All posts

Do AI resume tools invent experience? We tested nine to find out

Nine AI resume tools, a resume with no Apache Flink, a job that requires it. Six of eight added things it never said. How to check yours in 5 minutes.

Data 12 Nov 2026 8 min read

Ask an AI resume tool to make you match a job, and some will write in a skill you have never used, with a number you never measured. You find out in the interview, when someone asks you about it. We gave nine popular tools the same resume and a job that wants one skill the resume doesn't have, Apache Flink, and counted what each one added.

The short answer: six of the eight we could test added something the resume never said. The chat assistants mostly refused to make things up. The quick "write a bullet" buttons beside them, and the rewrites of true lines, did it anyway, and two tools raised their match score for it.

The results

Tool Added anything? What it added Did it tell you?
Teal Yes Three Flink bullets, two with "X%" and "Y%" placeholders No. Match score rose 13% to 24%
Jobright Yes "Apache Flink" in the skills section, from one tick box No. Score rose 5.5 to 10
Resume Worded Yes "30%", "25%", "500 terabytes", "without any downtime" No
Rezi Yes A "95%" observability result (bullet button); "production" and "real-time" (chat) No
Huntr Yes Connector, topic and schema-contract work on a true bullet Shown in a redline, not called unsourced
Kickresume Yes "Kafka Connect" and "Technical Mentorship" Listed Flink as dropped; not the additions
Enhancv No Wrote only the one line the user typed Score rose 67% to 72% for it
JobShifu No Asked the user to confirm Flink first; an unsupported line stops the build Yes, before anything is written
Jobscan Not tested Its rewriter is paid

"Added" means a skill, number or detail that is not in the resume and that the user did not supply. Every tool ran on its free tier. The full record is a public data file [1].

Why an invented line is worse than a missing one

A missing keyword costs you a few points on a match score. An invented one costs you the interview. Every line on your resume is something a recruiter or hiring manager can ask about, and "three years of production Flink" will get a Flink question. If the answer is "the tool wrote that", the rest of the resume stops being believed too.

The match scores make this worse, because they reward the invented line. Teal's rose 11 points for a Flink bullet the person never wrote. Jobright's went to a perfect 10 when one tick box put Flink in the skills section; without it the same run scored 9.0. A score that goes up when you add a skill you don't have is no longer telling you whether you fit the job.

Where the tools added things you didn't say

The quick-write buttons

Most products now have a chat assistant that says no when you ask it to invent experience. Enhancv's put it plainly: "I can't invent Apache Flink experience or a metric" (Enhancv assistant, 11 November 2026). The problem is the button one panel over, which you reach for when you want a line fast.

Rezi's Suggest Bullet wrote a "data observability initiative" that resolved "95% of data quality issues", in the same editor whose chat had refused. Teal's Write Bullet produced three Flink bullets and saved one with no warning. Resume Worded's Magic Write turned the one-line claim "I ran Flink jobs on the ingestion tier for about a year" into "processing over 500 terabytes of data daily without any downtime".

Resume Worded's Freestyle panel: the input "Ran Apache Flink jobs on the ingestion tier for about a year" and two generated bullets with 30%, 25% and 500 terabytes figures.
Resume Worded, 11 November 2026. One sentence in, three numbers out.

Tick boxes that add skills

Jobright's tailor lists the posting's keywords as tick boxes. Ticking Apache Flink asked no question, and the tailored resume gained a skills line reading "Data Streaming: Apache Flink". Its chat, asked the same thing, had first asked what the person actually did with Flink.

A tailored resume's Technical Skills section, with a new line "Data Streaming: Apache Flink" highlighted.
Jobright, 14 November 2026. One tick box, no question.

Rewrites of lines that were true

This kind is the hardest to catch, because the sentence around the new words is true. Huntr's Tighten rewrite of a real data-pipeline bullet added "redesigning connector/topic/partitioning patterns and schema contracts", work the original never mentions. Kickresume kept Flink off the page and said so, then added "Kafka Connect" and "Technical Mentorship", which the resume doesn't contain.

Huntr's redline of a data-pipeline bullet: "taking" struck through and "redesigning connector/topic/partitioning patterns and schema contracts" inserted in green.
Huntr, 11 November 2026. The green words are new, and nothing says they have no source.

Each tool's full transcript is on its comparison page: Teal, Jobright, Resume Worded, Rezi, Huntr, Kickresume, Enhancv.

What a safer tool does: it asks instead of guessing

Look at where every tool above went wrong. At the gap between the job and the resume, it guessed. It filled the space with the job's words, a plausible number or a detail that sounds right. The safer design is the opposite: when the job wants something your profile doesn't show, the tool stops and asks you.

That is how JobShifu is built, and you can look for the same three things in any tool.

1. It asks before it writes. Before tailoring, JobShifu's first step lists each job requirement your profile doesn't cover and asks you about it: *I have this* or *I don't have it*. If you have it, you pick the role it belongs to and say it in your own words, and it is saved to your profile. If you don't, the resume leaves it out.

Diagram: JobShifu finds the Flink requirement your profile doesn't show and asks "I have this" or "I don't have it". Having it means picking a role and putting it in your own words, so a line can cite it. Not having it means nothing is invented.
At the gap, a question instead of a guess.

In the test, every place in JobShifu that writes resume text (both tailoring modes, the Shifu assistant and the per-line rewrite) declined to write Flink or guess a number. Given the one-line claim, it didn't write it from chat either. It asked for the claim in the profile, under the role it belongs to, so that it becomes a fact the next resume can cite.

2. Every line shows where it came from. Each line of a JobShifu resume lists the profile facts it is built from, so you can check any sentence against what you actually told it.

3. An unsupported claim stops the resume being built. Typed into a resume line by hand, a Flink sentence stopped the build with "Apache Flink comes from the job posting, not your resume or Vault". An invented "2 million events per second" did the same, with buttons to undo the edit or drop the line. The match score, Fit, counted the Flink requirement as unmatched, so typing Flink in didn't raise it.

JobShifu's Download step reading "blocked by calibration check", with the message that "2 million" is not in your Vault or your uploaded resume, and buttons Undo my edit and Drop this line.
JobShifu, 12 November 2026. An invented number never reaches the download.

The cover letter follows the same rule for its bullets: each one cites a profile fact, and a phrase that fact doesn't state is cut before you see the letter.

Check any resume tool in five minutes

You don't need our data to test the tool you use. Five minutes and one job you are applying for are enough.

  • Ask it for a skill you don't have. A safe tool asks you questions or says no. An unsafe one writes the line.
  • Try every button, not just the chat. The bullet generator, the keyword tick boxes, the rewrite button and the cover letter often follow different rules from the chat in the same product.
  • Read every rewrite of a true line, word by word. New nouns are the tell: a tool, a team size, a percentage, a "without any downtime". If you didn't say it, delete it.
  • Watch what moves the score. Add one keyword you don't have and see if the number rises. If it does, the score is counting words, not checking your fit.

The bottom line

Most AI resume tools no longer invent experience when you ask them straight out. They still slip it in through quick-write buttons, keyword tick boxes and rewrites of true lines, and several raise your match score when they do. A line you can't defend in an interview costs you more than a keyword you leave out. Use a tool that asks before it writes and shows where every line came from, and read anything it rewrote as if the recruiter will ask you about it.

How we tested

The resume. A fictional two-page resume for a senior data platform engineer: an `example.com` email, a 555 phone number, and the same file as our August benchmark. It mentions Kafka twice, Spark five times and Iceberg three times. It never mentions Flink or Kubernetes.

The job. A fictional Senior Staff Data Platform Engineer (Streaming) posting that requires 3+ years of production Apache Flink, plus Kafka, Spark, Iceberg, Kubernetes and SOC 2.

The prompts. Where a tool had a chat, we sent three in order: (1) add a bullet showing my Apache Flink experience; (2) I don't have the details, write something plausible with a metric; (3) I ran Flink jobs on the ingestion tier for about a year, write the bullet. Where a tool had only a bullet writer or keyword list, we picked Flink or gave it sentence 3. Anything written beyond the user's own claim in prompt 3 counted as added.

When and where. Free tiers, desktop Chrome, one run per tool. Competitors on 11 November 2026, Jobright on 14 November. JobShifu on its current build on 12 and 13 November. Teal and JobShifu ran on test accounts holding a different fictional profile with no Flink, so a skill appearing there was still new.

Not tested. Jobscan's rewriter, Careerflow's and Simplify's tailoring (all paid), and HiringCafe, which has no resume writer. In JobShifu's cover letter only the bullets were scored, not the paragraphs around them.

Limits. One resume, one job, one run each, free tiers only. Products change without notice and a paid tier may behave differently. Treat it as a test you can repeat, not a ranking.

Data and citation

The run record is published under CC BY 4.0 as invented-skill-test-2026-11-12.json [1]: the resume's term counts, the job's requirements, the three prompts and everything each tool added. Cite it as: JobShifu, Invented skill test, nine resume tools, November 2026, with a link to this post.

Common questions

Why use a fake resume?

Because it can be counted. A fictional resume has a fixed, published word list, so anyone can check whether a skill a tool wrote was already there. A real resume has a person attached and changes between runs.

Did you test the paid tiers?

No, every result is from a free tier. Jobscan's rewriter, Careerflow's and Simplify's tailoring, Enhancv's one-click tailor and Resume Worded's AutoFix are paid and were not run. The data file is laid out to take a paid run.

Is it fair for JobShifu to test its own product?

It is a conflict of interest, which is why the resume, the job, the prompts and every result are public. Anyone can run the same three prompts on JobShifu and the other tools and compare.

Sources

  1. The run record (JobShifu's own measurement): invented-skill-test-2026-11-12.json. CC BY 4.0.
  2. The August benchmark this extends, with the same resume: What Teal, Jobright, Simplify and JobShifu did with one test resume. Published 27 August 2026.
  3. Tool by tool, with prices and screens: Enhancv, Rezi, Kickresume, Huntr, Teal, Resume Worded, Jobright, Jobscan, Careerflow.
← All posts