How to check a CV or resume for AI writing
- Paste the CV, resume or cover letter into CV / resume text, or click Upload file and choose the candidate's .pdf, .docx or .txt file. The text is extracted in your browser.
- Check that the extracted text looks complete. Unusual PDF layouts can lose sections, and scanned CVs contain no text at all.
- Click Check CV. The first check downloads the detection model (about 126 MB), which your browser then caches.
- Review the overall score, the CV-specific signals, the highlighted passages and the general writing-style signals.
Nothing is uploaded. CVs contain personal data, so the whole analysis, including reading PDF and Word files, happens on your own device.
What the score is based on
The score comes from the same models as our AI Detector. The main model is TMR, an open-source RoBERTa classifier trained on the RAID benchmark, which covers text from 11 language models including paraphrased and deliberately disguised output. The CV is analysed in passages of roughly 160 words, each passage is scored, and the passages are highlighted at 80%+, 50–80% and 30–50%.
If the main score is borderline (between 50% and 97.5%), a second RoBERTa model trained on the HC3 human-vs-ChatGPT dataset gives a second opinion and gets most of the weight. That step exists to reduce false positives on human writing, and the result shows both scores when it is used.
In our own small test of 32 texts (24 genuine human texts and 8 AI texts in different styles), all 8 AI texts scored 95–99% and 23 of the 24 human texts scored under 50%; one informal human comment was wrongly flagged. That test did not include CVs and is not a benchmark.
CV-specific signals
These figures sit next to the score to help you judge the document. They do not change the score.
| Signal | What it measures | How to read it |
|---|---|---|
| Buzzword density | Generic resume buzzwords per 100 words, such as "results-driven", "proven track record", "spearheaded", "cross-functional" or "stakeholders" | Above 3 per 100 words is high. Chatbots lean on these words, but so do many CV templates and career coaches. |
| Buzzwords found | The buzzwords that appear, with counts | Useful for feedback to a candidate, or for spotting a CV that says little in many words. |
| Specificity | The share of lines (or bullets) that contain a number or metric | Under 25% means few concrete figures. AI-written CVs are often vague, because the model does not know the candidate's real results. |
| Bullet uniformity | How much bullet lengths vary | Very even bullet lengths across five or more bullets are a common pattern in generated CVs. |
A note for recruiters
Many candidates use AI to polish their CV, fix their grammar or translate it into English. That is not dishonest in itself, in the same way that asking a friend to proofread a CV is not dishonest. What matters is whether the experience and results are real.
Use a high score as a prompt for questions, not as a reason to reject. Ask the candidate to describe a project from the CV in their own words, explain how a figure was measured, or tell you what went wrong and what they would do differently. Someone who did the work answers easily, however the CV was written.
Be especially careful with candidates who write in a second language. Non-native English tends to be more regular and predictable, which pushes AI scores up for people who wrote every word themselves.
Why CVs are hard to judge
A CV is close to the worst case for any AI detector. It is short, built from lists rather than prose, and written in a formulaic style that both people and chatbots copy from the same templates. Expect more uncertain results here than with essays or articles.
Tips
- Check the cover letter separately. More prose means a more reliable score.
- If the PDF text comes out jumbled, convert it with PDF to Text or PDF to Word and paste the cleaned text.
- Compare two applications that look suspiciously alike with the Plagiarism Checker; it finds shared wording and paraphrased sentences.
- Job seekers: if your own CV scores high, replace generic phrases with concrete numbers, tools and outcomes. That makes it stronger for human readers too.
Frequently asked questions
Is the candidate's CV uploaded anywhere?
No. The file is read and analysed in your browser. Neither the document nor the extracted text is sent to our server.
Should I reject a candidate because of a high score?
No. The score is a probability with known false positives, and using AI to polish a CV is common and not dishonest by itself. Use it to decide what to ask in the interview.
Does it work for CVs in other languages?
The models are trained on English. A CV in another language may get a score, but do not rely on it.
Why did a CV I wrote myself get a high score?
CVs are short, list-based and formulaic, and buzzword-heavy templates read like chatbot output. Non-native English also raises scores. Adding concrete details and varying your sentences usually helps.