Can AI Be Trusted with Complex STEM Conversion? Where Automation Helps — and Where It Breaks
“Can’t you just use AI to convert it?” is now the first question on almost every publishing project. It’s a fair question — AI document tools have improved enormously in the last two years. It’s also the wrong question, because the honest answer isn’t yes or no. It’s where.
Here is the short version, for anyone who needs an answer before the meeting: AI is an excellent accelerator for STEM conversion and a dangerous replacement for human judgement. It produces a first draft brilliantly fast, and it makes confident, silent errors in exactly the places complex science can least afford them. The model that actually works is AI-assisted, human-verified — never AI-only.
Where AI genuinely helps
We use AI every day, and it earns its place. On clean, well-printed material, modern OCR and vision-language models extract text quickly and accurately, and they can produce a usable first-pass structure — headings, paragraphs, lists — to build on. For mathematics, tools now output a first draft of each equation as LaTeX, which is a real time-saver. Some engines also return per-page confidence scores that flag which pages look risky and need a human to check them.
In other words, AI is superb at the fast, high-volume 80% of the job: turning a stack of pages into editable, roughly-structured content, and giving an expert a running start instead of a blank screen. For simple, modern, text-led books, that first pass is often close to usable.
Used well, it also lifts some of the tedious parts of accessibility: drafting alt text for simple images, suggesting heading levels, and flagging obvious structural problems for a human to confirm. None of this is the finished product — but as scaffolding, it genuinely shortens the path.
Where AI breaks — the expensive part
STEM content is defined by the 20% that is hard, and that is precisely where today’s automation struggles:
- The meaning of mathematics. A model can often reproduce how an equation looks while misreading what it means — a subscript read as an exponent, a dot taken as multiplication instead of a decimal, an ambiguous or custom symbol guessed wrong. And a LaTeX or Markdown draft is not the valid, accessible MathML, with correct semantic structure, that a compliant e-book actually needs.
- Silent, plausible errors. The danger is not that AI fails loudly — it is that it fails quietly, producing clean, confident, well-formed output that is subtly wrong. Research on these systems documents both outright “hallucinations” and failure modes where a page degrades into repeated or invented text. A flipped sign or a dropped limit of integration in otherwise tidy output is far more dangerous than an obvious blank, because nothing signals that anything is wrong.
- Complex tables. Multi-row headers, spanned cells and footnoted values routinely get flattened or mis-associated, which breaks both the data’s meaning and its accessibility.
- Figures and alt text. AI can describe a photograph; it cannot reliably explain what a scientific figure means in its context — which is exactly what accessible alt text requires. Meaningful description of a graph, a circuit or a reaction scheme needs domain understanding, not image captioning.
- References, metadata and tagging. Mis-parsed citations, invented identifiers and mis-tagged elements tend not to surface until much later — at a hosting platform’s ingestion gate, or in an accessibility audit.
A concrete example
Consider a single line of a physics text. The model reads x² (x squared) as x₂ (x sub 2), or renders a vector product a·b with the dot silently dropped. Each is a one-character slip. Each is invisible in output that otherwise looks perfect. And each changes the mathematics completely — the sort of error a subject-aware editor catches in seconds and an automated pipeline ships without a flicker of doubt. Multiply that by the hundreds of equations in a single STEM title, and the scale of the risk is clear.
Why these errors are so dangerous in STEM
A broken image announces itself. A wrong exponent inside clean-looking text does not. In science, medicine and engineering, a silent error doesn’t just sit there — it propagates: into a student’s understanding, a clinician’s quick reference, a result someone tries to reproduce. The cost of catching such an error late — after publication, at ingestion, or when a reader relying on assistive technology hits it — dwarfs the cost of getting it right the first time. For scholarly and technical publishing, “mostly accurate” is not a standard; it is a liability.
The stakes are higher still for accessibility. When a reader who depends on a screen reader meets an equation that AI mis-encoded, they don’t get a slightly-wrong answer — they often get nonsense, or silence, with no way to know what was there. Getting the structure right isn’t a nicety for these readers; it is the difference between a usable book and a closed door.
The model that works: AI-assisted, human-verified
The answer isn’t to reject AI — it’s to put it in its proper place. At KPR Knowledge we let AI do what it is genuinely good at and keep a human exactly where judgement is required:
- AI accelerates: first-pass OCR and text extraction, a draft of the structure, candidate LaTeX for equations, and confidence scores that route attention to the riskiest pages.
- Experts verify and finish: our specialists check and correct the mathematics, rebuild complex tables, write meaningful alt text, encode valid accessible MathML, and apply the correct JATS/NLM or EPUB structure.
- Validation proves it: every file is checked with the appropriate tools — epubcheck for EPUB, ACE by DAISY for accessibility, schema validation for XML — so what ships is verified, not merely generated.
The automation does the fast, broad work; the expert does the small, decisive part that determines whether the output can actually be trusted. That combination is faster and safer than either humans or machines alone — and it is the only approach we would put our name on for complex STEM.
And the economics favour doing it properly. AI-only conversion looks cheaper on the quote, but the true cost arrives later: titles rejected at a retailer’s ingestion gate, a failed accessibility audit under the European Accessibility Act, corrections after publication, and the reputational cost of an error in a scholarly work. Fixing a silent mistake once a book is on the market is many times more expensive than verifying it before release. “Cheap and fast” that has to be redone is neither.
What to ask any conversion vendor
If you’re commissioning STEM conversion, these questions separate a trustworthy supplier from a risky one:
- Do humans verify the mathematics and tables, or is the output AI-only?
- How do you catch silent errors — what exactly is your QA and validation step?
- Do you deliver valid, accessible MathML and JATS, or just LaTeX and Markdown?
- Can you provide an accessibility conformance report (for example from ACE by DAISY)?
- Who handles the hard 20% — a subject-aware expert, or no one?
The bottom line. AI has genuinely changed STEM conversion — it has made the first 80% faster than ever. What it has not changed is who is accountable for the last 20%. Trust the automation to accelerate; trust a human to verify. That is how you get work that is both fast and correct — and it is exactly how we work.
Frequently asked questions
Can AI convert my STEM book automatically?
Partly. AI can produce a fast first pass, but a trustworthy, accessible final file needs a human to verify the mathematics, tables, figures and tagging. AI-only output should not be published in STEM without review.
Is AI accurate enough for mathematics now?
It is impressive at reproducing how equations look, but not reliable at their meaning, and it does not by itself produce valid accessible MathML. Every equation still needs expert verification.
Does AI produce accessible, EAA-compliant output?
Not on its own. It can draft text and alt text, but EAA/WCAG-compliant MathML, correct reading order and valid structure are standards work that requires human expertise and validation.
So should we avoid AI entirely?
No — that would be leaving real speed on the table. Use AI to accelerate the work, not to replace the judgement that keeps complex content correct.
This article reflects our practical experience converting complex STEM content; specific tool capabilities change quickly, so evaluate any automated pipeline on your own material.