OCR Redacted Documents: What the Black Boxes Do to Text

When you run OCR on redacted documents that someone else prepared, you get text for the visible words and nothing reliable where a box sits. The converted text may close the gap, fill it with stray characters or break the line. Mark every redaction in your working copy, log each one, and stop if words turn up under a box.

What does OCR output where a redaction box covers the page?

Where a box covers the page, converted text shows one of four things: nothing at all, a run of junk characters, a label that was printed on the box, or a break that splits the line or the paragraph in two. None of them tells a reader that words were removed. Only the page image does.

Take one line from a produced email in a fictional matter, Halvorsen v. Brightwater Freight. On the page it reads:

On March 4, ▇▇▇▇▇▇▇▇▇▇ told the dispatcher that the load would be released to ▇▇▇▇▇▇▇▇ by noon.

Depending on the box and the tool, the text for that line may come out in any of these ways:

The first row is the dangerous one. It is short of a subject and an object, yet it scans as a sentence if you are reading quickly, and in a long paragraph a closed gap is easy to pass over.

In a production, that closed gap can hide a privilege label or a work-product doctrine stamp you were meant to log. You also watch admissibility language in a cover letter that a box sliced in half. For impeachment later, the image plus your marker is what shows a sentence was never whole.

Software treats a box as it would any dark shape: a picture, a rule, a table cell or a very bold character. Which of those it picks can change from one box to the next on the same page, so do not expect one consistent placeholder you could search for.

To see what a particular tool does, run a page. Schema Law reads each page into text and tables and shows the parsed page on screen, so you can hold it against the scan before exporting. It does not mark where a redaction was, and nothing marks a passage as uncertain. The free plan's one document a day of up to three pages is enough for the test.

Black bars, white boxes and marker redactions on a received document

Redactions arrive in several styles, and the style decides how visible the gap is in converted text. A black bar may leave junk characters that at least hint at a deletion. A white box on white paper leaves nothing in the image or the text, which makes it the easiest to miss.

White boxes need a second look

A white box in the middle of a paragraph looks like generous spacing. In a table it looks like an empty cell. In a list of names it looks like the list simply has fewer entries. The converted text gives no hint. The way to find white boxes is to page through the image looking at the shape of the text: ragged gaps inside justified lines, a paragraph that ends short of the margin and resumes mid-sentence, a column with one cell blank among full ones.

Marker redactions and partial letters

A hand-drawn bar may not cover a word exactly. The tops of tall letters or the tails of "g" and "y" may stick out. Recognition can turn those fragments into characters of their own, such as a row of apostrophes or commas. Treat any such string beside a bar as noise, and do not try to read meaning into it. What is under a bar is not yours to reconstruct.

Labels are worth keeping

When a box carries a label, such as a privilege designation, a category from a protective order or a statutory code, that label is information the producing party chose to give you. Record it exactly. For federal agency releases, the post on FOIA documents explains redaction boxes and exemption codes and how those codes get misread.

A bar is not always a redaction

Dark highlighter that a photocopier turned black, a strip of tape, a fax streak or a line drawn through text can all look like a bar. Square edges, consistent height and a pattern of covering names or numbers point to redaction. If you cannot tell, log it as "possible redaction" and ask the sender.

Why do redaction boxes break sentences, tables and search hits?

A box takes out part of a structure that the remaining text still depends on. A sentence loses its subject, a table row loses a cell, and a phrase you search for loses one of its words. The converted text keeps the remains and presents them as if they were whole.

Sentences

A closed gap changes meaning as well as grammar. "Payment was approved by ▇▇▇▇▇ and not by the board" can become "Payment was approved by and not by the board", which a hurried reader may mend in their head to something the page never said. Longer redactions do more damage: when a whole clause is gone, the two halves left behind can join into a statement with the opposite sense.

Tables

In a table, a redacted cell that produces no text can cause every later cell in the row to shift one column to the left. A phone log with the number column redacted in some rows then shows durations under "Number" and nothing under "Duration" for those rows. In a spreadsheet export this looks like data. Before sorting or totaling any exported table from a redacted document, compare the cell count of each row with the header, and look at the image for each row that comes up short.

Search

Three things go wrong with search:

• Phrase searches miss. A search for "released to Corwin Haulage by noon" fails if the name is under a box, and also if a junk character sits where the box was.
• Proximity searches mislead. Two words that were a paragraph apart on the page can be adjacent in the text after a block redaction, and a search for one word near another will hit.
• A search that finds nothing proves little. The term may be absent from the document, or present and redacted. Only the log of redactions tells you how much of the document you could not search.

Counts and comparisons

Word counts, page-level summaries and text comparisons between two versions of a document all treat the gap as nothing. If you compare a redacted production copy with an unredacted copy from another source, every redaction shows up as a deletion, mixed in with any real differences. Mark the redactions first, and the comparison becomes readable.

Marking each redaction in the converted text of a document

Put a bracketed marker in your working copy at every place the page image shows a redaction, using one fixed form throughout the matter. A marker tells the next reader that the page holds something the text cannot, and it gives search a string to count.

A notation to use

Rules that keep the notation honest

1. Work from the image, page by page. The text will not show you where to put markers; the image will.
2. Mark in a working copy. Keep the file as received and the raw export unchanged beside it.
3. Use square brackets and capitals. They are unlikely to collide with the document's own words, and they survive export to plain text.
4. Describe size only as the page shows it. "Approx. 6 lines" or "one cell" is an observation. "Name of supervisor" is a guess, and guesses belong in your notes, not in the text.
5. Delete junk characters the box produced in the working copy, and replace them with the marker.
6. Carry the marker into every quotation. A quoted sentence that spans a box shows the box.
7. Write the convention down in the matter's file note, so a second reviewer marks the same way.

A marked line then reads like this: "On March 4, [REDACTED] told the dispatcher that the load would be released to [REDACTED] by noon." Nobody can mistake it for a complete sentence.

The same convention for other things text cannot hold

A redaction is one case of a wider problem: the page shows something, and the text has no way to say so. The same bracket form works for the rest.

Posts on particular marks go further, including whether OCR can read an embossed seal and stamps over text. Keep the vocabulary short. Six or eight markers used the same way every time are worth more than twenty used loosely.

Counting and logging the redactions in a production you received

A redaction log is a table with one row per redaction: where it is, how big it is, what label it carries and what the visible text around it says. It turns "this production is heavily redacted" into a list someone can act on, whether in a meet-and-confer letter, a motion or a follow-up request.

Columns for the log

The entries are invented to show the form. Record what is visible and stop there. The log is a description of the page, and its value in any later dispute depends on its containing no speculation.

Building it without retyping everything

1. Finish marking the working copy first, so every redaction carries a bracketed marker.
2. Search the working copy for [REDACTED and [PAGE REDACTED and note the number of hits.
3. For a long production, export the marked text to a spreadsheet, or paste it in, and filter for rows that contain the marker to draft the page and context columns.
4. Fill in style and extent from the image.
5. Page through the image once more for white boxes, which leave nothing to search for until you have marked them.

If the documents carry Bates numbers, log by Bates number and not by PDF page, since the PDF page changes whenever files are combined. The post on how to OCR Bates numbered documents covers getting those numbers into the text reliably.

What the totals tell you

Sum the log by document, by label and by style. A producing party's privilege log, where one was served, can then be compared line by line: redactions with no matching entry, entries with no matching redaction, and labels that differ between the two. What obligations a producing party has to explain its redactions depends on the rules of your court and any order in the case, so that question goes to the lawyer handling the matter. The log supplies the facts for it.

If you are the one applying redactions, the sequence is a separate subject, covered in the post on whether to OCR before or after redaction.

What should you do when OCR finds text under someone else's redaction?

Stop reading, do nothing further to the file, and tell the lawyer responsible for the matter. A redaction that left its words in the file is a sign that you may hold something the sender did not mean to give you, and your jurisdiction's professional conduct rules may require a step such as notifying the sender.

This comes up without anyone trying. A search hit lands inside a blacked-out area, or exported text contains a name the page image hides. The post on hidden text explains why a black box drawn over words leaves the words in the file. From the moment you notice, the question is no longer a technical one.

The rule to look up

Look for your jurisdiction's conduct rule on documents received by mistake. The wording and the duties are not the same everywhere. As one example, Washington's Rule of Professional Conduct 4.4(b) reads:

A lawyer who receives a document or electronically stored information relating to the representation of the lawyer's client and knows or reasonably should know that the document or electronically stored information was inadvertently sent shall promptly notify the sender.

Comment [2] to that rule says that the rule's term for a document includes embedded data, which it notes is commonly referred to as metadata, and that whether the lawyer must take further steps, such as returning the document, "is a matter of law beyond the scope of these Rules". Washington's text is quoted here only to show what such a rule looks like. Find the rule, comments and ethics opinions for the jurisdiction that governs your lawyer, because some go further than notice. Whether text left under a failed redaction counts as "inadvertently sent" under your rule is a legal judgment for the lawyer, not for the person who noticed it.

Court rules and orders may apply as well. In federal civil discovery, Rule 26(b)(5)(B) sets out what a receiving party must do after being notified of a privilege claim over produced information, and a protective order or clawback agreement in your case may add its own procedure. The companion post covers that rule under running OCR on a document someone else redacted.

What the person who noticed should do

1. Stop reading the passage. Do not search for more instances.
2. Do not copy, export, forward or summarize the text.
3. Do not run the file through any other tool to see what else appears.
4. Write down what you saw in neutral terms: the document, the page, how you came across it, the date and time. Leave out the content.
5. Tell the responsible lawyer at once, in person or by phone if your office prefers that for sensitive matters.
6. List every place the file now sits: the review platform, shared drives, email, and any service it was uploaded to along with its exports.
7. Wait for instructions before deleting anything. Deleting may be right, or it may be the wrong step if the file has to be sequestered or returned.

What not to conclude

Finding recoverable text does not mean the sender waived anything, and it does not mean you may use what you saw. Those are questions of law that turn on your jurisdiction, your court and the orders in your case. The handling steps exist so that whatever the lawyer decides can still be carried out in full. If the file went to an online service, its deletion controls matter here; Schema Law's commitment that you can delete any document at any time is stated under security on the home page.