Dima Bolkunov says he reads AI-generated text at least a third slower than human-written text of the same length. Add the verbosity most models default to, and total reading time climbs higher still. His point is simple arithmetic: it does not take many readers for the writer’s time savings to disappear on the other side of the document. In a plant, those readers are your shift leads, your auditors, and your CAPA owners.

That makes AI writing an operations problem, not a style debate. Ambiguous phrasing in a deviation report becomes a question, then an iteration, then a finding. This article breaks down where the cost actually lands in regulated manufacturing, and how to use AI for documentation without pushing the work downstream.

The Deviation Report Nobody Can Act On

You have read this document. The grammar is clean, the structure is tidy, and three paragraphs in you still cannot tell what went wrong, which batch was affected, or who decided to release it. A supplier email that says “we are addressing the root cause through enhanced process controls.” A shift handover that mentions “some variability on line 3.” Nothing to act on.

Bolkunov explains why. A model has to make assumptions to stitch text together, and those assumptions push it toward “more ambiguous, less precise language to avoid logical mistakes.” The writer rarely notices, because the exact meaning is already in their head.

Your quality system does not run on intent. It runs on what the document actually says, and ambiguity there is rework waiting to be scheduled.

What Bolkunov Actually Argues About AI-Generated Text

Bolkunov sets a hard definition up front: AI writing means text with at least 50% of the words generated by AI. That keeps the argument concrete. Using a model to tighten a sentence you wrote is not what he is talking about. Handing a prompt to a model and shipping the output is.

His background matters here. His bachelor’s thesis was about identifying an author from a short paragraph by comparing word usage statistics, and it worked. Readers pick up far more than word choice, including sentence structure and how someone tells a story.

Why averaged style strips out the context a reader needs

AI flattens style toward an average, and prompt tricks do not fix it. Once the style goes, so does the author behind the text. Bolkunov’s point is that style is not decoration, it is signal.

Even if you have never met the writer, their phrasing tells you something about them and the situation they were in. That is the irony he lands on: the tool built to add context ends up removing it.

The ambiguity trap: why the writer is the last person to notice

Precision drops because vague language is the safer bet for a model trying to avoid logical mistakes. The writer cannot see the gap. They hold the exact meaning in their head, which quietly gives an ambiguous word a specific one.

Writing also forces you to challenge your own thinking. Skip to the output and you skip that check entirely.

The Reader-Side Cost: One Writer, Forty Slower Readers

Where the savings actually land on a P&L

An engineer saves ten minutes drafting a work instruction. That saving is real and it sits on one person’s calendar. The cost lands somewhere else entirely.

Forty people read that instruction: operators on three shifts, a line supervisor, a quality engineer, an internal auditor. If each loses three minutes to slower reading and padded prose, you have spent two hours to save ten minutes. Bolkunov puts it plainly: “You do not need many readers for the writer’s time savings to evaporate on the readers’ side.” In a plant, documents are read dozens of times and written once. The ratio is never in your favour.

Second-order effects: questions, iterations, and lost trust

Then the slow leak starts. Someone Slacks the author asking what “enhanced monitoring” means in practice. A collaborative SOP goes through two extra review cycles because reviewers are arguing about wording rather than process. None of that time shows up on the writer’s side of the ledger.

Trust erodes last and costs most. Once your team learns that a given author’s documents require decoding, they start skimming or ignoring them. That is a quality risk. A work instruction nobody reads carefully is functionally a work instruction that does not exist, and no audit trail will save you when that becomes obvious.

The Thinking You Skip When You Skip the Draft

Bolkunov makes one point that should worry anyone who signs off on investigations. Writing something down is how you find the holes in your own reasoning. Gaps and inconsistencies that feel resolved in your head become obvious the moment you try to put them in a sentence.

AI writing takes away this opportunity, as you jump directly to a result.

That matters most in documents whose entire value is the reasoning. A process validation rationale. A capital investment case. A 5 Whys that has to survive a regulator asking why five actually becomes four.

Root cause analysis and CAPAs as thinking artefacts, not paperwork

A CAPA is not a form. It is the record of an argument: this happened, because of this, so we changed this, and here is how we know it worked. If the argument is weak, the document is worthless no matter how clean the prose reads. Auditors are trained to attack the logic, not the grammar.

Generated text hides weak logic very well. The model smooths over the jump between a cause and a corrective action because smoothing is what it does. Your investigator never feels the friction that would have told them the causal chain breaks in the middle.

Use the model after the thinking, not instead of it. Draft the causal chain yourself, in rough sentences, then let a tool tighten the wording and check for gaps.

The Counterarguments Worth Taking Seriously

Standardisation as a feature: when interoperability beats voice

A commenter named PEG asks whether the same argument was made against dictionaries and the standardisation of spelling. Standardisation reduced variation, the point goes, but made writing more interoperable. That is a fair hit, and in manufacturing it lands harder than in literary criticism.

Nobody wants personal voice in a calibration record or a batch release checklist. Controlled documents exist to be identical across shifts, sites, and languages. If a model produces a work instruction that matches your template every time, that is a win. The distinction is between documents that transmit a fixed procedure and documents that carry judgement. Standardise the first. Protect the second.

Is imprecision a model limitation or a structural one?

Tejasvi S. Tomar argues the deeper issue is temporary: “Other issues like imprecise language, critical thinking, elaborate prose are temporary limitations of current LLMs and prompting.” He also says he used AI to read the article, which is the honest version of where this is heading. Readers will offload reading the same way writers offloaded writing.

Partly right. Verbosity and clumsy prose are fixable with better prompting and tighter models. The part that will not be prompted away is that a model has to guess at intent it was never given. If the writer never specified which batch, no model invents the right answer. That is a structural gap in the input, not a flaw in the output.

Where to Point AI at Documentation Instead

A simple rule: AI reads your documents, humans write the conclusions

Models are excellent at the work nobody wants: pulling fields out of inspection records, classifying complaint text, translating a work instruction into Polish, finding the three deviations from 2023 that match the one on your desk this morning. That is reading at a scale no quality engineer can match.

Writing is where it breaks. A commenter on Bolkunov’s post, Tejasvi S. Tomar, said he used AI to read the article. That is the right instinct applied backwards. Let the model compress input. Keep humans on output that carries a judgement, a root cause, or a signature.

What to automate first in a quality or ops function

Start where the source data is already structured and verified, so the model is extracting rather than inventing.

• Field extraction: pull batch numbers, dates, and equipment IDs from scanned records into your QMS.
• Defect clustering: group free-text operator comments to surface recurring failure modes nobody logged as a trend.
• Retrieval before investigation: surface prior CAPAs and audit findings on the same process step in seconds.
• Translation of approved text: same content, more languages, no new claims.

Then draw the line. Anything that states a conclusion, assigns a cause, or gets signed stays human-authored. Use the model’s summary as input to your thinking, never as a replacement for it.

What Teams Will Standardise on by 2027

Writing at work is going to split into two lanes, and the split will be written into quality systems rather than left to individual judgement. Lane one is throughput text: meeting notes, status summaries, translations, first-pass supplier correspondence. Averaged style is fine there because nobody is making a release decision off it.

Lane two is judgement-bearing documents. Investigations, validation rationales, CAPA effectiveness checks, management review conclusions. Expect SOPs that name these document types explicitly and require human authorship with a traceable owner. Expect disclosure norms too, a simple flag on internal documents stating how much of the text was generated, because Bolkunov’s own threshold (at least 50% of the words generated by AI) is the only line anyone has proposed that is actually measurable.

Three questions to settle before your next documentation rollout

First: which document classes are allowed to be machine-written, and who signs off on that list? If the answer is “whoever is drafting decides,” you do not have a policy. Second: how do you measure success? If your business case counts drafting hours saved and ignores reading hours added, the number is wrong before you start.

Third: what happens when an auditor asks who wrote this and why it says what it says? That question has one acceptable answer, and it involves a named person who can defend the reasoning. Settle these three now. They get much harder to settle after a few hundred documents are already in the system.

Source: bolkunov.com