Artificial Irreverence: the Limits of Language Model Lawyers*

*This article first appeared in the San Francisco Apartment Association’s (SFAA) November 2025 issue of the SFAA Magazine.

Everything these days is an AI story. As you faithful SFAA members know, San Francisco has seen double digit percentage increases in year over year median rents – the highest annual increase in the country in fact. Speculation as to the cause of the sudden shift in the market range from the evergreen “supply scarcity”, return to office policies, and “AI”. This reductive term is now a broad label to describe new industry tools, the companies that compete to create the best ones, the demand for qualified employees to develop them, and the conspiracy theories about how this will finally be the technology to change everything we know about technology. As a cause of rent increases, artificial intelligence poses an ironic narrative. Booming industry means hiring the best from around the world, and they need housing. But the industry is pulling in office workers to develop the platforms to replace office workers.

Is the speculation about the role of AI in our industry overblown? Technological advancement has often led to critical efficiency gains. Imagine trying to calculate a banked rent increase using a slide rule and ledger paper. The firm I started at had a typewriter. Gray as I may be getting, I’m not quite of the generation that drafted pleadings on those clunky old things. Instead, it was used to interlineate text on judicial council forms (e.g., to add the word “amended” in front of the pre-printed title). We can now edit pre-printed forms in seconds with Adobe.

In my July 2023 column, Write on Cue, I mused about lawyers using AI the wrong way (like generating briefs based on imagined case law). I maintained that legal ethics and intentionality would find lawyers ignoring the lay opinions of language models in favor of their own legal analysis. (It’s one thing to advance a theory that the judge finds unpersuasive, but we certainly don’t want to find ourselves at a hearing explaining to the judge that one of our arguments must have been an AI hallucination that we didn’t properly review.)

I’ve seen legal research make leaps forward before. I was one of the last generations of law students who learned “paper research” (following the trail of key words from the latest, printed copy of an index to navigate through hundreds of volumes of chronologically bound decisional authority). These days we can just ask Westlaw a question and probably find the case we could have been looking for the hard way.

AI feels different. Earlier this year, I attended a presentation on Westlaw’s AI add-on, which will analyze pleadings, discovery and client documents in order to summarize data, answer questions and draft work product. I had an existential crisis in that meeting, because this kind of software would eliminate menial tasks for young associates. Associates do more than just the menial, and they can often be put to better use than focusing on tedium. But attorneys learn through the grind. When you have to squint through the fine print in a boilerplate lease, you’re going to develop a relationship with the provision you’re looking for.

I thought it would be a good time to revisit the topic because it keeps popping up in my practice. In an increasingly common story arc, AI refers clients to me (thanks AI), those clients ask AI to give them some background on the law, I provide finer points and practical applications to the clients, and then I work on the matter. I use AI tools to help analyze a client file to provide me with backgrounds, biographies, timelines and key facts. (I’ll even ask for the occasional primer on a novel area of law that supplements my normal practice.)

Or, at least, this would be an ideal experience. I’ve also had clients use AI to dictate an incorrect version of the law, draft correspondence and other work product they want me to edit for them, or even try to educate me on negotiation tactics they’ve learned from AI. (Telling a lawyer that you want to teach them how to negotiate, after spending a few minutes Googling it no less, should provoke a response like, “do I go to your job and tell you how to sweep?”.)

But I tried to leave my biases at the door when I recently sat down with Chat GPT to interview it as my replacement.

JAG: What advantages do large language models have over human lawyers?

ChatGPT: We don’t have student loan debt.

In all seriousness, I did actually engage it quite a bit in working on this column. I asked it about its familiarity with San Francisco landlord-tenant law, which it knows quite well, actually. As any good “lawyer” would, it started from the most general principles with an invitation to discuss nuance. (Frankly, it’s responses are on the “flashy” end of thorough, like a younger attorney eager to convince me how much they knows.)

Of course, this L-AI-YER* is only as good as its inputs. (*By the publishing deadline, I had not come up with a punchier name, but ChatGPT thought it was phonetically consistent and playfully satirical, evoking the anxiety around AI replacing lawyers.) It will freely share its pedigree, but much of its qualitative analysis seems to be coming from marketing materials (like law firm blog posts) instead of living through actual controversies.

In terms of market trends, it can repackage articles and op-eds with a convincing thesis. But, for instance, when asked about recent eviction levels, it adopted a story that sourced its data from the Rent Board’s eviction notice report. This data is only published for the year ending February 28^th (so it lags), and it is both over and under-inclusive (where the notices don’t necessarily correspond to the number of eviction lawsuits or actual tenant displacement). A more accurate count comes from the Superior Court itself, which recently shared with us that, while there were 2,923 unlawful detainers filed in 2024, there were already 3,474 filed just through mid-September of this year. (ChatGPT didn’t think to ask the Superior Court.)

As many readers are aware, a recent eviction case out of Los Angeles, Eshagian v. Cepeda, completely upended our conventional wisdom about the content of a notice to pay rent or quit, to the extent that landlord lawyers have debated how to comply, we’ve been following the fate of active cases, we asked the Supreme Court to depublish the opinion, and in the last stage of grief, we just coped with it and put all these silly new things in our notices.

I challenged ChatGPT to draft a three-day notice, but it missed the change. However, after asking it about Eshagian and discussing the requirements, it did much better. This is the difference between asking an associate to update work product versus asking them to generate competent work product from scratch. The former takes mentorship, the latter wisdom. (ChatGPT did, however, immediately agree with me that the Court got Eshagian wrong, so at least it knows how to suck up to the boss.)

Turning to the practical, I asked it how to navigate a tricky dispute. Landlords owe duties to tenants, like maintaining quiet enjoyment (i.e., avoiding disturbances that affect their use of the unit). When disruptions come from other tenants’ use of the property, the landlord becomes guardian, peacekeeper and disciplinarian. Severe penalties ward off a landlord’s wrongful acts that displace rent-controlled tenants for pecuniary benefit. Does passivity in policing a troublesome tenant create the same liability as to the affected one? And does that potential outcome compel a landlord to evict as a precaution? Most evictions respect a dynamic between a landlord and a tenant, but this dispute is multilateral.

ChatGPT cleverly observed the various vectors of conflict, warned about the consequences of both action and inaction. It then recommended a mediative approach and volunteered to role-play correspondence between a hypothetical landlord attorney and tenant attorney. In this exercise, it earns high marks.

In my August 2025 column, I discussed the imbalanced economics of landlord tenant disputes, and how savvy landlords can build toward inflection points that foster reasonable resolution. Countless decisions lead to these junctures, not just in the content of a notice or complaint, but from the very outset in managing client expectations from the perspective of experience. Large language models (LLMs) are trained on vast volumes of text, and they function by predicting the next part of the pattern. But effective lawyering isn’t about reciting past outcomes, it’s about imagining new ones and synthesizing the tools to get there.

Before all my free time went to writing these columns, I frequently blogged about landlord-tenant law on my website Costa-Hawkins.com (perhaps you’ve heard of it). This exercise could not have made sense as a marketing tool, given the post-to-client ratio. But I always found its value in having an experience with a particular “occurrence” of law. If I read a new case or statute well enough to post about it, I had to understand it critically. And if I did that, I could call upon it later as one more of my tools. In my practice, I take every case as an opportunity to expand my abilities, altering my intuitions as I work through the experience of a novel application. There are rarely surprises, but I can attack a problem in a new way to see if that optimizes the outcome. I don’t want to predict the next step in the pattern so much as remember that I created the pattern in the first place.

AI is a shockingly capable tool for parsing content, gathering the basics of an area of law, and even recommending well-worn work-throughs. It will no doubt become more specialized and capable as a legal supplement. Maybe my Fall 2027 column will damn my hubris, but I’m not any more concerned about AI displacing lawyers than I was two years ago. If anything, I’m excited to eliminate inefficiencies to focus on what I love most – thinking through complex legal theory and putting it to work in court. For now, AI knows how to be thorough and circumspect but not inventive and incisive. You might say that LLMs aren’t worthy of an LL.M. (This is actually a clever pun about large language models sharing an acronym with a Master of Laws degree. ChatGPT thought this was funny, but it had notes. Whether AI will replace comedians is outside the scope of this column.)