Ouch. That's a way to tell me something without telling me something.
Clearly the confidence comes across whether intentional or not, even in my own home, and it doesn't always stand the test of time.
So, let's flip that. Today let's talk about everything I've gotten wrong this year. Well, not everything, we have to respect the word count. Let's cap it at AI.
The list filled in faster than I would like to admit, and the thread running through all of it is that I kept overfitting to early signals, landing on confident views in places where the evidence did not really exist yet. In nearly every case, what corrected me was our own research or a colleague pointing at something I had walked straight past.
I assumed researchers were on a path toward better tooling
Much of what gets written about AI and research assumes that adoption improves as it matures. Institutional licenses arrive, entitlement-aware integrations get built, and researchers move off consumer chatbots onto systems made for scholarly work, with authentication and audit trails intact.I had no particular evidence for that when I started saying it. It just sounded like how things would go.
Research we conducted this year points somewhere else. Across experience levels, from graduate students through established principal investigators, researchers are working overwhelmingly in free or personal-account versions of general-purpose AI tools. The enterprise and institutional deployments, where the access controls and privacy protections live, are not where most of the work is happening. Seniority did not predict sophistication, and neither did discipline. I had assumed both would.
The pattern is not unique to research. LayerX, drawing on browser telemetry across its enterprise customers, found that 47% of enterprise AI conversations run through personal identities rather than corporate-managed accounts. More than 60% of ChatGPT and Claude conversations happen on non-corporate accounts, while enterprise-oriented deployments like Copilot M365 and Gemini Enterprise run over 90% through corporate identities. People reach for the version they already have, which is the least surprising finding imaginable once somebody puts it in front of you.
I also had the causation backward. I thought researcher adoption would pull governed infrastructure forward. It looks more like infrastructure has to get good enough to make the governed path the easier choice, and until then convenience wins. So the useful work over the next year will be making the well-governed option good enough that choosing it stops feeling like a tradeoff.
I mistook the room I was in for the wider field
This is the one I am least comfortable writing down.Spend a year in this industry's lecture halls, conference rooms, and webinars and you will hear a great deal about Model Context Protocol. It comes up constantly. I absorbed that and started describing MCP as something that had arrived, without ever checking whether the people I was hearing from were representative of the industry as a whole.
The challenge is that the people in those rooms are publishing professionals, technologists, and the researchers whose work sits close to publishing operations: people who hear about this daily and in many cases have a professional stake in it. I was sampling the most interested population available and treating its fluency as a read on the field.
Even inside that room, though, the fluency is thinner than the volume implies. In a webinar my colleague Stephanie Lovegrove Hansen hosted in May, 47% of the publisher audience said they had heard of MCP but could not clearly explain how it works, 27% had not encountered the term at all, and only 9% were actively evaluating or implementing an integration. And when we talked with researchers, authors, and reviewers, the people who keep the scholarly record running without needing to think about how it gets transmitted, awareness and use of MCP were minimal.
My first read was that researchers were behind. Wrong again, and it took a colleague to show me why. Our VP of AI, Jeremy Little, has argued in that same webinar series that MCP adoption depends on three kinds of readiness: the technology, the providers building endpoints, and the users themselves. Of those, he thinks user readiness is the smallest obstacle, because researchers are already using AI tools for exactly the work MCP is meant to improve. They do not need to learn a protocol, they need the results of one. Which means researcher awareness of MCP was never the right thing to measure. Plumbing that requires the user to understand it is not finished plumbing, and if we get this right, most researchers shouldn't ever need to learn the acronym.
I gave advice that was less actionable than I thought
In May I argued that publishers and societies evaluating MCP partnerships should be asking harder questions about mechanics: who controls the corpus, who controls the ranking logic, what the reporting measures, and what happens to entitlement infrastructure when a new intermediary sits between the researcher and the content. I would ask most of those questions again. What I would change is that I presented them as matters for the industry to consider, in the deliberate way scholarly publishing has historically approached consequential infrastructure.Over the past several months, capital has moved quickly into the layer between publisher content and AI systems, including investment flowing directly from AI platforms into companies building licensing, metering, and access controls on the publisher side. Those arrangements are being designed and priced now, in individual commercial negotiations, ahead of any shared framework for evaluating them. The answers are getting written deal by deal, and terms set by early movers are likely to become the terms everyone else inherits.
I flagged that risk in the same piece, in the context of MCP access getting bundled in as a throw-in rather than priced as a capability of its own, and then wrote as though there would be a window to deliberate before it took hold. There was not. What I should have written is that these questions belong in the next contract conversation, not in a working group.
I thought the uncertainty was a phase
Coming into this work, I assumed the fog would lift. Standards would mature, protocol questions would resolve, business models would settle, and we would move from improvisation into something closer to a plan. I built expectations around that, and it has not happened.A year in, the fog is starting to look like it might be the climate rather than the weather.
The July MCP specification release candidate was the largest revision since the protocol launched, taking MCP fully stateless, with real implications for anyone whose infrastructure assumed otherwise. That arrived roughly ten weeks after I published my piece about MCP mechanics.
What has changed is my sense of how long any single piece of analysis should be expected to hold. The commitments that make sense are different too: shorter horizons, and choices built to survive being wrong. That is a large part of why we run the AI Lab the way we do, putting prototypes in front of clients early and learning in the open rather than waiting for certainty that is not coming.
I have stopped treating the discomfort of that as a problem to be solved. It is the condition of the work right now, and this community handles it better than it gives itself credit for. Scholarly publishing has always brought real care to the infrastructure of knowledge, and that care has never depended on certainty. It depends on continuing to ask, and on sharing what we are learning while we are still learning it, including the parts that need correcting three months later.
The sign in my office is staying up. My kids had me pegged, and I've stopped reading it as a boast. "I doubt it" is a posture, not an assessment. We're an evidence-based community, and on the subject of AI, we haven't reached the threshold we'd demand anywhere else. Something to consider next time I take to the keyboard.