For the first time since I had been in Barcelona I went to have a look at the cathedral — a modern cathedral, and one of the most hideous buildings in the world. It has four crenellated spires exactly the shape of hock bottles. Unlike most of the churches in Barcelona it was not damaged during the revolution — it was spared because of its ‘artistic value’, people said. I think the Anarchists showed bad taste in not blowing it up when they had the chance, though they did hang a red and black banner between its spires.

— George Orwell, Homage to Catalonia (1938)

Orwell calls it “the cathedral”, but the building he describes is Gaudí’s Sagrada Família, which is technically a basilica — Barcelona’s cathedral is a medieval Gothic building.

To my surprise, the writing agent stepped back and returned the task to me: actually, I have to think this through, not just prose edits it can execute unilaterally. And the machine was right. The limitations section should be written by hand. It is essential to feel the writing frictions, to weigh up the word arefully, and to fully own the conclusion.

They were careless people, Tom and Daisy — they smashed up things and creatures and then retreated back into their money or their vast carelessness, and let other people clean up the mess they had made.

— F. Scott Fitzgerald, The Great Gatsby (1925), Ch. 9

Sarah Wynn-Williams borrowed that line for the title of her memoir, Careless People: A Cautionary Tale of Power, Greed, and Lost Idealism (2025). It is one of the better books I have read recently — a first-hand account of the culture inside Meta that is as readable as it is alarming.

The AI revolution genuinely excites me — it changes what we do, how we live, how we work. But the people currently shaping it matter as much as the technology itself. The culture Wynn-Williams describes at Facebook is not unique to Facebook but, presumably, hard-baked into the culture of the Silicon Valley. From what I can read in the media, OpenAI and other companies building the most capable AI systems resemble Meta closely: The unchecked power, the retreat into certainty, the willingness to let others clean up the mess. That is the part worth losing sleep over.

Two things tend to surprise people most. First: urban housing is actually less expensive today than it was 100 years ago, once you account for income growth. The story of a relentless affordability crisis looks very different when you zoom out far enough. That doesn’t mean there isn’t a real problem right now — there is — but the long-run trend is not the one most people assume.

Second: why do buyers consistently pay a premium for historical architectural styles? The instinctive answer is aesthetics — people just like the look of old buildings. The data suggest something else is going on. It seems to be less about the architecture itself and more about the signals that historical styles send: quality, durability, neighborhood stability. The style functions as a proxy for attributes that are hard to observe directly.

Both findings are good reminders of why historical data matter. A decade of observations can tell you a lot about cycles; a century of observations can tell you something about the underlying structure of markets.

The full interview is available on the Lincoln Institute website: Taking the Long View on Real Estate Investments

This year’s edition takes place on 12–13 September 2026 at Pembroke College, Cambridge.

The format is deliberately unhurried: each paper gets 45–50 minutes, with time for a formal discussant and open Q&A. Topics range across real estate finance and economics broadly — risk management, capital structure, sustainability, international investment, ML and AI in real estate, and more.

Submissions are due 1 June 2026. Around eight papers will be selected, with decisions by 15 June. If you have work in progress that fits, send it my way: htl24@cam.ac.uk.

The full call for papers has all the details.

A small writing change makes the necessary quality checks a bit easier for me:

In academic writing, it is uncommon to say “I”. The convention is “we”, even in sole-authored papers. In the age of AI-enabled research, I think that should change to “I”. When the text says, I did this. I tested that. I find this, it becomes clear who is responsible. The author, alone. No “we”. No hiding behind “Claude wrote this”. It is my text. When I claim authorship, I also own the errors and that’s why I cannot be complacent.

Yes, the text above sounds like AI slob. The short sentences. One line paragraphs. A good example of the indirect AI effect on our communication, as documented by Yakura et al. (2025).

The way I “do stuff” has changed tremendously. The long nights of getting the data ducks in a neat row are disappearing. Visualisations and interactive data explorations are so much easier now. I love those interactive HTML slides that let me quickly zoom into spatial estimation output, much faster than importing everything into GIS.

But the longer I work this way, the more I realise that the way I “think about stuff” has not fundamentally changed. A lot of the real work still happens away from the keyboard: while cycling into work, doing the dishes, sitting in a meeting, or mentally turning over a research step. Running empirical tests requires as many critical checks as before. We still need to think carefully about data characteristics and limitations, the data generation process, measurement, context, theory, and links to other sources.

Just taking the output from Claude Code, however smooth it looks, is a recipe for disaster. And that will not change, even with better models. Quality control is as important as ever.

To ensure quality, a researcher needs to know their empirics, the literature, the theory, and how real-world data were generated in the real world. AI promises the automation of many steps in the research process. But this does not mean an erosion of true research skill. On the contrary. The better the tools become, the more important it is to identify mistakes and to earn another “Good catch!” response from Claude Code. Maybe we are not redundant (yet).

For a writer, some AI help with ideas or outlines may be acceptable. For a researcher, I would draw the line more narrowly. The core contribution is not just the prose; it is the research question, framing, hypotheses, interpretation and judgement.

If you want AI to help write or polish an abstract, fine in my book. That is a summary of work already done. But use it to suggest the ideas? Not really.

In research, the ideas are the thing. The question, the angle, the theoretical move and the interpretation of evidence are what make the work yours.

I might change my view on this in the future.

AI is everywhere in academic research. Kobak et al. (2025, Science Advances) tracked words that language models overuse — “delve,” “nuanced,” “meticulous” — across 14 million biomedical abstracts and found at least 13.5% of 2024 papers were processed by an LLM. The same pattern shows up in the real estate literature, as a quick replication on 100K real estate papers indexed by OpenAlex shows.

That is the writing layer in the research process. The more consequential shift is deeper. A growing number of papers rely on AI not for drafting but for execution — work that could not exist without machine learning carrying out the core analysis. Bartik, Gupta and Milo (2025) read thousands of municipal zoning codes and built regulation measures that no research team could produce by hand. Calainho, van de Minne and Francke (2024) replaced linear hedonic models with ML on 30,000 New York transactions and showed systematic gains in out-of-sample accuracy. Shen and Ross (2021) extracted a description-quality measure from MLS listing text that captures soft information about property quality invisible to structured data. Leow and Lindenthal (2025) applied the Gu-Kelly-Xiu ML asset-pricing framework to REIT factor returns and showed substantial forecast improvements over OLS.

In each case, AI enables a measurement or prediction the research requires. Remove it and the paper disappears. But the role is still that of a skilled research assistant (RA): executing tasks specified by a human. The principal investigator (PI) — the person deciding what to study and why — remains human.

AI outperforms human researcher in many dimension (speaking for myself, obviously). The question is whether it can shine higher up the value chain. Can LLMs suggest research topics that are genuinely innovative and plausibly doable — functioning more as a PI than as an RA? Do humans still have a competitive edge?

The new working paper tests this. I mapped the full published corpus of Real Estate Economics (1,676 articles, 1973–2026) and real-estate-relevant subsets of JREFE, JUE, AER, JF, and RFS into a shared semantic embedding space. The result is a coordinate system for the field — not a literature review, but a map. Against that map, I generated 1,499 research ideas under eight conditions, varying what context the model received: nothing, the full corpus, individual cluster seeds, methods borrowed from economics and finance, methods from psychology. Each idea was scored on atypicality (a measure of unusual knowledge combination that retroactively predicts citations) and mapped back into the research space.

The figure below shows where generated ideas land. Grey dots are the full corpus; blue dots are REE papers; red dots are AI-generated ideas. Condition A is naïve generation from training data alone. Condition F draws on methods and paradigms from economics and finance journals.

Methodological scaffolding moves ideas outward into less-explored territory. Topical scaffolding alone does not. The best ideas — particularly those generated through method transfer from economics and finance — score comparably to the median published paper on the citation-predictive criterion. Some land squarely on papers published twenty years ago, having rediscovered questions the field already answered. But that is also true of human research proposals.

There is an uncomfortable regularity in how AI gets adopted: if a system offers a plausible-looking shortcut for a task it was never designed for, people will happily use it anyway, and it takes a lot of effort to convince them of the limits. Researchers will use LLMs to generate research ideas. They already do. The useful question is not whether this is a misguided idea but what the machines actually serve them when they try — and under what conditions the output is worth anything. That is what this paper is about.

Working paper (PDF) — Interactive idea explorer

At yesterday’s ECREDA conference in London, I tried to triangulate exactly this. Using real estate research as a testing ground, I explored how LLMs perform on research idea generation — varying domain knowledge and constraints, then scoring ideas for novelty and predicted citation impact. AI can expand the frontier of what gets considered. But the gym still matters.

Slides are available here.