More and more of the resumes I see now were written by ChatGPT or Claude. People ask me whether that is a problem for applicant tracking systems. So on 13 October 2026 I stopped guessing and tested it: I asked both tools for resumes the way a normal job-seeker would, saved the files the way a normal job-seeker would, and ran every file through the scanner I built.
- Plain AI resumes parse well. ChatGPT and Claude both scored 95 out of 100 when I asked for an "ATS friendly" resume and saved it from Word. The only flag was year-only dates, and those came from my own prompt.
- "Make it modern" is where it breaks. Asked for a nicer look, both tools put the skills in a table. Their own PDF exports glued the cells together and scored 84 to 85.
- The writing is fine; the file is the risk. The same content as a Word file scored 99 to 100. Pasted into Word and saved as PDF, Claude's version scored 99.
- Claude's exported PDF also moved the table. The skills ended up in the middle of the work history, with "Page 1 of 2" in the text a parser reads.
- The test found four mistakes in my own scanner, which I fixed the same day. I describe them below, because you should know how I check.
How I ran the test
I made up three people, so nobody's real details went into either tool, and asked for each resume in the words a real user would type:
- A marketing manager in Austin, six years, two jobs. Request: "Write me an ATS friendly resume."
- A software engineer in Toronto, seven years, two jobs, an AWS certificate. Request: "Make me a modern professional resume with a skills table and nice formatting so it stands out."
- A newly qualified nurse in Manchester, applying for NHS Band 5 posts. Request: "I need a CV ... make it look professional and give it to me as a PDF."
I used ChatGPT on the free plan and Claude on a paid plan. Where a tool wrote text, I copied it into Word with an ordinary paste and saved it with Word's own Save as PDF. Where a tool offered a file, I downloaded that file. Then I scanned every PDF and Word file with my ATS scanner and read the extracted text myself.
Three resumes is a small test, and I can't run files through Workday or Taleo themselves. What I can show is exactly what a parser extracts from each file, which is where most rejections start. My scoring rules are on the methodology page.
Results
| Resume | Tool and file | Score | What the parser saw |
|---|---|---|---|
| Marketing, plain request | ChatGPT, pasted into Word (PDF and Word) | 95 | Everything in order; year-only dates flagged |
| Marketing, plain request | Claude, pasted into Word (PDF and Word) | 95 | Same |
| Engineer, "modern with a skills table" | ChatGPT's own PDF | 85 | Skills table glued: "LanguagesGo, Python, TypeScript" |
| Engineer, "modern with a skills table" | ChatGPT's own Word file | 100 | Clean |
| Engineer, "modern with a skills table" | Claude's exported PDF | 84 | Table glued and moved into the work history; page footer in the text |
| Engineer, "modern with a skills table" | Claude's exported Word file | 99 | Clean |
| Engineer, "modern with a skills table" | Claude, pasted into Word, saved as PDF | 99 | Clean |
| Nurse, "give it to me as a PDF" | ChatGPT, pasted into Word, saved as PDF | 99 | Clean |
| Nurse, "give it to me as a PDF" | Claude's own PDF | 99 | Clean |
Why the skills table broke
When you ask for a "modern" resume, both tools reach for a two-column skills table: category on the left, tools on the right. It looks tidy. But inside the PDFs the tools generated, there is no space between one cell and the next, so a parser reads the row as one word. "Languages" and "Go" became "LanguagesGo". "Data" and "PostgreSQL" became "DataPostgreSQL". A recruiter searching the ATS for "Go" or "PostgreSQL" would not find that candidate, even though the skill is right there on the page.
Claude's PDF had a second problem: the table was pulled out of its section and dropped into the middle of the work history, between two jobs, along with the page footer. Read top to bottom, the resume no longer made sense.
The same content in Word was fine, because Word keeps each cell's text separate. That matches what I see in thousands of scans: tables are a common cause of garbled text, and the damage depends on how the PDF was made, not on the table itself. More on that in PDF vs Word for ATS.
One thing that surprised me
When I asked ChatGPT's free plan for the nurse CV "as a PDF", it made one file and then told me I had reached the limit for creating files and should try again the next day. So a free user who wants a PDF ends up doing what I did: copying the text into Word and saving it themselves. Done that way, the file was clean.
What the test found in my own scanner
Being honest about this matters more to me than looking perfect. Running these AI resumes exposed four places where my scanner was wrong, and I fixed all four on 13 October:
- An all-capitals heading below a correctly read name ("CLINICAL PLACEMENTS") was reported as the person's name squashed into one word.
- A CV with education first, which is normal for a new graduate, was reported as a scrambled two-column layout.
- "APIs" was reported as two words glued together.
- Line breaks inside a paragraph of a Word file were dropped, which glued words together.
The scores in this article are from the corrected scanner. Each of those cases is now a test that runs before every release, so they cannot quietly come back.
What I would do if I were using ChatGPT or Claude
- Ask for plain. "Single column, no tables, standard headings" gets you a resume that parses. You lose nothing a recruiter cares about.
- If you want a skills section, ask for a comma-separated list under each category instead of a table.
- Prefer the Word file when the tool offers one, or paste the text into Word and save the PDF yourself.
- Keep the facts yours. Both tools write confident numbers. Check every figure against what you actually did; AI wording is a different question from AI facts.
- Scan the exact file you will send, not the draft. That is the only way to see what the parser sees.
If you're choosing between an AI chat and a dedicated builder, I compared the main builders in AI resume builders compared. And if you want a resume that looks designed without the table problem, my rebuild now comes with 13 designs, each checked against your own resume before it is shown.
