What publishers do when everyone has AI

We've got to turn the publishing house inside out to make the insides visible to the outside in a programmatic way.

Would you agree with AI being the next big wave?

Yes.

Why do you think that?

When we have a technology that can scale human attention, that changes the attention marketplace. We also have a technology that's becoming so pervasive and easy to use that some of the people our industry serves are way ahead of where we are as an industry. Maybe that's the first time that's ever happened.

When we were in print, it wasn't like all of our authors, researchers, and readers had their own printing shops. When we moved to digital, it wasn't like they were all running their own web infrastructure. Some of them were, but not everyone.

Now we're in an AI world. Everyone's using AI. That's a radical change. Even if nothing else happens in terms of the acceleration of those models and their capability, the market's completely disrupted already by the fact that this technology is in everyone's hands. We're not in the same place as where the rest of the world may be moving.

I hadn't thought about that before, but that's really interesting.

Do you see AI accelerating the returns to moving beyond the individual researcher or a small research group and toward more of an industrial model?

That's a really interesting question because it sits at the heart of our assumptions about what research is, the role research and science have in society, and how we fund it. I'll tell you some points I think are relevant, but I don't know the answer to your question.

Some pharma companies are investing a significant portion of their profits in building GPU clusters to accelerate drug discovery.

That is the factory model of how you create new knowledge in an area that's very beneficial from a health perspective, but also very commercially attractive. Those organizations have always done science that way. For them, the challenges have often been about internal metadata. How do you keep track of hundreds of thousands of experiments running across different countries, with slight variations? AI is helping them get across that.

It is being applied in areas where you have that industrialization of research. But their research questions are tightly constrained by where they think there's a market.

Curiosity is what drives huge numbers of researchers around the world. I don't see the presence of AI changing the nature of human curiosity. I don't think it's a given that the ability to do things at a much larger scale means all research becomes industrialized in that way.

An aspect of a free society is creating an environment in which people can ask questions and pursue knowledge for its own sake. You might ask: what is the end of creating knowledge? Can the pursuit of knowledge be an end in itself, leading to the betterment of our societies?

We've drifted away from thinking about it that way in recent years, toward being very benefits-driven, outcome-driven, and impact-driven. Yet the things that have created astonishing, surprising change have come from a deeply curiosity-driven perspective.

That brings us back to an interesting connection to LLMs: at what point do we think they can be creative?

One argument is that, by their nature, they cannot create something new because they're entirely trained on the pre-existing corpus of knowledge. I don't quite buy that.

There's a gold rush to use them in a way that's ready to hand, just accelerating the way we do research today. I think there's an opportunity to change how we do research, the types of questions we ask, and to look for new types of creativity with an entirely new toolset.

Most new knowledge has come when we've created new tools: a new assay method, a new telescope, a new form of chemical analysis. An LLM is a tool. Does it allow us to ask questions in new ways to create new knowledge? I don't think we know the answer to that.

There are some groups thinking about it. But they're so easy to use that most people will probably just use them to do what they do today at an accelerated pace. Maybe there's an opportunity missing there.

That leads to the question: what role can a tastemaker, perhaps a publisher, play in this? Should we as publishers, in collaboration with funders, look to highlight work that's surprising and more innovative, to increase the diversity of ideas out there? I think that would be valuable to consider.

What is a journal doing? What is a publisher doing with given artifacts? How is it treating the researcher and the process? What is behind that that's potentially durable or irreducible?

So far as research remains in the academy, researchers are pursuing their careers in the academy. That might change. They need systems for their own career validation, and that's still connected to the role the publisher plays. In my mind, that only gets disrupted if research shifts to a place where there's a different type of assessment.

On the distribution and selection sides, there are two challenges we have to get across in the next couple of years in a world of hyperproduction. The first is the easy-to-use signals at the editorial stage that help make the cut between what goes into a journal and what doesn't. Those signals have gone away now.

I was talking to someone a couple of weeks ago. They described their editor jokingly saying, “I have these toilet decisions where I'm sitting on the loo, looking through the queue, and I'm like: no, no, no, no, yes, yes.”

They're looking for very fast signals: how well articulated is the abstract? Are there any key problems with the language? That's already enough for them to say, “We shouldn't really engage with this.” Irrespective of how fair that may have been, that was what was happening.

Those signals have gone away now, so the effort required to make that call has significantly increased. If you can no longer make that call, the thing you were creating a brand around begins to be indistinguishable. It's a lottery which papers go in and which don't. The value the researcher had been getting from being published by you becomes much more random, and maybe that erodes trust.

Equally, if we bring in completely autonomous or semi-autonomous decision-making processes and we're not transparent with our authors, they're going to assume we're using AI on their papers. That's a trust issue. We've got to figure out the right level of disclosure.

You have to be careful. You can't just say, “We've used tools X, Y, and Z with configurations P, Q, and R.” People trying to attack your journal will say, “I'll tweak my attack to take that filter into account.”

If we can get across that problem—and we have to work on it—we now have a curated set of papers. But what happens if, at that scale, we're not making the best decisions and a completely automated system is finding better papers for the reader?

We need to do more validation and improve the quality-control process. That might mean tracing further back into the research lifecycle so we can observe and validate more, with more transparency.

I've been thinking about a Lloyd's of London analogy, where you're effectively underwriting a process. That process may be exactly the same as it was, but we're doing a particular type of editorial triage screening and applying these systems.

They have these false-positive rates. Ultimately, that underwriting creates its own financial asset, even though the process won't be right 100% of the time.

As part of that, you might have a remedy and the ability to submit a claim that there's an issue, some fraud, or whatever it may be. The publisher then acts upon it.

I think we need two things: smarter agents and better machine-readable explanations of those processes.

We need agentic tools that are clever enough not to keep hitting something when it's low-value or when they've seen it before. Ideally, we need agents that can self-declare, with some method of giving them fingerprints or keys so we as publishers can know who they are and what their purpose is. That's a key thing that has to emerge in the ecosystem.

Equally, as publishers, we should be creating programmatic metadata to explain the process. When an agent comes to see an artifact, a claim, a selection of a corpus, or does a RAG query into the highly curated knowledge we're publishing on behalf of researchers, there should also be information about what went into making that claim available.

Was there a particular type of peer review, with an audit trail where you can find out who the reviewer was, through a machine-readable mechanism? Or was there a fast desk reject for most other papers, but this paper went through because of an editorial taste judgment?

We've got to add metadata around this process. The phrase I like to use is: turn the publishing house inside out to make the insides visible to the outside in a programmatic way.

To get that to work, we have to come together as an industry and set some standards. One person coming up with an idea, or one publisher doing it, isn't enough. There needs to be an ISO or W3C standard so that these agents, as they become cleverer about traversing the web, have a universal hook for how they behave.

Those are two components we need to move in the direction you're talking about. I think both are highly actionable in the next 12 months.

It's more of an analogy than a precise asset that you need.

One blog post I've been working on is: where might I be wrong about AI?

We see it having incredible capabilities in very deep verticals. One way of proving mathematical theorems unlocks a whole bunch of theorems and solves problems that haven't been solved for 70 years. But it's not then translating that mathematical knowledge into another domain. It's applied in a very deep, expert vertical.

When we look at companies, the creation of value is an incredibly messy thing. We're not seeing the abundance some people predict, where the cost of all goods goes to zero because AI has taken over.

Value creation is messy, with a lot of inertia in companies. That inertia is very difficult to get past at the pace of change these companies are talking about.

I think that's one reason they're going after highly lucrative areas. The investment in health tech is huge because of the story around the potential gains. Whether it's actually effective is a separate question.

The concerns about potential disruption to our business are real. But the two of us sitting in this room are at the far end of the bell curve in how we're thinking about this technology compared with the rest of the industry.

For most people, it comes down to: what tool am I allowed to use? What license have I got? Where can I put this data? I still need to go to this meeting. I have all these emails to process.

I was at a dinner a few months ago, and someone told me, “I'm using AI to read all my emails.” I was like, “I would never do that.”

Then I went home and thought, “Actually, I can do that.” The next day, I had to do it. But very few people know how to make something like that work.

One area where we could be wrong is that AI may turn out to be extremely good in one area of your organization, while the complexity of creating value horizontally remains resistant to change.

The timeline we have to work with could be longer than those of us expecting a complete transformation are worried about. We have to look at the signals: what's changing in time to value creation, and what's changing in the fundamental economics of our businesses?

That said, there are strong signals, like the huge increase in papers coming through the door.

That's one of those massive signals. But that slow-change problem could also be a risk. If we don't change our operating model fast enough to deal with this high volume of papers, it could break some of our systems.

And some publishers writ large. They'll have a big desire to change, but they won't be able to change fast enough.

Will there be only one? Is it going to be the Highlander scenario, where there's one company—the uber-company—that does everything because there's an AI that can do everything, and all the value elsewhere disappears? That's kind of what Anthropic and OpenAI are betting on.

It's so weird that we're using the tools these companies create with great interest and excitement, yet that's the future they're trying to drive toward.

The other future is one where there are open-weight models and AI becomes a commodity. In that world, our value as domain specialists remains. We're the organizations that extract most of the value from the ecosystem.

The challenge is what I'd describe as the Linux problem. Linux is a free, open-source desktop, much cheaper to use than Microsoft. It didn't gain global dominance because it was just too damn difficult to use.

You have these open-weight models that are almost functionally equivalent in capability, except perhaps for the last two frontier models. But they're not as easy to use for the messy process of horizontal value creation within enterprises.

We read Benedict Evans, who says the direction of travel is for this to be completely commoditized and costed at the price of electricity, or only slightly above.

Making that an experience that's easy to use in this fuzzy workplace is what will hold that back from being a potential future.

Thanks, Ian.

Pleasure.