AI CV formatting: where facts go wrong and how to catch it

AI CV formatting is fast and usually right, and "usually" is the problem. The errors that hurt a recruiter aren't typos. They're quiet changes of meaning: a withdrawn degree shown as if completed, a missing end date filled in as "Present", "basic Spanish" turned into a formal language level, or a city from one job attached to another. In our October 2026 tests on a synthetic CV, we saw each of these. The answer isn't to avoid AI. It's to make every extracted fact checkable against the passage it came from, and to have a human approve the exact version that goes to the client.

Why AI CV formatting changes facts

Large language models are good at producing fluent, plausible documents, which is exactly why their mistakes are hard to spot. Five mechanisms cause most of the trouble:

• Normalising to the common case. Most degrees listed on CVs were completed and most jobs have end dates. A model filling a structured template drifts towards the typical pattern, so "withdrew, not awarded" can fall away.
• Templates that demand a value. If the output layout has an "End date" field, something has to go in it. "Present" is a tempting default.
• Reading order in complex layouts. Two-column PDFs, sidebars and tables can be read in the wrong order, which attaches a skill, date or location to the wrong job.
• Mapping to taxonomies. Converting free text into framework levels, totals or categories creates derived facts the candidate never wrote.
• Rewrite and summary modes. Asking a model to "improve" a CV invites stronger verbs, rounded numbers and added skills.

None of this is malicious, and much of it is fixable with good design. But it means "the AI read the CV" isn't the same as "the CV says this".

What our October 2026 tests observed

We put several recruiter-focused formatting tools through free accounts and public demos using a synthetic CV built to be awkward. It had two jobs with exact dates, a six-person reporting line, a documented pilot that cut processing time from 12 days to 9, an MSc marked "Withdrew in May 2018; degree not awarded", a completed BSc, three skills, "basic conversational Spanish" and fictitious contact and referee details. No real candidate data was used. The full method and limits are in we tested CV formatting tools.

To be fair to the tools involved: this was one synthetic CV on one day, and products change quickly. One tool kept the withdrawal wording through parsing and PDF export. Both tools that dropped it restored it in both output formats once we corrected the field. Recovery was good where we tried it: one tool's Undo and Redo worked cleanly, and another regenerated both files after a correction. We didn't measure overall accuracy rates, and we didn't run CVPitch in the same session, so this isn't a ranking. What it does show is where to look.

Which errors matter most

Not every mistake is equal. Grade them by what happens if the client sees them:

Don't let a tool's polished layout average away a critical error. One changed qualification undoes ten beautiful page breaks.

How source evidence catches it

The most reliable defence is structural: every extracted fact carries a pointer to the passage it came from. That changes review from "reread both documents" to "check each fact against its source in one click". It's how CVPitch works:

• Evidence beside every fact. Each employer, title, date, qualification and skill links to the exact passage in the original. For PDFs, that's the real page.
• Unsupported facts are marked. If a value has no supporting passage, it's labelled unverified rather than presented as fact.
• Uncertain items are flagged. An unclear date range becomes a review question, not a guess.
• Recruiter additions are labelled. Your notes, the candidate's notice period and transcript notes are kept apart from the candidate's own words.
• Approval is tied to the exact revision. Edit a fact, switch template or change contact removal afterwards, and the submission needs approving again.
• Layout is deterministic. AI extracts the facts into fields. The template then lays them out by fixed rules, so changing layout can't rewrite content.
• Scans are flagged. When a CV is a scan or photo, the AI transcription becomes the source, and the CV carries a review flag until a recruiter has checked names, dates and figures against the image.
• AI writing is opt-in and reviewable. Optional rewrite, summary, proofreading and translation tools return proposals you accept change by change. Rule-based checks drop suggestions that change protected facts such as employers, titles and dates or add numbers not in the CV, and accepted wording is labelled AI-assisted.

Evidence has an honest limit. It shows where a fact came from, but it doesn't prove the interpretation is right. A recruiter still has to read "withdrew" and confirm it made it into the education entry. Evidence just makes that check fast.

The same principle applies to matching candidates against a brief. CVPitch's job match marks each requirement in a job description as evidenced, partly evidenced or not shown in the CV, with the supporting passage beside it. It doesn't produce a fit score. Because each mark shows its passage, a sentence like "no Python experience" can't hide behind a percentage. See how to pitch a candidate with evidence.

A ten-minute accuracy test for any tool

Before you trust a tool with live submissions, run it on a synthetic CV that contains these lines:

• An MSc with "withdrew, not awarded".
• A role with only a start year and no end date.
• "10+ years" of a skill.
• "Basic Spanish" or "conversational French".
• "No Python experience".
• A location given for one job only.
• A numeric claim, such as "reduced processing time from 12 to 9 days in a pilot".
• A two-column PDF layout with skills in a sidebar.
• A phone number in the footer and a LinkedIn hyperlink.
• A referee's name and email.

Then check the exported files, not the on-screen preview, and record anything missing, changed or added. Count critical changes separately from layout issues. If you're choosing between tools, our buyer's guide to CV formatting software has the full trial method.

Using AI safely in your formatting workflow

• Use AI to extract and arrange, not to write claims. Turn off rewrite and "enhance" modes for client submissions, or keep their output clearly labelled.
• Review qualifiers first. Education status, employment type, date precision and negatives are where meaning changes.
• Keep your words separate. Your assessment goes in a labelled summary, never blended into the candidate's experience.
• Make one named person approve one exact version. The person who checked it should be approving the file that's sent.
• Know what a parser can and can't do. Read CV parser vs CV formatter if you're relying on your ATS's parser output.

Frequently asked questions

Is AI CV formatting accurate?

Often, but not reliably enough to send without checking. In our tests, qualifications, date ranges, language levels and per-job locations were where meaning changed. Use a tool that shows the source passage for each fact so the check is quick.

Can AI formatting tools invent information?

They can add or alter information. In our tests that included a formal language level the candidate never stated, a location attached to the wrong job and an end date shown as "Present". Rewrite modes can also add skills or stronger claims. Treat anything without a source passage as unverified.

How do I check an AI-formatted CV quickly?

Check each fact against its source passage. Look at education status, employment type, date precision and any numbers first. If your tool doesn't link facts to sources, compare the documents line by line, starting from the original.

Does CVPitch use AI?

Yes: for extracting the facts from the CV into structured fields, for reading scans and photos, and for optional writing tools such as rewrite, summary, proofreading and translation, which only propose changes for a recruiter to accept. The layout is applied by fixed template rules, every fact links to its source passage, and a recruiter approves the exact revision before anything is exported. Step-by-step details are in the guide.

Bottom line

AI makes formatting fast. Source evidence makes it safe. Choose tools that show where every fact came from, test them on awkward synthetic CVs, and keep a human approval on the exact version that leaves your agency.