Why a Database of 10 Million Cases Makes Your AI Worse

Cable companies used to brag about having 500 channels. You watched two of them. The other 498 were clutter that made it harder to find the thing you actually wanted.

A lot of legal AI marketing sounds the same way now. "We have 10 million cases in our database." For a high-volume defense attorney, how many of those matter? A few hundred, maybe. The rest are 498 channels you'll never watch. Here's the part the pitch leaves out: those extra channels are actively making the AI worse at the thing you actually need it to do.

The broader legal-AI debate is dominated by fear of hallucinated cases, arguments about RAG versus fine-tuning, and leaderboards ranking whose legal database is biggest. Almost none of it applies to high-volume defense practice. That world is different, and once you understand how, the hallucination problem mostly dissolves.

High-Volume Defense Is a Different Animal

When we say "high volume," we mean firms handling the same type of case hundreds of times a year. In NY, this includes: MCA defense, no-fault, and 5102(d) threshold motions. The economics only work because the cases are similar. The legal theories are the same and the arguments repeat.

Research, as a percentage of total work, is tiny. A senior attorney in one of these niches already knows every case and statute that matters. They cite the same authorities, motion after motion, because those are the authorities that win. New controlling cases come along occasionally, and when they do, the attorney reads them, decides whether to fold them in, and updates their template. That's it. The universe of law that matters for the niche is small enough to fit in an attorney's head, and small enough to fit, in structured form, in a playbook.

This changes what the AI is being asked to do. The bulk of the work in a high-volume practice isn't finding the law. It's extracting facts from messy case files, matching those facts to a known playbook, and assembling documents. Most of the legal-AI discourse, including most of the hallucination discourse, is solving for a problem these firms don't really have.

If You Already Know the Law, Stop Asking AI to Find It

The simplest way to eliminate fabricated-case hallucinations is not to use AI for case citations at all.

When the universe of relevant law is small and already curated by the attorney, there is no reason for the AI to be pulling cases from memory, or searching a 10-million-case database, or reasoning about which authority to cite. The attorney has already answered those questions. The AI's job is to take a known, vetted argument and adapt it to the facts of a new case.

"Do not modify any case citations" is a trivial instruction for a modern LLM. Bounded text manipulation is what these models are genuinely excellent at. Take this argument, keep the citations exact, adjust the facts to match the new case. That's a task LLMs handle reliably. The risky thing was asking the model to generate legal authority in the first place. Take that job away and the entire class of fabricated-case hallucinations disappears.

The Boilerplate Nobody Wants to Talk About

High-volume defense attorneys are refreshingly honest about this: a lot of their briefs are boilerplate. Entire sections are identical, word for word, across hundreds of motions. That isn't a weakness of the practice. It's how the practice is high volume. The arguments have been refined over years. They work. Rewriting them from scratch on every new case would be a different kind of malpractice. It would waste hours, introduce errors, and produce inconsistency across the firm's work product.

The broader legal-AI conversation mostly ignores this, probably because it's unglamorous. But if a section of a brief is identical in every filing, why is the AI rewriting it? Copying has a 0% hallucination rate. AI paraphrasing does not. The correct tool for "this paragraph is the same every time" is Ctrl+C, not a language model. Every paragraph the AI never touches is a paragraph that cannot hallucinate.

Where the AI Actually Decides Something

Not everything is boilerplate. Some arguments only apply to a subset of fact patterns, and that's where the AI earns its keep.

A concrete example from MCA defense: the usury argument has a branch that only opens up if the defendant actually requested reconciliation from the funder. That doesn't happen in every case. When it does, you want the reconciliation-branch argument included. When it doesn't, you don't.

The right job for the AI here is narrow: look at the facts, determine whether the reconciliation branch applies, and select the pre-written argument. It is not writing the argument. It is not reasoning about whether reconciliation matters under New York usury law. The attorney already answered that question, years ago, and wrote the branch. The AI is choosing from a menu.

Reduce the AI's decision surface to selection rather than generation, and the failure modes shrink to something a paralegal can review in minutes.

The Real Hallucinations Are Somewhere Else

None of this means hallucinations go away. They just move.

The real AI failure modes in a high-volume defense workflow look like this: missing a claimed body part in the bill of particulars, misreading a date on an agreement, pulling the wrong figure out of a settlement statement, attributing a statement to the wrong deponent. These are ordinary LLM errors, the kind that happen when a model fails to accurately extract structured information from messy documents.

The good news is they're much easier to manage than fabricated case law. Every factual claim can be tied back to a specific page of a specific document. A paralegal reviewing the draft can see exactly where each fact came from and verify it in seconds. The errors are visible and local, not buried inside a plausible-sounding legal argument that requires a Westlaw search to disprove.

Why Bigger Databases Make This Worse

When you hand an AI free rein over a massive legal database, you don't know what it pulled and you don't know what it missed. The output looks polished. It looks comprehensive. A fabricated case is easy to spot, because the citation doesn't resolve. A plausible-but-incomplete selection of authority, drawn from 10 million cases by an opaque retrieval algorithm, is not. You traded a visible problem for an invisible one.

For high-volume defense, where the attorney already knows which 200 cases actually matter, this is a bad trade. A curated set of vetted authorities beats 10 million uncurated every time. The AI doesn't need more options. It needs fewer, better ones.

The Work Is Somewhere Else

The fear of hallucinated cases has sucked most of the oxygen out of the legal-AI conversation. For high-volume defense, it mostly shouldn't have. The case-law hallucination problem is nearly solved, not by better models or bigger databases, but by not asking AI to do the part it's bad at in the first place. The real work, and the real opportunity, is in the messy middle: extracting facts accurately, matching them to a known playbook, and assembling documents that look like the ones the firm already files.

That's the problem we built DocketDrafter around.