Writing
Notes from the record.
Articles by Jereme Lytle, in four groups: what hallucination is, what the work takes, how to evaluate a tool, and where client data goes. Every factual claim is tied to a public source, listed at the end of the piece. The words themselves are defined in the glossary.
Hallucination
What a fabricated or misgrounded citation is, why it happens, and what courts have done about it.
A working link is not a verified source
A 2026 study of fourteen models: deep-research links open over 94 percent of the time, but the source supports the claim only 39 to 77 percent of the time.
Why language models hallucinate, in the words of the people who build them
OpenAI’s September 2025 paper says hallucination is built into how models are trained and scored. What that means for anyone who signs what a model wrote.
How to check an AI citation in five minutes
A procedure you can run before you file: prove the authority exists, prove the passage supports the sentence, then check the quotation word for word.
What a citation gate is, and why “grounded” AI still fabricates
A citation gate has to ask two questions, not one: does this authority exist, and does it hold what the sentence attached to it says.
The sanctions record: courts and AI-invented authority
What judges have actually ordered when a filing cited cases that do not exist — from a $5,000 penalty in 2023 to reprimand, removal and a bar referral.
Why retrieval does not stop hallucination
Retrieval changes what a model reads. It does not change the fact that the answer is written afterwards — and the measurements bear that out.
Defender practice
The defender workflow the record actually arrives in: body cam, dockets, and the memos written at night.
Body cam review: how long it actually takes
What the public record says about footage volume and review hours in criminal cases, why keyword search is not review, and what watching per issue means.
What a hearing memo is, and who writes it
The pretrial conference memo, the motions memo and the trial-prep memo: what each contains, who writes them today, and where the hours actually go.
Hearing memos from the docket: what autonomous means in a defender’s office
Docket in, memo out, with no one writing a prompt — and a precise account of what the attorney still does, and what the rules say she may not delegate.
Evaluating tools
Buying and trusting a tool: the questions to ask, the words to pin down, and the rules that bind the lawyer.
The AI scientist without a gate: what an independent evaluation found
A fully automated AI scientist, evaluated independently: 42 percent of its experiments failed and some papers carried invented numbers.
ABA Opinion 512 and generative AI: what it says, and what it does not
What Formal Opinion 512 requires on competence, confidentiality, supervision and fees, what it leaves to the facts, and how a defender office complies.
What "deterministic" means when a vendor says it
The difference between a deterministic decision layer and a model at temperature zero, what to ask a vendor, and what a reproducible rerun actually proves.
How to evaluate legal AI: questions to ask before a pilot
Nine questions an office can put to a vendor in one meeting, and the two tests to run on a closed file before anyone signs anything.
Client data
Where the client file goes, what the terms say about it, and what running a model in your own building takes.
On-premise AI for law firms: hardware, models and licenses
What local really means, what an office needs to run a model in its own building, open-weight licenses in plain words, and what leaves with each setup.
Client data, local models, and why on-prem is the default for defense
Confidentiality duties, what provider terms actually promise, and why a court in someone else’s case can change how long your prompts are kept.
Shorter answers to the questions that come up first: questions defenders ask.