Before entering the legal research industry, I worked as a professional photographer for years, back when photographs were still shot on film. My assignments ranged from fashion catalogs and runway shows to magazine features and commercial work.
People often mistook photography skills for simply pressing the shutter. What they saw was the final, finished image, not the contact sheet—the page of small prints showing every frame from the roll, including the blurry focus, awkward expressions, bad timing, and shots with nothing particularly interesting to say. Pressing the shutter was the easy part. The real work was in the edit: deciding what actually held up, what honestly reflected what had been captured, and what belonged on the cutting-room floor.
There’s something similar going on with how people think about AI and legal industry research.
As legal news and LinkedIn posts continue to show us all, most warnings about AI and legal research focus on hallucinated cases, fake citations, and lawyers sanctioned for trusting an answer without checking it. This remains a real problem. But it’s not the problem I run into most often.
I research the business of law: how departments operate and buy services, how firms manage and price their work, what legal technology and AI actually do, and where the market is heading.
That kind of research has a different AI problem. Not hallucination, but of making weak claims sound stronger than they are.
A law department says it’s using AI. Does that mean a pilot, a sanctioned tool, a workflow change, or something fully operational? A law firm says it has a pricing function. Does that mean one person with a spreadsheet, a formal pricing team, a real process, or partner-by-partner negotiation with better reporting around it? A legaltech company says it has an AI platform. Is that proprietary technology, a licensed tool, a services layer, or a mix of all three? Those distinctions matter, and they are exactly where AI is dangerous, since it often doesn’t differentiate between these situations
Of course, AI can help with the early parts of research. It can organize notes, summarize long reports, compare public descriptions, clean up interview themes, and help test whether an argument is clear. We use it for all those things. But that first layer is not the research. Whatever AI turns up, we treat it as a lead, not a finding.
The real work is deciding what the information means, how strong the support is, and how much confidence the reader should place in the conclusion. That is where AI still needs close supervision.
Sometimes the strongest research finding is not a bold conclusion. Sometimes the right answer is that a claim could not be verified, the evidence is mixed, or the available information does not support stronger conclusions. AI does not naturally like those answers. Given a choice, it tends toward confidently wrong: It wants to finish the thought and hand you a clean, polished paragraph.
The polish itself is the danger.
A rough research note usually shows its uncertainty: the gaps (sometimes intentional), the thin sourcing, the claims still to be checked. A polished AI paragraph hides all of that. It can pass off a company’s positioning as independent fact, turn a few examples into a market pattern, stretch a survey finding beyond its data, and flatten a messy interview theme into a neat conclusion.
Researchers make mistakes, sometimes spectacular ones, sometimes in a chain where each bad finding feeds the next. Better citation tools and grounded AI systems certainly help. They can confirm that a source exists and reduce some kinds of errors. But a citation doesn’t tell you how much weight a source deserves, whether a marketing claim is representative, or whether a conclusion goes beyond the evidence. Those are judgment calls.
The same is true of scale. AI is exceptionally good at scaling a research process. If the process is weak, AI simply helps you reach weak conclusions faster and with more confidence. If the process is disciplined, AI can make it dramatically more efficient without replacing the judgment that makes the work valuable. That is why I treat AI output as a lead, not a finding.
In photography, nobody mistakes the contact sheet for the finished print. Research deserves the same discipline. AI can help us capture more frames, organize them, and even suggest which ones look promising. But the responsibility for choosing what belongs in the final picture—and what doesn’t—still belongs to us. Otherwise, we risk publishing something that looks polished, convincing, and confidently wrong.