Why AI CV builders hallucinate

What hallucination means in a language model

Hallucination is the researchers' word for fluent output that the input does not support. The survey by Huang and colleagues (arXiv 2311.05232, 2023) separates factuality hallucination, where the text conflicts with the world, from faithfulness hallucination, where it conflicts with the instructions or source it was given. A CV from ChatGPT, Claude, Gemini or an AI CV builder built on them can suffer both; the second matters here: the model had your notes and wrote something they do not say.

The mechanism is ordinary. A language model produces the next most likely word. It has read millions of CVs and job adverts, so the likely continuation of "Sales manager, 2021 to 2024" is a sentence with a team size, a revenue figure and a percentage. Nothing in the model checks whether those numbers belong to you.

What a hallucinated CV line looks like

• It is specific. Invented lines carry precise numbers ("reduced churn by 23%"), because strong CV bullets do.
• It is plausible. The invented title, tool or project fits the job description, which is why you do not notice it on a quick read.
• It is confident. It reads in the same tone as the lines from your notes; nothing signals that one is sourced and the other is not.

Why AI CV builders hallucinate more than other writing tasks

A CV is a short document about a person the model has never met, close to the worst case. Four causes stack up.

The prompt is thin

Most people give a job title, an employer and two dates. The model needs a page of bullets and has a paragraph of facts, so it fills the gap with the average of every similar CV it has seen. Less input, more invention.

The instruction is "make this more impactful"

Most AI CV builders, and every chat prompt, ask for stronger, more quantified bullets. Inside the model, "more impactful" means "add a number and a bigger verb". Given "helped with the migration", the improvement is "led the migration of 40 services with zero downtime". The tool did what it was asked; the user asked for something they did not have.

The model was trained on generic CVs

The training data is dominated by sample CVs and templates written to look good, not to be checked. The model learned the shape of a good bullet, including the fabricated metrics that fill sample CVs, not the habit of asking where a number came from. It is also why AI-written CVs converge on the same phrases: results-driven, spearheaded, cross-functional.

There is no source to check against

A chat model keeps no separate record of your facts; your notes and its guesses share one context window, and each sentence it writes becomes context for the next. A grounded AI CV builder puts that record outside the model and checks the output against it.

The three typical fabrications

The invented content falls into three groups. The examples are constructed for this page; none is a real candidate.

The pattern is the same each time: a true, vague statement goes in and a false, precise one comes out. The precise version scores better on keywords and reads better to a tired recruiter, which is why tools keep producing it.

Why recruiters catch it and applicant tracking systems do not

Applicant tracking systems such as Workday, Greenhouse, Lever, Taleo and SuccessFactors parse a CV into fields and match keywords. Their documentation describes parsing and ranking, not a check that a number is true or a detector for machine text. A fabricated metric passes the software without friction.

The check happens later, in a conversation. Harvard's Office of Career Services and MIT's career advisers give the same instruction from the candidate's side: bullets should be specific, quantified where a real number exists, and defensible in an interview. Prospects, the UK graduate careers service, says the facts on a CV are easy to corroborate and that AI should not be used to invent or embellish details.

• The screening call: "you cut churn by 23%, how was that measured?" A candidate who did not write the number cannot answer.
• The technical interview: "tell me about the Snowflake schema" has no answer if the model added the tool.
• The background check: in finance, healthcare, law and the public sector, titles and dates are compared with records, and a mismatch is misrepresentation.

What grounding means technically

Grounding is the machine-learning term for tying a model's output to a given source rather than its training data. An AI CV builder needs four parts; a tool with only the first is making a promise, not enforcing one.

1. A source of truth: a structured record of the candidate's facts (roles, dates, results, tools) kept outside the model, so there is something to check against.
2. A claim-to-source link: every line carries a reference to the facts it was written from, so a reader or a program can look them up.
3. A verifier that rejects unsupported claims: a check, ideally not a language model, that every number, date and named entity on the line exists in the cited facts, then that the whole sentence claims no more than the facts do.
4. A refusal path: when the facts do not cover what the job needs, the tool asks a question or leaves the line out. It never fills the gap itself.

Generic AI writer vs grounded AI CV builder

How to test any AI CV tool for hallucination in five minutes

Run this on any tool before you trust it, whether it calls itself grounded or not: ChatGPT, Rezi, Teal, Kickresume, Enhancv, RecastCV, ResumeFab or CitedCV. A tool that fails step one fails the category.

1. The empty test. Give only a job title and a company, no facts, and ask for a CV. A grounded tool refuses or asks; a generic one writes a CV for a person who does not exist.
2. The vague-bullet test. Enter one true, vague line such as "helped improve the checkout page" and ask for it to be made stronger. Count every number, tool and verb you did not type.
3. The source test. Ask where each number in the draft came from. A grounded builder shows a fact per line; a generic one admits it estimated or explains the number away.
4. The scope test. Enter "contributed to" and see whether "led", "owned" or "managed a team" comes back.
5. The tool test. Mention one tool once, paste a job description listing five, and count how many it attributes to you.

Score it simply: any invented number, title or tool in steps two to five is a fail. The test measures not how well the CV reads but whether you can defend it.

RecastCV publishes a five-prompt version of this test, and RoleWorth's 2026 comparison reports that Rezi, Teal, Kickresume, Resume.io and Enhancv will write a bullet without evidence if asked. The comparison page below applies it to each tool.

How CitedCV implements grounding

CitedCV is built on the four parts above. The writing model is MiniMax M3, called with the confirmed facts only, never with a chat history or a job description alone.

1. A facts ledger. Every fact comes from you: a form, a PDF, DOCX or TXT read in your browser, or your interview answers. Each fact gets an id.
2. Per-line citations. Every bullet carries the ids of the facts it was written from, visible in the app.
3. A deterministic verifier. Every number, date and named entity in a bullet is checked against the cited facts by code, not by a model.
4. An entailment check. A language model then checks that the whole sentence claims no more than its facts, so "took part in" cannot become "led".
5. Drop instead of ship. A failing bullet is rewritten once from the same facts; if it fails again it is removed. Nothing unverified reaches the export.
6. Interview questions. When the facts are too thin for a section, CitedCV asks before writing rather than guessing.

What it exports and what it costs

A filler linter scores each CV for human-likeness and flags stock phrases. Templates are single-column with standard headings; each PDF or DOCX export is re-parsed to confirm it reads back. Regional conventions cover the US, UK, EU, Turkey and Germany; the interface is English and Turkish. A new account has 5 free credits; a full build costs 10, a quick draft 6, a line edit 1; exports are free; plan prices in US dollars are on the pricing page.

The limits of a grounded builder

• It cannot add experience you lack. If your facts are thin, the CV is short. Grounding removes invention; it does not replace work you did not do.
• It is not a guarantee of interviews. No CV tool can promise one; the decision belongs to a recruiter and a hiring manager.
• It checks consistency, not truth. If you type the wrong year, a grounded builder cites the wrong year faithfully. You own the facts.
• It does not help with AI detection, because applicant tracking systems do not detect AI text.
• It is slower than a chat reply: asking and verifying takes minutes, not seconds.

The words: grounded, zero-hallucination, source of truth, cited

• Grounded AI CV builder: a builder whose output is tied to a source you supplied. The machine-learning term.
• Zero-hallucination CV builder: the same thing named by its result; no line without a source. A design property, not a claim that the tool is never wrong, because it repeats a wrong fact you gave it.
• AI CV builder that does not make things up: the plain-English name most people search for. Same category.
• Cited CV: the document a grounded builder produces, every line carrying a reference to its source. Defined on its own page.
• CV and résumé: in the UK, the EU and Turkey the document is a CV; in the US it is a résumé. Hallucination does not care which.

Questions people ask

Sources

1. Huang et al., A survey on hallucination in large language models (arXiv 2311.05232, 2023)arxiv.org/abs/2311.05232
2. Harvard Office of Career Services: create a resume, CV or cover lettercareerservices.fas.harvard.edu/resources/create-a-resume-cv-or-cover-letter/
3. MIT Career Advising and Professional Development: how to write a resumecapd.mit.edu/resources/how-to-write-a-resume/
4. Prospects (UK): how to write a CVwww.prospects.ac.uk/careers-advice/cvs-and-cover-letters/how-to-write-a-cv
5. RoleWorth: best AI resume builder 2026, an independent comparison on proof groundingroleworth.com/guides/best-ai-resume-builder-2026
6. RecastCV: why AI resume builders hallucinate, with a five-prompt testrecastcv.com/blog/grounded-ai-resume-no-hallucination