Skip to main content
School of Engineering JudgmentThe verification gateA mapping call, at no cost
WhatsApp058-4887876
A mapping call, at no cost

All articles

Reading a table from an image: where AI gets it wrong

Published

Table in a standard: 45.27 and 40.65 of 100. Same content as text: 98.94 and 87.27. Check every value.

A generative AI tool reads a table far less accurately than it reads the same content as text. So a value that came out of a table in an image stays a draft until someone opens the original document and checks it. This is failure class G1, table from image.

Tables are where that error has a cost. Unit tables, code tables and data sheets usually arrive as images. The value read from them goes into a report, a calculation or a purchase order.

What the benchmark measured

A peer-reviewed benchmark called AECBench tested nine large language models on 4800 questions (Liang et al. 2026). Engineers in architecture, engineering and construction wrote the questions. Two of the tasks required reading a table inside a building code.

Across two measured tasks, accuracy in percent: 45.27 against 98.94, and 40.65 against 87.27.

The first figure in each pair is baseline accuracy. The second is the accuracy when the model got the same table as running text written by an expert. That gap is the finding: the models knew the content, and failed to read it out of a table.

The study concludes that the models show deficits in interpreting knowledge from tables in building codes. Drawings show the same pattern. A study of HVAC piping-and-instrumentation diagrams concludes that directly applying language models to digitising them is highly challenging (Lu et al. 2025). Usable results there needed an engineered pipeline, and an uploaded image with a question was not enough.

Three things these numbers do not say, and they are worth saying out loud.

So remember the pattern and leave the number behind. A tool can know the material and still read the wrong row.

What it looks like on the job

The example below is built for illustration. It is not a measured case.

A technician photographs a data table from Maker A and asks for the value on his model's row. The tool returns a number and units, in a confident tone. In the same table as text, the value on that row is different. The tool read the neighbouring row.

Nothing in the answer flags this. The number is plausible, because the neighbouring row is a similar model. The units are right, because they come from the same column. Anyone who glances at it and says "looks fine" has checked nothing.

What does not help

Three habits look like checking and are not.

Asking again. In a study on licensing exams, the answer consistency of one tool was 97 percent. So a wrong answer is wrong on the next attempt too (Choi, Lee, and Kim 2025). The same study classified 172 errors, and 20 of them came on questions that require reading a figure or schematic.

Asking another tool. A tool is never an authoritative source, and the output of a second tool is not one either. Two tools that agree are still two outputs.

Getting suspicious only when something looks odd. The output carries no external marker of error, so the check cannot depend on suspicion. It runs on every value and leaves a record.

The check to run tomorrow

Take one table you are working with this week and run this sequence on it.

  1. Before you open the tool, read the row yourself and write down the value you expect. An estimate you did not write down does not count.
  2. If the table exists as text, give the tool the text and leave the photo out. The benchmark found the gap between a table in a standard and running text. We have no figure for a photo, and text is the careful choice.
  3. Ask the tool to write the model, the column and the unit next to each value. Add the instruction: do not invent, write MISSING.
  4. Open the original document and check four things: the row, the column, the unit, and the footnote if there is one.
  5. Log one result of four: match, gap, not verifiable or missing. Correct a gap to the source and log it as G1.

Step 3 does not replace step 4. It only shortens it, because it tells you where to look.

What if there is no original document, only the photo? The gate blocks the value. Mark it "not verified" and keep it out of the deliverable. If the deliverable depends on it, the deliverable stays a draft. That is the outcome the method specifies, and it is not a failure of the method. In that case, get the data sheet from the manufacturer or measure directly.

What to record, and where

For most work orders three fields are enough: source, value checked, and who checked. The template is at work order in 3 fields.

A signed deliverable records more, in a log of 13 fields. Three of those do the work on a table: what the tool said, verbatim. The name of the authoritative document. The location inside it: page, clause or table. An auditor can then open the same table and see the same value.

The level follows the consequence of an error, and the difficulty does not enter into it. Reading a value off a table sounds simple, and that does not lower its level. A simple calculation that enters a signed survey is level 0, no tool at all. A table value going into the same survey deserves the same thought before you open the tool.

Who needs to know this

Whoever orders the part, whoever signs, and whoever receives the report. A technician who copied a value from a table sees no problem, because the number looks plausible. The person who signs after him sees none either, if the report carries no source and location. That field does more than any talk about being careful: it shows the next reader where to look.

Where the check stops

The check does not make the deliverable correct. If the source is wrong, the deliverable is wrong, and the log shows exactly where.

It needs the competence to read the table. Someone who cannot tell two columns apart will miss the error with the original document open in front of them.

It also adds time to every deliverable. Whether that time is less than the cost of a signed error is an empirical question, and we have no figure for it yet.

And the evidence here measures the failure, not the check. The working paper behind the taxonomy is not peer-reviewed, and it reports no collected data. G1 is one of seven failure classes, and one of the two that rest on peer-reviewed research.

References