Every matter file has a bad page
Open almost any litigation file and you'll find one. A medical record that was faxed, printed, and scanned again. A photocopy of a photocopy where the left margin fades into gray. A treating physician's handwritten notes. A client's phone photo of a letter, shot at an angle on the kitchen table. A date stamp sitting right on top of the sentence you need.
Personal injury practices see more of these than most. Records come from many different providers, each with its own scanner and its own idea of legible. Your paralegals already know which pages are trouble. The question is whether your AI tool knows too, and whether it tells you.
What OCR actually does with a bad scan
A scanned page is a picture. Before any software can summarize it, something has to turn that picture into text. That step is optical character recognition, or OCR, and on a clean typed page it works well.
On a bad page, OCR guesses. A smudged 3 becomes an 8. A faint "not" drops out of a sentence. Two columns get read straight across and stitched into nonsense. Handwriting comes out as a string of plausible letters that mean nothing at all.
Here's the part that matters. The text that comes out the other end looks the same either way. There's no font for "I wasn't sure about this." A misread date sits in the output with the same confidence as a correct one.
Where it goes wrong
Now hand that text to a language model and ask for a summary. The model is very good at producing fluent, reasonable sounding prose. If the input has holes, it tends to write right over them. A garbled line becomes a tidy sentence. A missing date gets filled in from context. The summary reads beautifully, and nothing in it tells you that page eleven was mostly static.
That's the dangerous combination: a weak input and a confident output. A rough draft that looks rough gets checked. A polished draft that looks finished gets trusted. When the error is a treatment date or the name of a medication, it can travel a long way before anyone catches it, if anyone does.
This is a big part of what we've called the verification tax. If you can't tell which statements rest on solid pages and which rest on guesses, you end up rechecking everything, and the tool saves you very little.
What an honest refusal looks like
The better behavior is simple to describe. When a tool can't read a page well enough, it should say so, plainly, in the output a person will actually look at.
- It names the page. Which document, which page, and roughly what the problem is: handwriting, low contrast, text cut off at the edge.
- It keeps the guess out. Nothing from that page gets folded into a summary or a timeline as though it were read cleanly.
- It points you to the original. One click should take a person to the page image, so they can read it themselves.
- It shows up where people work. A warning buried in a log file that nobody opens doesn't count.
"I couldn't read page eleven" is a useful sentence. It tells a paralegal exactly where to spend a few minutes. A confident sentence built on page eleven's static tells them nothing, and it might send the case in the wrong direction.
Why this matters most in chronologies
Think about a matter chronology built from medical records. A timeline with a marked gap, "visit on this page, date unreadable, see original," is honest work. A timeline with a confidently wrong date looks complete and is quietly broken. The second one can shape a demand letter, a deposition outline, or an expert's understanding of the case before anyone notices.
Every entry in a chronology should link back to the page it came from. When it does, an unreadable page is a short detour. When it doesn't, you're back to trusting the summary, and that's the thing you can't afford to do.
How to test a tool before you buy it
Vendor demos run on clean documents. Your files aren't clean, so test with your worst ones. Pull a handful of pages your team already knows are trouble, from a closed matter where the client's permission and your own policies allow it, and see what the tool does with them.
- Does it tell you which pages it couldn't read, by name and page number?
- Can you click from any statement in its output straight to the source page?
- What does it do with handwriting: read it, flag it, or invent something?
- If you blank out a date on a test page, does the summary admit the date is missing, or does it supply one?
- Where do the warnings appear: in the review screen, or somewhere nobody looks?
A tool that passes these checks will sometimes look less impressive in a demo. It'll say "unreadable" where a competitor produces a smooth paragraph. That's the tool you want reading client files.
How we approach it
We designed the Tiber River Legal Workbench to treat an unreadable page as something to report. It runs on hardware the firm controls, produces source-linked review aids, and is designed to flag the pages it couldn't read so a person can go to the original. It's built to mark gaps rather than fill them with guesses, and it doesn't decide anything. The attorney decides what goes into the file and what goes out the door.
Tell us which pages give your team trouble
The Tiber River Legal Workbench is a matter intelligence appliance for litigation and personal injury practices, running on hardware your firm owns. Your paralegals know the faxed records and handwritten notes better than any vendor does. We're inviting a small number of firms to shape it as design partners.
Start the conversationOr write to josh@tiberriverconsulting.com