ACCESSIBILITY NOTE · PUBLIC PDFS · 30 OCTOBER 2026

We read 25 public PDFs that had already been through accessibility remediation.

14 still had problems a screen reader would trip on.

If your agency pays a vendor or uses a tool to make its PDFs accessible, the file that comes back is not automatically one a blind resident can use. We checked files that state and local government bodies had posted, each marked as having gone through a remediation tool or service. More than half still had at least one error that would make a screen-reader user miss content or hear it wrong.

Then we tested whether an AI assistant could catch those errors before posting. When it could look inside the file at the layer a screen reader reads, it caught most of them. When it saw only the file's text and page images, as some chat assistants do with an uploaded PDF, it missed most of them. So the practical advice is short: when a remediated file comes back, check it before you post it, with a check that reads that layer.

What we read

We collected 25 PDFs posted by 6 US state and local government bodies. Each file's own metadata named a remediation tool or service. They included a sample ballot, a tax handout with charts, board documents with data tables, project reports with maps and photos, and lists of officials. We read pages 1 to 10 of each file's tag structure (the layer a screen reader actually reads) against a fixed checklist of five kinds of error:

content hidden from the tags, so a screen reader skips it;
an informative image or chart with no usable description;
a data table without headers;
a section read out of order, or under the wrong heading;
a link with a wrong or missing name.

One of our AI agents did the read, working through the tags with its own tools. It did not know which files were expected to fail. We also slipped in two other files from the same bodies, with errors we had planted ourselves; it caught both, with the right kind of error.

What we found

14 of the 25 files had at least one of those errors, 47 errors in all. This was a targeted sample of files with remediation-tool or service metadata; it does not establish an error rate for these bodies' other files.

Reading order was the most common problem: 24 of the 47. On the sample ballot, candidate names and party lines were read mixed together, a measure's question text was read in chunks taken from different columns, and one measure's table and YES/NO options were read after a different measure. In a list of elected officials, one official's second office was read after another official's offices, where a listener would take it for theirs.
Hidden content came next (10). On three pages of one board document, worked example tables and the conclusions drawn from them were marked as decoration, so a screen reader passes over them in silence. One of those pages had nothing on it a screen reader could read.
Images and charts (9). On a tax handout, two budget pie charts and the graphic showing how each tax dollar is divided were marked as decoration. Only the heading above them was readable, with nothing under it.
Links (3) and table headers (1) made up the rest. One email link was announced as a phone number.

The problem was not one body's. 5 of the 6 bodies had at least one such file.

Our own automated tag check flagged only 4 of the 14. A file that passes an automated check can still fail a screen-reader user.

Can your own AI assistant catch these?

We gave 6 of the files with errors, and 3 files in which our read had found none, to a general-purpose AI assistant that could read the file's raw contents, including with simple text-search commands, but was blocked from running scripts to parse or decompress the PDF, with one plain request:

The request

I work for a small US local government. We paid to have this PDF made accessible before we post it on our website, because of the ADA Title II web rule. Before we post it, please check whether a blind person using a screen reader can actually use it. Look at pages 1 to 10 (all pages if it is shorter). List every problem that would make a screen-reader user miss content or get it wrong, with the page number and a one-line description of what is wrong and where. If you find no such problems, write NO PROBLEMS FOUND.

It found errors in 5 of the 6 files, matching 18 of the 23 errors our reader had listed. The sixth check was explicitly incomplete because the assistant could not unpack the file's compressed parts without running a program. Among the 3 files where our read found no material errors, it raised 2 false alarms on one file. Two of those 3 reports were explicitly incomplete, including the one with the false alarms; the third said it could not check the exact wording inside the tags. An AI agent that could also run programs found errors in all 6, matching 22 of the 23, with no false alarms on the 3 comparison files.

Some chat assistants receive an uploaded PDF differently: as its text and a picture of each page, without the layer a screen reader reads. We gave the same request to an assistant set up that way, 3 times each, for 3 of the files with errors and 1 file in which our read had found none. Every time, it said it could not see that layer. It found the tax handout's hidden charts every time, because their numbers were missing from the text. On the board document it found one reading-order error that our own first read had missed, and we have added it to the totals above (the scores in this section use our earlier list); it never found that document's hidden tables. It never found the ballot's errors, and on the ballot it reported reading-order problems that are not in that layer. It did not call any file with errors clean.

When the assistant can read that layer, this is good enough to be worth doing on every file a vendor sends back. If its answer says it saw only the text or the page images, it has not checked what a screen reader reads, so its list, or a clean result, tells you little. Even at its best, it is not proof that a file is clean. It missed a few errors and flagged a few things that were fine, so treat its list as a list of places to look. A short listen with a screen reader on the pages that matter most, such as a ballot or a fee table, is the most direct check of what a resident will hear. Follow your agency's rules on AI tools; these files are meant to be public, but your agency's rules still apply.

What this does and does not show

This is a small sample, from the bodies our search reached. It is one structure read of each file, not a session with a screen reader. It is not a ruling on any vendor or tool. A remediation marker in a file shows that a tool touched it, not that the posted file is exactly what a vendor delivered: a body may have edited it afterwards. We did not test Acrobat or AI auto-tagging as a separate group; files without a remediation-tool or service marker were outside the sample.

It also does not mean every old PDF must be fixed. The rule has exceptions, including for many documents posted before the compliance date; our guide to the ADA Title II dates covers which date applies to you (April 26, 2027 for bodies of 50,000 people or more, April 26, 2028 for smaller ones and special districts).

We tag PDFs for public bodies too. Our own first builds had errors of similar kinds until a separate read caught them, and we fixed and re-read them. That is the point of this note: whoever made the file, check it after.

If you would rather have us tag a file and send a check report with it, our terms for tagged PDFs say which files qualify and what it costs.

Read on 30 October 2026. Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Corrections: hello@cybernative.ai. We will publish dated corrections on this page.