You have a word that doesn’t sound right. You paste the sentence into ChatGPT and ask how to spell it. AI produces an answer. It’s wrong. The reporter catches it in review. Everyone moves on — and nobody asks why it happened.

Here’s why it happened.

ChatGPT is not a dictionary. It’s a probability engine. It predicts the most statistically likely answer in general English. And general English is a wide dial — every frequency at once. When the model doesn’t know which station you’re on, it defaults to whatever is loudest across all of usage.

Access is louder than excess. Statue is louder than statute. A surname that sounds like Cohen has multiple competing spellings — and the model picks the most common one, not the right one for your case.

The fix isn’t a better dictionary. The fix is narrowing the dial. And narrowing it is a skill — one you build, own, and carry to every job.

See It Work

Take a fictional case: Voss v. Meridian Excess Partners — an insurance coverage dispute over layered excess policies.

Without context, you ask: The witness said something that sounds like “access layer” — is that right? The model says yes. Access layer is common in technology. Confident answer, wrong domain.

Now you paste the case caption and run the extraction prompt below. The Case Context Block comes back: excess insurance, layered coverage, attachment points, retention. Same question. The model says: “Excess layer” is almost certainly correct — referring to a layer of excess insurance coverage above the retention. “Access layer” is a technology term and doesn’t fit this domain.

Same question. Same model. Different universe. And in a coverage dispute where the word excess defines the entire layer structure at issue, the difference between those two answers is the difference between a clean transcript and one that misrepresents the testimony.

Build the Universe Before the Job Starts

Paste whatever you have into a chat and run this prompt. Don’t underestimate how little that needs to be — party names alone narrow the universe. A company name tells the model an industry. An industry tells it a vocabulary. A witness name helps but isn’t required. Two sentences and a case caption can be enough to shift the model out of general English and into your domain.

I am preparing for a deposition transcript. Below is all the information I have. Extract and structure it into a Case Context Block: what the case is about, the witness and their role, the primary legal issues, the industry or technical domain, the time frame, terminology likely to appear, what this case is NOT about, and potential lookup traps — similar-sounding names, technical terms, acronyms, foreign names. Present it as a block I can paste at the top of this chat and reuse throughout the job.

That Case Context Block becomes the permanent header for the job thread. Every lookup from that point forward runs inside your case — not all of English. You don’t need the reporter to provide it. You build it from whatever lands in your inbox.

After the block, add one short line: what this case is not. Not criminal. Not personal injury. Not medical malpractice. Every exclusion eliminates a competing frequency. You’re not just pointing the model at the right station — you’re switching off the wrong ones.

One critical rule: keep everything in the same thread. The model has no memory between sessions. Open a fresh chat and the context is gone. Start one thread per job. Stay in it.

Make It Cautious, Not Just Helpful

Paste this immediately after the Case Context Block:

Assume this conversation relates only to this case. Prefer consistency with previously established spellings and terminology. If multiple spellings are plausible, rank them by probability and explain why each fits or doesn’t. Do not guess — only provide answers you are reasonably confident in given the case context. Do not invent affiliations or facts. If you are uncertain, say so clearly rather than offering a best guess.

Without this, the model defaults to helpful — confident whether or not that confidence is earned. A helpful guesser picks the most common spelling and moves on. A cautious analyst ranks three options and explains the tradeoffs. You want the second one.

Prompts to Use During the Job

The comprehensive lookup prompt — use this for most situations:

Here is a sentence from this deposition: “___________________________.” The word or phrase that may be incorrect is: “.” Using the established case context: (1) confirm whether this word or phrase makes sense in this domain; (2) if not, provide the three most likely intended alternatives ranked by probability and explain why each fits or doesn’t; (3) flag any phonetically similar alternatives — including negations, near-homophones, or compressed words — that would significantly change the meaning; (4) if it may be a name, acronym, foreign term, or product name, identify the most likely candidates. Note any uncertainty clearly.

Drop in the sentence, drop in the flagged word, send it. The model handles the triage.

Add phonetic clues whenever you have them: sounds like three syllables, hard C or K sound, ends in -tion, may be an acronym. Phonetic constraints combined with case context narrow the field more than either one alone.

For targeted lookups when you know exactly what you need:

Name clarification:

In this case context, what are the most likely spellings of a surname that sounds like “_____”? Rank by probability.

Acronym expansion:

In this industry context, what are the most likely expansions of the acronym “_____”? Which is most consistent with this case?

Foreign name or term:

The witness used a word that sounds like “_____.” It may be a foreign name, place, or technical term. In the context of this case, what are the most likely candidates?

Medication or product name:

The witness referred to what sounds like “_____.” In this case context, what pharmaceutical, medical device, or product name is most likely intended?

Technical phrase check:

Does the phrase “_____” make sense in this domain? If not, what’s the closest likely technical term?

The Mishear You Won’t Catch

Everything above assumes you know something is wrong. The word doesn’t parse, the spelling looks off, the term feels unfamiliar. You ask. You find it.

The harder problem is the mishear that passes the eye test. The sentence is grammatical. The word fits the context. Nothing triggers a lookup. And it’s completely wrong.

Programmatically unable vs. programmatically enabled. Opposite meanings. Both land cleanly in a transcript. The scopist who doesn’t know to look won’t look — and the error survives to the final.

This is where AI earns something the lookup workflow can’t: a prompted sanity check on technically dense testimony. After scoping a complex answer, paste it and ask:

Does this sentence make sense in context? Are there any words here that have phonetically similar alternatives that would significantly change the meaning?

The model won’t catch everything. But it will flag the pairs it knows — words that compress at speed, technical terms with near-homophone cousins, negations that can drop a syllable and reverse the sentence. That’s the category where confident transcription causes the most damage. A few seconds of verification on high-stakes answers is cheap insurance.

When AI Is Still Wrong

The context workflow dramatically reduces wrong answers. It doesn’t eliminate them. If an answer comes back and something still feels off — the name doesn’t look right, the term doesn’t ring a bell — trust that instinct. Flag it for the reporter and move on. You don’t need to be certain something is wrong to flag it. Uncertainty is enough.

It Compounds

If the same case comes back, find the old thread instead of starting a new one. The context is already there. Don’t throw it away.

Same parties, same experts, same counsel — you’re back up to speed in seconds instead of starting from scratch. That’s the compounding advantage, and it costs nothing except not hitting “New Chat.”

Fewer errors in the rough means less for the reporter to correct. Less to correct means you’re more valuable, more trusted, and more in demand, and the client (both the reporter and the attorneys) are getting a better product.

You’re not automating your work. You’re sharpening your judgment — using a tool that rewards the scopist who takes five minutes to build the universe over the one who types “how do you spell that?”

Build the context. Own the process. The reporter benefits. So do you.