PAC 2026.1 (26.1.0.0) · Checked
PAC 2026 "AI-assisted check" warnings on a PDF with no PDF/UA failures: fix or leave?
In PAC 2026.1, everything on the AI-assisted check tab is a warning, not a failure. It does not change the PDF/UA result, and PAC's exported PDF report leaves it out. PAC's own documentation says they "should therefore be understood as supporting information, not as official technical criteria", and that they may indicate incorrect tagging. We tagged three public reports and found no such miss: PAC never showed us where an item was, and by our reading every item we could read was a false alarm or a judgment call, so we changed nothing. This page shows how to tell a real miss from a false alarm.
The messages
This is the exact text PAC 2026.1 (version 26.1.0.0) showed in "Results in detail" for our files:
The AI-assisted check detected a Table element, but no corresponding structural element exists. Score: 0.87
The AI-assisted check detected a TOC element, but no corresponding structural element exists. Score: 0.83
A structural element is tagged as H1, but the AI-assisted check identified it as P. Score: 0.87
A structural element is tagged as H3, but the AI-assisted check identified it as P. Score: 0.85
A structural element is tagged as P, but the AI-assisted check identified it as H. Score: 0.69They sit under the "AI-assisted check" tab, which opens with "The AI has detected potential issues in this PDF document.", in three groups: "Classification of the elements", "Unrecognized structural elements" and "Missing structural elements". PAC's documentation calls the first group "Match of structural elements".
What PAC says they mean
From PAC's AI-assisted checks pages:
- Classification of the elements (Match of structural elements): the AI reads a region as a different element from the one it is tagged as, with a probability of 50% or more. Below 50%, PAC skips it.
- Missing structural elements: in PAC's words, the AI "detected an element in the document, such as a paragraph or an image; however, this region is not tagged in the document", again at 50% or more. In our county report the regions we matched to the Table items were tagged, as headings and paragraphs, and PAC still showed "detected a Table element". By our reading, the element type in the message is what the AI expected there, and the item means it found no structure element of that type, so we read the message more broadly than PAC's page does.
- Unrecognized structural elements: a tag where the AI sees no visible content. PAC's example is invisible text tagged as a paragraph. None of our files had one.
PAC also says that the AI recognizes tables and lists as a whole, not their rows, columns or items; that it does not check reading order or alternative text; and that "Individual AI results cannot be manually ignored or hidden". The pages describe a probability, and the items show a "Score". We read the score as that probability, but PAC doesn't say so in those words.
What we saw ourselves, in 26.1.0.0 on three files:
- PAC's PDF/UA result was 0 failed on all three, with 3 or 4 warnings in its Structure tree group, and the AI tab still showed its own warnings on each (3 and 5 on the two runs that finished).
- Every AI item was a warning, including scores of 0.83 to 0.87.
- The AI items were not in the PDF report that PAC exports.
- The item text names no page.
So someone who receives only your PAC report won't see these items. Anyone who opens the file in PAC 2026.1 with the default settings will see them if the AI flags anything; it did on each of our three files.
Older guidance says 80% is an error
The first release, PAC 2026 (26.0.0.0, 27 November 2025), could show AI items as errors. Its documentation listed a missing element at a probability of 80% or more as ERROR, and 50% to 80% as WARNING (archived 12 March 2026). The 26.1.0.0 release notes (25 March 2026) say: "The AI tab no longer displays errors; it now only shows warnings." The same release also lists a "significantly improved AI model".
If a webinar, checklist or blog post from before April tells you that a high score means a failure, that described 26.0. If you still run 26.0, the same file can give you different items, because the model has changed. We tested 26.1 only.
Decide each item
- Fix the PDF/UA result first. The AI items come in addition to it, not in its place.
Find the region. The item names an element type but no page, so look for content of that kind:
- "detected a Table element": in our files, ruled boxes with a side heading beside a paragraph.
- "tagged as H1 … identified as P": your H1s.
- "detected a TOC element": your contents page.
- "tagged as P … identified as H": in our files, a letterhead line, an "Attachments" line and running titles.
We couldn't test whether clicking an item in PAC's preview highlights its region, so try that first.
Ask one question per item: with the tag as it is, does a screen-reader user lose meaning or a way to move around the document?
- Table: does a cell's meaning depend on its row or column heading? If so, it is a data table: tag it Table, with header cells. If it is layout, such as a label beside a paragraph, leave it. Our reading is that a layout tagged as a table makes a screen reader announce rows and columns that carry no data relationship; we did not test this with a screen reader.
- Heading or paragraph: does the line open a section that a reader would want to reach with heading navigation? If so, it is a heading. If not, it is a paragraph.
- Contents page: if it is tagged as a list, in order, and each entry links to its section, our reading is that retagging it as TOC is a choice, not a repair.
- Fix the items that fail that question, then run PAC again. A warning you have ruled on can stay, because PAC doesn't let you hide it.
- Write down each call: the item text, the score, the region, your decision and why. Whoever checks the file next sees the same orange tab, or, with only the PDF report, nothing at all.
- If you think the AI is wrong, axes4 asks you, if possible, to send the PDF with an explanation to support@axes4.com (Q&A, 20 January 2026).
Since 26.1, a registry value named EnableAIChecks set to 0 turns the AI checks off (PAC's technical details). We haven't tested it. It changes only your PAC: a reviewer with the default setting still sees the tab.
What we found on three public reports
The regions below are our reading. PAC never showed us where an item was. We matched each item to content of its kind, using the tag tree and the page images.
| Report (our tagged build) | PAC 2026.1 PDF/UA result | AI tab | Item, our reading of the region, our call |
|---|---|---|---|
| County internal-audit report, 7 pages | 0 failed, 4 warnings | 5 warnings, 208 passed | 4 × "detected a Table element": the ruled boxes on pages 4 to 7, where a shaded side heading (Objectives, Scope and methodology…) sits beside a paragraph. This is a layout grid, tagged as headings and paragraphs. There are more such boxes than the 4 flagged. Left as is. "tagged as P … identified as H" (0.69): a letterhead line, "Attachments", or a running title. Every section on pages 2 to 7 already has its own heading. Left as is. |
| State agency annual report to the legislature, 9 pages | 0 failed, 3 warnings | 3 warnings, 243 passed | "tagged as H1 … identified as P" (0.87): the only H1 is the cover title. Left as is. "detected a TOC element" (0.83): the contents page, tagged as a nested list of links. Left as is. "tagged as H3 … identified as P" (0.85): a bold "Note about data" label over a call-out, or one of the appendix lines. Either tag reads correctly. Judgment call, left as is. |
| Federal inspector-general report, 18 pages | 0 failed, 3 warnings | not final in our capture | The AI check was still running at about half its checks (3 warnings so far). We didn't capture the item text, so we make no call here. |
That is 8 items read and none changed. These are three files, so the result says nothing about how often PAC's AI is right.
How we ran this
We used PAC 2026.1, version 26.1.0.0, from the installer on axes4's download page, with the AI checks on (the default). We ran it on Windows Server 2025 and under Wine 11 on Linux, with PAC itself blocked from the network. The files were our own tagged builds of three public reports.
- The Windows runs gave the PDF/UA results and the AI tab counts. The item text and scores come from PAC's "Results in detail" window under Wine, where the same counts showed. For the county report, the Wine run used an earlier build of our file that differs only in whitespace inside the page content; the final build gave the same counts on Windows.
- We read the exported PDF report under Wine. Under Wine, PAC's page preview doesn't draw, and on Windows our automated runs didn't open the item rows. That is why we have no item locations.
- Not tested: PAC 2026.0, the
EnableAIChecksswitch, whether the preview highlights an item, and files other than these three.
Sources: PAC's AI-assisted checks and technical details pages and release notes, read 4 November 2026; the archived missing structural elements page of 12 March 2026; axes4's Q&A of 20 January 2026. When PAC releases a new 2026.x version or changes these pages, we re-check this page and date the change here.
Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something here is wrong, write to hello@cybernative.ai; we correct this page and date each correction.