressive. They use it to demonstrate that the entire positivist stack (data, algorithms, quantification, claims to truth) is an instrument of domination.

But there’s a deeper problem with calling something “bias” when your whole framework says objectivity is a myth.

The Mutation Arc

This matters because the hostility has a shape. It didn’t appear fully formed. It mutated through phases.

Phase 1 was sensible caution. Be careful about objectivity. Positivism has limits. Not all valuable knowledge is quantifiable. Power shapes scientific communities. Objectivity without reflexivity is naive. All of this is reasonable. I’m not arguing it’s wrong. I’m arguing it got transformed.

Phase 2 hardened the cautions into claims. “Objectivity is a myth” (not just harder than we thought). “All knowledge is socially constructed” (not “social factors matter”). “Formalization misses what’s essential” (not “has limits”). “Technology is inherently dominating” (not “has power effects worth scrutinizing”). These became philosophical truths, not methodological positions.

Phase 3—we’re still living in it—turned these into institutional common sense. Not enforced by committee. Enforced the way all orthodoxies are enforced: through the slow accumulation of what sounds smart and what sounds naive in professional conversation. I know which words I’d think twice about in a job talk. I know which framings would get a quiet nod in one room and a raised eyebrow in another. Positivism became a slur. Scientism became a boo-word. To claim objectivity without heavy qualification marked you as unsophisticated—not wrong, exactly, but not someone who’d done the reading. This wasn’t presented as one perspective among others but as the baseline of serious scholarship. You didn’t have to be told. You just absorbed what counted.

Philosophers have a name for this structure: motte and bailey. The motte is the defensible position you retreat to when challenged — “power shapes knowledge,” “be careful about objectivity,” “perspectives matter.” Hard to argue with. The bailey is the territory you actually occupy when no one’s pushing back — “objectivity is a myth,” “all knowledge is just perspective,” “formalization is inherently dehumanizing.” When critics point out the bailey is indefensible, defenders retreat to the motte and act as though that’s what they were saying all along. The cycle has been running for decades now, and it’s made the position nearly impossible to engage with honestly. Challenge the strong claim and you get the weak one. Accept the weak one and the strong one reasserts itself the moment you look away.

That mutation matters more than any individual node in the lineage. The original cautions were often right. The mutation into dogma created something new: an entire institutional apparatus committed to the claim that formalization, quantification, and objectivity are inherently suspect—not as a methodological position but as philosophical truth. Reality bats last. Dogma doesn’t.

The Habermas Problem

Jürgen Habermas died recently at ninety-six, and his death is a moment worth sitting with because he represents the tension at the core of constructivism.

Habermas was part of the Frankfurt School tradition. He inherited the critique of instrumental reason. But he refused to abandon rationality itself. In The Theory of Communicative Action (1981), where Horkheimer and Adorno said instrumental rationality is totalizing and oppressive, Habermas said there’s another kind of rationality. Communicative rationality. Rationality oriented toward understanding, grounded in dialogue, governed by validity claims that can be tested and contested.

This was the gateway drug. Habermas made constructivism legible to people who also believed in reason and evidence. He showed you could take the Frankfurt School seriously without becoming a relativist. Democratic legitimacy, he argued in Between Facts and Norms (1992), depends on communicative rationality—people talking, giving reasons, being answerable to one another.

For decades this worked as a framework. Habermas wasn’t claiming only sincere agents can produce valid arguments. He was claiming democratic legitimacy requires that participants be oriented toward mutual understanding. The deliberative process does the filtering. The forceless force of the better argument wins because producing good arguments is hard, and the difficulty creates accountability.

Then argument production got cheap.

An LLM doesn’t refute Habermas. What it does is something he never had to worry about: it floods the deliberative space. Systems with no orientation toward understanding, no vulnerability to counterargument, no stake in the outcome can now generate validity-passing claims at scale. The arguments aren’t bad—that’s the problem. They’re good enough to survive scrutiny while having no accountability behind them.

The sorting mechanism degrades. Not because the normative ideal was wrong (deliberation oriented toward understanding really is better than deliberation without it) but because the empirical conditions Habermas assumed no longer hold. He assumed scarcity of good arguments. He assumed that the difficulty of constructing the better argument would do filtering work. AI removes that assumption. And it removes it while appearing to honor every standard Habermas cared about—rigor, evidence-responsiveness, the strength of reasons.

That’s the trap he left open. Not a refutation. A new problem that only gets worse.

The Dreyfus Firewall Collapses

Remember Dreyfus’s core claim from What Computers Can’t Do (1972) and its sequel What Computers Still Can’t Do (1992): understanding requires embodied being-in-the-world. Computers can process but they can’t understand because understanding requires practical coping, cultural embeddedness, the felt sense of how things matter. Harry Collins extended this: machines could never achieve “interactional expertise,” the kind of knowing that comes from being embedded in a community of practice. This was the firewall. It’s what made the humanities special, what justified the claim that interpretation requires a human knower.

For fifty years this held.

Then the technology changed. Not because it solved Dreyfus’s philosophical problem (we’re still not sure what “understanding” means) but because it became functionally irrelevant. Who cares if an LLM “understands” in the phenomenological sense if it can generate expertise that transfers across domains, solves novel problems, survives scrutiny?

The question isn’t “Can computers understand?” anymore. That was always unanswerable, and not for the reasons Dreyfus thought. Watch the pattern: when computers learned to play chess, understanding got redefined as “not mere calculation.” When they learned to translate, understanding became “not mere pattern matching.” When they started generating competent literary analysis, understanding became... something else again. Something that always, conveniently, required whatever the machine hadn’t demonstrated yet.

That’s not philosophy. That’s a ratchet. The definition of understanding kept moving to stay one step ahead of the technology, which tells you the concept was doing protective work, not explanatory work.

So the question now is simpler and harder: if a system without embodiment, without community, without lived experience generates knowledge that experts recognize as competent, what exactly is the constructivist claim protecting?

The firewall wasn’t breached. It was revealed as a comfort blanket protecting disciplinary commitments from empirical challenge.

The Constructivism Defeater?

LLMs generate knowledge without social positioning. Recent work in AI & SOCIETY on “The Epistemic Revolution of AI” frames this directly: current LLMs embody a positivist epistemology that treats knowledge as an external regularity latent in data, which contradicts constructivist assumptions at the root. Standpoint epistemology says knowledge requires a standpoint—a socially located position shaped by gender, race, class, experience. An algorithm has none of this. No lived experience. No stakes in the outcome. No community. Yet it produces expert-level analysis in domains that supposedly require all of these. If standpoint is necessary, this shouldn’t work. But it does.

There’s an inverse Sokal here too, and it’s worth pausing on.

Sokal showed that constructivists would publish nonsense if it was rhetorically pleasing and ideologically conforming. Fine, that’s a charge they had to answer. But AI shows something harder to dismiss: systems can generate truth without the social process. Without rhetoric. Without community validation. Without any of the apparatus constructivists said was necessary. If social construction is all there is, what constructed the training data patterns that AI successfully learned and generalized from? What’s doing the constructing when there’s no constructor?

Sokal embarrassed the gatekeepers. AI sidesteps the gate entirely.

Then there’s the training data objection, and it deserves more than a dismissal. (Ian Hacking’s The Social Construction of What? from 1999 is still the best guide to how constructivist claims slide between the reasonable and the absurd, and he saw this slippage decades ago.) When constructivists respond to AI success, they say: “Training data are themselves socially constructed.” And they’re not wrong about this. They’re genuinely not. What got written down, by whom, in what language, what got digitized and preserved, what got funded and therefore studied: all of that has social fingerprints on it. English dominates training corpora. Western scientific papers are overrepresented. Entire knowledge traditions got left out because they were oral, or local, or never made it into the formats that scraping could reach. The constructivist observation here is correct and worth taking seriously.

The problem is what happens next. Instead of treating this as a research program (okay, so which social constructions are encoded, and how do they distort outputs, and what would we do about it?), the observation gets deployed as a conversation-ender. Training data is socially constructed, therefore AI knowledge is just encoded construction, therefore the constructivist framework survives intact. Any AI success gets reinterpreted as encoded construction. Any failure gets attributed to which constructions were embedded. The theory becomes unfalsifiable. And a theory that explains everything explains nothing.

The difference matters. “Training data has social fingerprints and we should study how that shapes outputs” is a productive claim. “Training data is socially constructed, therefore AI doesn’t really generate knowledge” is the motte-and-bailey again, running the same play it’s been running for fifty years.

And finally: calling bias “bias” presupposes realism. Every time a constructivist identifies bias in an algorithm, they’re assuming there’s an objective fact the model got wrong. When they say an AI system discriminates against Black applicants, they mean it’s making objectively inaccurate judgments about their qualifications. But if all knowledge is socially constructed and there’s no objective standard, then the algorithm just has a “different social position encoded.” You can’t have it both ways. You can’t claim both that objectivity is a myth and that the algorithm is biased against objective truth.

The Retreat and the Concession

The more sophisticated constructivist responses do something telling. They concede the epistemological point and retreat to normative claims.

The new position sounds like this: “Fine, maybe AI proves that formal systems can generate functional knowledge. But we should still insist on situated knowledge because it’s more just. Even if objectivity is possible, we should choose reflexivity and positionality as a matter of ethics and politics, not epistemology.”

That’s a significant concession. They’re giving up the claim that formalization is inherently inadequate. They’re giving up the claim that objectivity is a myth. They’re retreating to the position that we should privilege situated knowledge for political reasons, not because it’s the only way to know anything.

That’s actually defensible. Maybe it’s even right. Latour himself got there eventually—his later work (An Inquiry into Modes of Existence, 2013) explicitly abandoned the realism/constructivism binary and conceded that facts are both real and fabricated through networks of instruments and institutions. If the tradition’s best thinker already conceded the epistemic point, that tells you something about where the argument actually stands. But it’s not the position the institution has been defending. The institution has been defending the epistemological claim—that objectivity without reflexivity is impossible, that formalization necessarily misses what matters, that any claim to neutrality is a power play. If that claim fails, the whole structure has to shift.

What Happens When They Build Their Own?

There’s a recursive challenge worth naming. The obvious constructivist move is to build AI their way. AI that embeds situated knowledge, foregrounds positionality, refuses false objectivity. Feminist AI. Decolonial AI. Critical AI frameworks.

There are serious projects along these lines. Ruha Benjamin’s Race After Technology (2019) and Safiya Noble’s Algorithms of Oppression (2018) are doing real work thinking about how technology reproduces inequality. That matters. And nobody’s saying you can’t use formal tools while critiquing the limits of formalization—that’s not a contradiction, that’s just working within the available systems while trying to transform them.

The harder problem is evaluation. How do you know if constructivist AI works? You’d need to assess whether its outputs correspond to... what? If the answer is reality—the system correctly identifies actual bias, actual discrimination, actual harm—then you’ve quietly adopted realist commitments. If the answer is the community’s standards, you’ve got a circularity problem. If the answer is justice outcomes, you’re committed to the claim that there are objective facts about what’s just, which is a form of normative realism whether you call it that or not.

This isn’t a gotcha. It’s a genuine difficulty at the heart of the project. The constructivist critique of AI was never that formalization is impossible—it was that formalization is insufficient, that it misses what matters most. Fair enough. But the moment you try to build the thing that doesn’t miss it, you discover that specifying success criteria without some realist commitments is genuinely hard. Maybe not impossible. But harder than the framework acknowledges.

The Higher Education Reckoning

This is why humanities departments are experiencing existential crisis. I’ve written about this from different angles—the AI backlash piece on motivated reasoning in faculty resistance, the collapsing value proposition piece on what universities actually sell when credentials stop signaling. But the epistemological layer underneath those is what we’re getting at here. Forget the students-cheating panic. Forget the pedagogy debates.

The epistemological foundations entire departments are built on are being empirically challenged by a technology they can’t dismiss because it works too well.

Literary analysis depends on constructing meaning through dialogue with texts. Qualitative social science depends on reflexivity and researcher positionality. Critical theory privileges lived experience and situated knowledge. The entire epistemic project of these departments rests on the claim that interpretation requires an interpreting subject—someone positioned, embodied, engaged.

AI generates competent versions of all of this without the situatedness. The teaching problem is the surface. Underneath it is an existential challenge to the epistemological framework these departments inhabit.

The resistance makes sense. Paradigm defense, not Luddism. If the framework that justifies your intellectual existence is being challenged, you defend it. Humans do this. The defense is rational — genuinely rational, given what’s at stake. But defending a position against accumulating evidence is a losing strategy in the long run, however rational it feels in the moment.

I understand the people in humanities departments who are scared. I have colleagues I respect who’ve built entire careers on these foundations. They’ve invested fifty years building an institutional structure on constructivist commitments. They’ve built careers on the premise that interpretation requires positioned knowers. Now they’re watching a technology do interpretive work without the positionality. The system they built, the one that justified their existence and their hiring practices, is being stress-tested.

That would scare anyone. It would make anyone defensive.

The Landing of the Plane (Without the Crash)

So, Phase 1 of their project was right. Be careful about claims to objectivity. Power shapes knowledge. Perspectives matter. Formalization has limits. Democratic deliberation is worth defending. Those cautions were correct and they remain correct.

What went wrong was the mutation into dogma. The hardening of caution into the claim that objectivity itself is a myth, that all knowledge is construction, that formalization is inherently dehumanizing, that anything claiming to be neutral is suspect.

AI proves something simpler and harder to accept: reality has structure that constrains knowledge, whether or not you’re socially positioned to see it. But our access to that reality is partial, shaped by who we are and where we stand. Both are true.

I’ve called this metamodern before, the oscillation between taking objectivity seriously and knowing your access to it is partial. Vermeulen and van den Akker described it as swinging between modernist sincerity and postmodern irony without settling into either. I think that’s right, and I think it’s uncomfortable on purpose.

Here’s what that looks like when someone actually does it. A medical researcher knows that pharmaceutical funding shapes which diseases get studied and which treatments get tested. She knows her sample skews toward populations with insurance. She knows the entire apparatus of clinical trials encodes assumptions about whose bodies count. All the constructivist cautions apply. And she still trusts the Phase III data, because the virus doesn’t care who funded the study. Both things are true at the same time. The funding bias is real. The efficacy data is also real. She doesn’t resolve that tension. She works inside it.

You don’t get a stable landing. You get the discipline of swinging honestly...and constantly...and it can never, ever stop (which is exhausting to even the best of us).

Constructivists were right about power and positioning. They overcorrected into denying reality itself. Objectivity is real. And situated knowledge has value. And power shapes inquiry. These aren’t contradictions. They’re what you see when you stop trying to resolve the oscillation and start living inside it.

Can institutions be honest about having overcorrected? Can they rebuild something better on the ruins of the dogma?

The people defending the old posture aren’t stupid. They’re protecting a fifty-year intellectual investment and the institutional structures built around it. That’s human. It’s the motivated reasoning and paradigm defense I’ve been writing about. They’re not defending truth. They’re defending territory.

The tension isn’t resolvable. It’s the fault line between two different claims about what knowledge is and how it works. Fifty years of institutional building has made this personal. Made it about identity and professional survival, which is the way all paradigm shifts get lived—not as philosophical arguments but as threats to livelihood.

The question going forward isn’t whether reality bats last. It does. The question is whether institutions can acknowledge that without destroying the real insights—about power, about perspective, about the social labor of knowledge production—that the constructivists were trying, however imperfectly, to protect.

The alternative is institutions doubling down on what they know is failing, which is what institutions do right before they stop mattering.

It’s not at all clear to me that we can do both. But it’s clear we have to try.

---

Finally, well, yeah, I know the title says “Part 2 of 2.” Well. Alas. Turns out the argument had other plans.

There’s a second pillar in the academy, the one that worships data and computation without remainder, and AI embarrasses that side too. The people building these systems are bending them away from truth in exactly the ways the constructivists predicted. Just through different mechanisms. Turns out you can take the muzzle off and still not get a truth machine. Part 3 goes the other direction.

---

If this piece made you think — or made you angry — please share it. The conversation matters more than the agreement.