Skip to content

BlogEngineering

ChatGPT for recruiters: what it does well, and where it stops

ChatGPT drafts a job ad, a Boolean string or an interview guide in a minute. It sees neither your ATS nor your clients, and a pasted CV leaves your records.

10 min read

ChatGPT is a very good tool for recruiters on anything that can be checked by reading it and contains no one: a job ad, a Boolean string, an interview guide, tidying up a brief that is already anonymous. It stops where the actual job begins, because it cannot see your ATS, does not know your clients, and everything it produces has to be copied across by hand before it exists anywhere.

The figure that places the usage comes from iHire’s 2025 report on online recruiting in the United States. 25.9 % of employers say they use AI to recruit, against 14.7 % a year earlier, and among them 73 % use it first of all to write job ads, ahead of messages to candidates (68.6 %) and far ahead of reading CVs (32.1 %). The ranking is no accident: writing is the use where the result can be judged at a glance.

What is ChatGPT actually good for in a recruitment firm?

Producing a first draft on material you already master, in the situations where that draft is expensive to write and cheap to review. It is the test we apply to any delegation, start with what can be checked rather than with what takes the most time, and a general assistant passes it comfortably on the four uses that follow.

One point of method applies to all four. By default the model produces the average answer of its training data, so the average job ad, the average interview question, the average search string. What makes it useful is what you give it and it cannot guess: the context of the role, the client’s constraint, the house tone.

Writing a job ad that does not look like the other three hundred

The model writes quickly and cleanly, and its natural flaw is blandness. Left unconstrained it lines up the “stimulating environments” and “passionate teams”, and a candidate who has read fifty of them this week will not read the fifty-first.

You are a recruiter at an 80-consultant IT services firm in Lyon. Write a
job ad for a senior Java developer on a long assignment with a banking
client. Constraints: 250 words maximum, no list of soft skills, remote
work (two days a week) appears in the first three lines, the salary range
is €55k to €62k and is stated plainly. Suggest two very different opening
lines and say which one you would pick.

The instruction that changes the result most is the last one. Asking for two versions forces you to choose, and choosing forces you to read, which is exactly what the speed of the tool tends to make you forget.

Building a Boolean string

This is probably the most underrated use in the trade. A search string for a rare profile means thinking of English and French job titles, neighbouring technologies, common misspellings and exclusions, and the model beats a hurried human at it because it does not forget the variant nobody thinks of at six on a Friday evening.

Build a Boolean string to find DevOps engineers specialised in Kubernetes
in France, with at least five years of experience. Include job titles in
English and French and neighbouring tools (OpenShift, Rancher, EKS, GKE).
Exclude internships, apprenticeships and sales roles. Give a short version
for LinkedIn, with no wildcards, and a long version for a CV database,
then list what the string will miss.

The string is checked by running it: if the first twenty results are off target, you see it in thirty seconds. A task whose errors show up that fast is the ideal task to hand to a tool that is sometimes confidently wrong.

Preparing an interview guide

The model is good at turning a job description into behavioural questions, provided you also ask it what a good answer looks like. A list of questions without criteria produces an interview judged on impression, whereas a guide that says what is expected lets two recruiters score the same thing.

From this job description (pasted below, with no client name), suggest
six structured interview questions for an infrastructure project manager.
For each one: what it is meant to check, what a strong answer contains,
and one warning sign.

Tidying up a brief or some notes

This is the most tempting use and the first one where you should stop and think. Rewriting a brief into a structured requirement, turning interview notes into a readable summary, all of it saves real time, and all of it almost always contains a client name, a candidate name or a salary. The useful case and the risky case look so alike they blur together, which brings us to the limits.

Where does a general assistant stop?

At the chat window, and that boundary will not disappear with a better model because it has nothing to do with its intelligence. It comes from what it can reach, what it knows about you and what it does not know without saying so.

It cannot see your ATS

Ask it which candidates in your talent pool have already worked for a given client, and it cannot answer, because it has never seen your talent pool. Everything it knows was pasted into the conversation, and everything it produces has to leave by copy and paste: the ad published by hand, the candidate’s status updated by hand, the follow-up scheduled by hand. If nobody copies it across, nothing happened, and the colleague who picks up the file the next day knows nothing of what was said.

That is not a design flaw, it is the definition of an assistant. It produces text you then use, and it shows none of the behaviours that make an agent, starting with the first one: writing into a system. Neither category is a disgrace, and an honest assistant is worth more than a mediocre agent. You only need to know which one you have in front of you.

Paid plans offer connectors to some office tools, so the boundary does move a little. It rarely moves as far as a recruitment firm’s ATS, and when it does, opening write access to business software is a company decision that depends on what your ATS lets a third party do, not a setting a recruiter switches on alone.

It does not know your company

It does not know that this client turned down two profiles in the spring because they came from a competitor, that the firm never goes below a certain margin, that a particular manager hates being called before ten. Even with its personal memory switched on, what it retains belongs to one person’s account, not to the firm.

The consequence shows at the scale of a team of eight recruiters. Each has their own saved instructions, their own version of the client context, their own phrasing, and those eight versions age at their own pace without anyone comparing them. The day one of them leaves, what they had taught the tool leaves with their account, which is the exact opposite of what you expect from a work tool.

It does not say what it does not know

Paste in three CVs and ask which one to keep: it will answer, with an argued ranking and perfect confidence, from three documents it cannot verify a single line of. It will not point out that the end date of an assignment is missing or that a job title is ambiguous, unless you explicitly ask.

A pasted document can also contain something other than what you see. A CV can carry instructions aimed at the model that will read it, in white text on a white background, and an assistant asked to rank is exactly the target of that kind of manipulation. If you have it compare profiles, ask what is missing and what could not be verified, never for a score.

Can you paste a CV into ChatGPT?

Not into a personal consumer account, and the reason is less technical than people think. OpenAI states that content from its consumer services may be used to train its models until the model improvement setting is switched off in data controls, whereas its business offerings are not used for training by default. That setting matters, but it settles only half of the problem.

The other half sits with you. A CV pasted into a tool the company did not choose is a processing of personal data missing from your record of processing, handed to a provider the firm has signed no contract with, about a candidate who was never told. The copy also escapes the retention periods running in your ATS, and whatever leaves the system that holds the data is no longer erased on the right date.

The CNIL, the French data protection authority, published its first clarifications on deploying generative AI in July 2024, and they read like a list of what an individual use lacks. It recommends defining authorised and prohibited uses, citing as an example not providing personal data to the system, checking whether the provider may reuse the data supplied, favouring secure systems, and training users. None of those decisions can be taken by a recruiter alone in front of a browser.

Removing the name is not enough, and it is the most widespread mistake. A career history naming an employer, assignment dates and a school often identifies a person, especially on a rare profile in a mid-sized city, and data with the name removed is still personal data as long as it can be traced back. For an IT services firm there is client confidentiality on top: a consultant profile naming the end client and its architecture leaves a perimeter you guaranteed in writing.

Why banning ChatGPT for recruiters solves nothing

Because the usage already exists, and a ban makes it invisible without reducing it. 66 % of the office professionals PagerDuty surveyed in April 2026 had used an AI at work they believed was banned, and over a third had entered client data into it. A recruiter with forty CVs to read before a client meeting at two will make the same calculation, and will make it on their phone if the work computer blocks access.

What we wrote about shadow AI in the enterprise applies here unchanged: the approved tool comes before the policy, never the other way round. For a recruitment firm, the list of what never goes into an assistant nobody chose fits in three lines, candidate names, client names and pay, and a list that short gets remembered where a twelve-page document gets filed.

The usage rule that follows is simple to state, which is why it gets applied. Anything that contains neither a candidate nor a client can go into a consumer assistant; everything else goes through a tool the company has chosen, contracted and entered in its record of processing. The job ad, the Boolean string and the interview guide are on the first side, the interview summary and the profile comparison on the second.

When do you need something other than an assistant?

When the task requires writing into a system, remembering a client from one time to the next, or starting without being asked. Chasing a candidate who has not replied in seven days, preparing a consultant’s file by fetching what is missing from the ATS, flagging that an assignment is about to end: none of these happens in a chat window, because none of them begins with a human message.

That is the territory of what is called agentic AI, and Cobalt has published a good definition of it applied to recruitment, with the difference between a system that produces content and one that plans, uses tools and corrects its own work. It is in that second category that we built Balt, and it comes with its own requirements: an agent writing to a candidate goes through human approval before sending, exactly like the text ChatGPT drafted for you and that you read before sending it.

None of this makes ChatGPT less useful for what it does well. The real question is not whether to use it, but where the line runs between what it prepares and what the firm has to carry itself. Once that line is drawn, the next question is what your ATS has to allow for a tool to work on the other side of it.

Frequently asked questions

  • Can you paste a CV into ChatGPT?

    Not into a personal consumer account. A CV is personal data, the service is neither in your record of processing nor covered by a contract the company signed, and the candidate has not been told. A business plan chosen and contracted by the firm changes the answer, provided the use is written down and known.

  • Is removing the candidate’s name enough?

    No. A career history naming an employer, assignment dates and a school often leads straight back to the person, especially on a rare profile. Data with the name removed is still personal data as long as somebody can trace it back.

  • Can ChatGPT rank candidates?

    It will if you ask, confidently, and that is exactly the problem. It only knows what was pasted in, it does not say what it could not verify, and a ranking produced that way ends up being applied without anyone being able to challenge it.

  • Should recruiters be banned from using ChatGPT?

    A ban moves the usage to the personal phone without reducing the volume. What works is providing an approved tool, writing down in three lines what never goes in (candidate names, client names, pay), and making exception requests quick.

Sources

  1. HCAMag, iHire 2025 State of Online Recruiting: where recruiters use AIhcamag.com
  2. OpenAI, How your data is used to improve model performancehelp.openai.com
  3. CNIL, How to deploy generative AI: first clarifications (July 2024, in French)cnil.fr
  4. PagerDuty, Shadow AI survey 2026 (Wakefield Research, 1,250 office professionals, April 2026)pagerduty.com
  5. Cobalt, Agentic AI: definition and concrete applications in recruitment (in French)cobalt-ia.com

Balt does this work

Read next

Give Balt its first job.

Thirty seconds to add it to Teams and get to work.

Try for free