Why rewriting clauses with GPT is tricky
When I first tried to use GPT to clean up legal clauses in a contract draft, I thought it would be as simple as pasting a chunk of text, typing something like “make this sound more professional but keep the meaning,” and then copying the result back into Word. Yeah… no. What actually happens is a mix of helpful polish and uninvited rewrites. GPT insists on fixing things that were not broken and sometimes completely changes the obligation part of the sentence. If your clause says the tenant “may” do something, it might come back saying the tenant “shall.” That’s a big legal difference — one is optional, the other is mandatory.
One very real frustration is when you paste in a dense sentence with defined terms, like “Party A” or “Effective Date,” and GPT helpfully decides to rephrase it as “the undersigned” or “commencement date,” breaking the internal consistency of the contract. This means later references don’t line up anymore. The fix I found is to include an extra note before my request, something like: “Do not change any defined terms or their capitalization.” Even then, about once out of every five tries, I get a clause where the defined term is subtly altered, and I have to manually fix it again.
Keeping legal tone without losing meaning
New users usually ask how to make GPT keep the legal formalities without turning everything into incomprehensible old English. The trick is to make it focus on clarity first, formality second. For example, I tried rewriting: “The Contractor shall procure and maintain insurance coverage adequate to protect the Company.” GPT’s first attempt was, “The Contractor is obligated to secure and sustain sufficient insurance coverage as deemed appropriate for the protection of the Company.” It sounded formal but basically bloated the sentence.
What I do now is add a plain-language summary before the rewrite request, like: “This clause means the contractor has to have enough insurance, period. Keep it professional but short.” Suddenly, GPT gives back: “The Contractor shall maintain adequate insurance to protect the Company.” That’s clean, still formal, and not drowned in filler words.
Also, beginners often overlook how sensitive GPT is to synonyms. Changing “may not” to “shall not” is something it thinks is harmless — but it’s a completely different level of prohibition. If I catch that happening, I just undo and run the same request again, because in my experience the second result is usually closer to what I intended :P.
Spotting silent wording changes
The most dangerous thing here is when GPT changes the risk allocation without you noticing. I had a clause that originally read, “The Supplier shall indemnify and hold harmless the Purchaser from any and all claims…” GPT once returned, “The Purchaser shall be indemnified by the Supplier from claims arising solely from the Supplier’s negligence.” That tiny “solely” essentially gutted my indemnification. If I had skimmed over it, the entire protection would have been narrowed.
This is why you have to run a side-by-side check each time. What I do is paste the original in one column of a table, paste the GPT output in the other, and scan for words that pop out — “only,” “solely,” “as determined by,” stuff like that. It’s slower, but it catches these little traps. Once you’ve seen three of these sneak into a draft, you stop trusting the output blindly.
Handling formatting mess after rewriting
You’d think GPT would just give you back clean text you can drop into Word or Google Docs, but it loves to decide how indentation or numbering should look — and that breaks the structure of your document. For instance, a perfectly aligned subclause “(a)” can suddenly come back as “1.” or no bullet at all. If your contract numbering is automated in Word, pasting GPT’s version can trigger re-numbering for the whole section, and now clause 12.4 is suddenly 1.1.
I avoid this by stripping GPT’s formatting entirely before pasting. On my Mac I paste into a plain text editor first, then into Word. On Windows, I use “Paste as plain text” in the right-click menu. This kills smart quotes, bolding, and random tabs, so my numbering system survives. Also, if you’re doing multiple rewrites inside a long contract, keep a master copy of the whole document and a working copy for testing — otherwise one paste error can scramble hours of layout work ¯\_(ツ)_/¯
Minimizing cross clause conflicts
Let’s say GPT rewrote your confidentiality clause to say “for a period of three years,” but your earlier non-compete clause already said “for a period of five years.” Boom — you just created an internal conflict without even realizing it. Contracts live and die on internal consistency, so any time I touch one clause, I have to scan for time periods, names, and cross references like “as set forth in Section 7.” GPT sometimes invents new cross references or changes a section number.
I’ve started putting a quick checklist in a separate note: defined terms, section numbers, time frames, parties’ names, governing law. Every time I run a rewrite, I check them against the rest of the document. Yes, it’s tedious, but it takes way less time than hunting down the source of a contradiction the day before signatures are due.
Making sure plain language still protects you
There’s a temptation to use GPT to strip all legalese and make clauses super readable. I tried this on a limitation of liability clause and, while the new version was much easier on the eyes, it silently dropped the liability cap amount and the exclusion for indirect damages. That would have been catastrophic if I hadn’t spotted it.
The balance here is to let GPT make the wording approachable but never let it drop or replace numeric caps or specific exclusions. If those make the text harder to read, I’d rather have a clear second sentence explaining them than lose them entirely. For instance:
Original:
“Neither Party shall be liable for any indirect, special, or consequential damages, and the total liability shall not exceed $100,000.”
Rewritten with manual fix:
“Neither Party will be liable for any indirect or special damages. In any event, the total liability will stay under the agreed cap.”
And then somewhere else in the contract, define that cap in a schedule or definitions section.
When to just draft it yourself
Sometimes the fastest way to fix a clause is to just rewrite it yourself after seeing GPT’s version. I had a termination-for-cause provision that GPT kept turning into three separate sentences, breaking the sequence of conditions. After two failed outputs, I realized it was faster to type my own clean sentence based on the parts I liked from GPT’s attempt.
The main lesson here: GPT is great for breaking writer’s block and giving you alternative phrasings, but never assume it understands the contract’s structure the way you do. That sense of “oh, this is good enough” is dangerous when legal meaning is on the line. I just treat it like a very eager but slightly sloppy paralegal — useful, but definitely not allowed to send the draft straight to the other party without my eyes on it.
Testing the workflow on small chunks
If you dump a 50-page contract into GPT and say “rewrite the problem parts,” it will either bail halfway or hallucinate changes in sections you didn’t touch. I only feed it one to three clauses at a time, max. That way, the memory of prior rewrites doesn’t contaminate new edits. If I absolutely need to keep context — for example, making sure a non-disclosure clause aligns with a reusable definition section — I paste that definition above my rewrite request so it doesn’t guess.
Doing it this way also keeps the mess small if GPT decides to swap party names or change timelines. It’s way easier to fix three clauses than reverse-engineer what it changed in thirty pages without any highlights.
It’s not perfect, but after enough of this, you start to predict the weird edits before they happen — sort of like knowing your coffee maker will always try to spit one last drop on the counter after you finish pouring 🙂