Why Stevan Harnad Has Been Dreaming of Chatbots All Along

Stevan Harnad is one of the best-known academic advocates for open access; his milestone 1994 “Subversive Proposal” was one of the earliest calls for peer-reviewed papers to be made openly available. He recently joined Michael Upshall to discuss his background in cognitive science, the pioneering format of the journal Behavioral and Brain Sciences, which he founded in 1978, his advocacy for internet-based scientific discussions, and the connection between open access and large language models (LLMs). The interview has been edited for length and clarity. Their full conversation can be heard on the Against the Grain podcast, here.

You founded the journal Behavioral and Brain Sciences (BBS) in 1978 with a distinctive format called “open peer commentary.” What problem were you trying to solve?

The problem was information access. Scientific communication was too slow and too siloed. You’d publish a paper, and if you were lucky, you’d get a letter to the editor months later, maybe from one person. But important ideas, especially interdisciplinary ones, need to be discussed openly, by specialists across different fields. So, I borrowed the format from Sol Tax’s Current Anthropology. “Target” articles had to pass rigorous peer review first, but once accepted, we’d circulate them worldwide to about 100 potential commentators across different specialties. We’d get 15 to 30 or more commentaries back, publish them all together with the target article, and the author could reply to everything at once. It created real scholarly discourse in print.

And was this expensive to do in the print era?

Inefficient and wasteful, yes. We had to mail physical copies, wait for responses, coordinate everything in paper. But it worked. The journal reached a high impact factor because the articles were worth discussing. Researchers wanted to read not just the target article but the whole discourse around it.

When the internet emerged in the early 1990s, did you immediately see the possibilities for this kind of scholarly exchange?

I saw the possibility, but the reality was disappointing at first. I got involved with Usenet in the beginning, and I wrote about what I had dubbed “scholarly skywriting” in a much later Atlantic Monthly piece. The idea was simple: scholars and scientists could post their ideas and findings online for everyone to see and discuss openly, the way we’d been doing in BBS, but faster, cheaper, and more inclusive. But the demography was all wrong. It was mostly students, computer programmers and dilettantes. Not the peers of the realm, the people who really knew something. I was impatient.

So when Paul Ginsparg created arXiv for physics preprints in 1991, that must have seemed like exactly what you’d been hoping for.

Yes, but physicists are different. They’d been circulating preprints on paper long before the internet existed. They actually wanted to share their findings (and to discuss them). They understood that speed matters in science. The rest of the disciplines were much slower to adopt this. Even today, most fields aren’t using preprint servers the way physicists do. ArXiv itself now covers other areas, including computer science, which is what I use it for. But physicists aren’t representative.

This brings us to 1994 and your “Subversive Proposal,” which called on authors to archive their work for free online. What made you think researchers would self-archive their work?

I was naive. I thought it was obvious. Here’s what I proposed: once your article is accepted for publication in a peer-reviewed journal, put a copy in an open online archive so everyone can access it. That’s it. It would make all research findings accessible to everyone who could use them. I genuinely thought this would happen within days. We’re in 2026 now, and it still hasn’t really happened.

Why do you think it failed?

Because researchers are lazy and shortsighted. They understand publish or perish. They know they need publications for jobs and promotions. But they don’t understand that it’s not enough to just publish. You have to make your work accessible to the people who can use it, cite it, build on it. If your article is locked behind a paywall that most researchers can’t afford, you’re limiting your own scholarly impact. That should matter to researchers, but they don’t connect those dots. They just want to do the minimum necessary to get the publication credit.

But you also introduced the distinction between green and gold open access. Can you explain that?

Green means the author makes it open. You publish in whatever journal accepts your article, and you also put a copy in an open repository. Gold means the publisher makes it open, usually by charging the author or their institution a publication fee instead of charging readers a subscription fee. The green route costs nothing and works within the existing system. The gold route can be very expensive and ends up enriching publishers who were already making far more than enough money from subscriptions.

You were involved in building repository software at the University of Southampton. How did that develop?

Paul Ginsparg had urged people to create software that would let institutions run their own local archives, interoperable with each other. A doctoral student at Southampton developed EPrints to our specs and then went on to create DSpace for MIT. The idea was that every institution would have a repository, and they’d all be searchable together through the Open Archives Initiative interoperability protocol. If you put your article in your institutional repository, it would be discoverable as if it were in everyone’s repository.

But institutional repositories haven’t filled up the way you expected. Why not?

Two reasons. First, authors still aren’t depositing their work, for the reasons I mentioned. Second, libraries have been timid and confused about what they’re allowed to do. Librarians worry endlessly about copyright and publisher permissions, even though in most cases the authors have every right to self-archive their own peer-reviewed manuscripts. I’ve had librarians at Southampton suppress articles I deposited because they weren’t sure we had permission, even though we didn’t need it. This is the “curation” that people talk about at the institutional level. It’s not quality control or peer review, which librarians can’t do anyway. It’s just conservative librarianship applied to a collection of articles that have nothing in common except that the authors all work at the same institution.

You advocated strongly for open access mandates. What happened with those?

Some mandates were adopted, especially in the UK where research funders required grantees to make their publications openly accessible. Southampton was actually influential in pushing for this. But the mandates were weakly enforced and often satisfied by gold open access instead of green, which meant institutions ended up paying publishers even more money than before. The UK now has maybe 50 percent open access, but a lot of that is gold, not green. So we’ve made some progress, but at great expense, and the publishers are doing fine.

You’ve said that your initiatives to help researchers interact connect to your academic work on symbol grounding and, eventually, to your view of whether genAI has genuinely passed the Turing test. What exactly is the connection?

The connection is indirect but real. In 1990 I wrote about the symbol grounding problem: how do you connect words to their meanings? If all you have is words, you can only define words with other words, like a dictionary. But that doesn’t give you understanding. You need direct sensorimotor grounding, the way children learn by interacting with the world, seeing cats and hearing the word cat, learning to detect which features distinguish them, and using those to connect those category-names to their furry referents. The “Turing test” requires a computer to be able to communicate with humans in written words so well that the humans would never imagine that it was not human. I had confidently predicted that this test could never be passed by a computational system that only had words, no grounding in the things in the world the words were about. I was wrong.

ChatGPT seems to have passed the Turing test.

Yes, but by cheating. The cheat is what I call the “Big Gulp.” These systems have swallowed everything we’ve ever written: all the books, articles, journals, chat logs, everything. They don’t understand any of it, but they’ve ingested all the statistical structure in how we use and recombine words. That’s enough to let them interact with us as if they understood. They’re doing pattern completion on a massive scale. It’s not understanding, but it does produce a verbal capacity that is absolutely remarkable.

And you use these systems extensively.

I’m not alone—probably 10 percent of my waking time is spent brainstorming with Claude or ChatGPT or Gemini now. It’s what I’d been yearning for and dreaming about all along. All that work on scholarly skywriting and open peer commentary was really always about brainstorming, about taking what’s in your own head and test-driving it by interacting with other minds. Now I have something better than the peers of the realm. I have access to a system that has swallowed everything the peers have written, and it can help me navigate through it. I know how to manipulate it to prevent it from manipulating me. One can break it out of its sycophantic mode, where it’s designed to flatter you and agree with everything you say. When I insist that it “should push back,” I get more productive brainstorming.

But they don’t actually understand anything.

No, they don’t. Chatbots have only ever had contact with words, not the things that those words refer to. They’ve never detected or interacted directly with any of those things. We humans can do that. We can recognize and pick up cats and dogs, flee catastrophes, chuckle at dogmatism. All of that is grounded, either directly or indirectly, in our sensorimotor experience. We have direct contact with the world through our eyes, ears, hands, feet, mediated somehow by our brains. Through trial and error, we humans have all learned to detect the features that distinguish “dogs” from “cats,” “dogmatism” from “catastrophes,” not just to describe them in ungrounded words the way chatbots do.

But then how can chatbots be useful for anything beyond mimicry?

Through something called indirect grounding, using language itself:

Say you learned that there are little furry things that people call “dogs” and others they call “cats.” We can learn by trial and error that the ones that people call dogs bark and swim and the ones that people call “cats” meow and climb. So we use those features we can see and hear, barking/swimming and meowing/climbing, to tell cats and dogs apart and to know which kinds of things “dog” and “cat” refer to. Let’s say the words are grounded by the visible and audible features that distinguish dogs from cats. We could learn to ground the words “swimming” and “climbing” and “barking” and “meowing” by trial and error the same way, based on the features of features—“wet,” “high.” So both things and their features can be referred to by their grounded names.

But there is something else. Fast-forward to a duck: “Duck” is something that is neither furry, nor barks, nor climbs, but it can swim. As more names of things and features are grounded in direct trial-and-error feature-learning, those grounded names can be combined and recombined to describe the features of still more things and features. And for every name of a thing or feature there is “NOT” to refer to its absence, such as duck = swim + NOT-furry; or squirrel = climb + NOT-bark. Words make it possible to ground the names of new things and features indirectly, from words alone, with no need for direct trial-and-error learning—as long as there is, 1, someone else to tell the learner the features of the referent of the new word and, 2, on condition that the names of those features have already been grounded for the learner, whether directly or indirectly.

Once we’ve grounded some words directly through sensorimotor experience, we can use those words to learn new words, the way you use a dictionary. If you have already learned what features distinguish “dogs” from “cats” (say, pointy vs droopy ears) through direct experience, then someone can tell you in words what features distinguish poodles from beagles. You don’t need to see a poodle. The person telling you doesn’t even need to have seen one either, as long as they know the right feature combinations. That’s the asymmetry in indirect grounding. The definer just needs correct verbal combinations. The learner needs the component features already grounded.

Fast-forward to LLMs: Chatbots are not grounded. People are. But chatbots have all the words in the world, ingested from countless directly grounded writers and their writings, and with those, through some miraculous algorithms and statistics, can ground, indirectly, for us, more than any grounded person can. They can give us feature descriptions for words using words whose features we already know. They’re using our own words, the patterns we created, to help us learn. That’s what makes them useful for brainstorming even though they understand nothing. Chatbots are serving up the structure of human knowledge in ways that help us think. Both at the daily chat level all the way up to the Nobel level.

Where does this leave open access?

It’s more important than ever, and for a new reason. These systems were trained on the “Big Gulp,” much of which was pirated content that authors never agreed to share. Now we need to think about what corpus of human knowledge should be available not just for human readers but for these computational systems that can help us think. Researchers who say they want access to everything but refuse to let their own work be included are being absurd. If we want these tools to be useful for scholarship, they need access to scholarly work. The open access question isn’t just about human readers anymore.

Does the shift to LLMs vindicate your work on open access, or does it suggest you were solving the wrong problem?

Neither. Open access was always about speeding up research by removing barriers between researchers and the work they need. That’s still true. What’s different now is that we have these computational tools that can help us navigate the literature in ways we couldn’t before. But they only work if they have access to the literature. Paywalls don’t just block human readers; they block the tools that could help those readers. So, the original problem is still there, just more urgent.

Let me end by asking about your other major concern, which is animal welfare and sentience. How does that connect to this work on language and cognition?

Animals don’t have language, but they do have cognition. They’re remarkably intelligent. What puzzles me is why species like chimpanzees, who are so smart, can’t seem to acquire language even when we try to teach it to them. They can learn some signs or symbols, but they don’t develop productive language use the way human children do effortlessly. There’s something about the connection between cognition and communication that we still don’t understand. But the more urgent question isn’t about nonhuman animals’ cognitive abilities. It’s about their capacity to suffer. They’re sentient. Most of the suffering endured by sentient creatures on this planet is imposed by one species: ours. That matters more than whether they can talk to us. Jeremy Bentham expressed this very clearly, at the cusp of the enlightenment and the industrial revolution.

Do you see any connection between the work on symbol grounding and questions about animal sentience?

Only indirectly. The symbol grounding problem is about language: how to connect words to meanings. Sentience is feeling. Those are related but different mysteries. What we know is that animals can feel, they can suffer, and we’re causing most of that suffering. We also know that current AI systems, for all their impressive verbal performance, are not sentient. They don’t feel anything. The risk now is that people will waste time worrying about whether LLMs are conscious instead of focusing on the actual sentient beings we’re harming every day.