Should I use AI in my job application?

In February 2025, Anthropic added a line to its job ads asking applicants not to use AI assistants when applying. The company that builds Claude wanted to see how candidates wrote and thought without software in the way.

Five months later it dropped the rule. Its current guidance tells applicants to go ahead and use Claude on their CV and cover letter, on the grounds that the company hires people who work well with AI and would like to watch them do it. It has said the policy will be reviewed again as the technology moves.

That reversal is a useful marker for where hiring has landed. The question is no longer whether a candidate used AI. Most employers assume they did, and in New Zealand they are mostly right.

Research carried out by Nature for SEEK, which surveys 4,000 New Zealanders a year, found 35% had used AI to draft a cover letter, 25% to write a CV and 17% to prepare for an interview. What employers are working out now is which parts of the process they are willing to see it used on.

They are working it out under pressure. Greenhouse, the recruitment platform, averages more than 250 applicants per advertised role, and chief executive Daniel Chait has described the result as a doom loop, with jobseekers paying around US$20 a month for tools that fire off applications at scale while recruiters reach for AI to sort the pile.

In a 2026 ZipRecruiter survey of more than a thousand talent acquisition staff, 48% said AI had increased the number of applications they receive per role. Whatever a hiring manager thinks about AI in principle, they are reading more applications than they used to and giving each one less time.

Two-thirds can tell

A 2026 survey by CV-writing service TopResume, covering more than 800 hiring managers, found 67% could pick an AI-written cover letter and 54% thought less of the candidate for it. TopResume sells the service AI competes with, so read the numbers with that in mind, but the direction matches what recruiters say publicly.

Read those two figures together and the story is less about detection than about what gets detected. The letters that get flagged are the ones with nothing in them. Smooth sentences, broad enthusiasm, no mention of the company beyond its name. A hiring manager working through 250 applications is not scanning for machine syntax. They are scanning for whether this person knows what the job is, and a hollow letter fails that test regardless of who or what wrote it.

It is worth being clear about what “can tell” means, because it is not forensic. Recruiters are inferring from the writing, not running it through a scanner. Where they do use detection software, the results are shakier than the marketing suggests. Stanford researchers testing detectors in 2023 found they misclassified writing by non-native English speakers at far higher rates than writing by native speakers, in some tests flagging most of it as machine-written.

In a country where a large share of applicants write English as a second or third language, that is a live fairness problem, and it argues against treating a detector score as evidence of much at all. Most experienced recruiters use a flag as a reason to read more closely.

For a candidate, the thing to fix is not the polish. It is whether anything in the letter could only have been written by you.

The line has moved to the live parts

Where employers are still firm is anything happening in real time. Amazon’s recruiter guidance tells candidates not to use generative AI tools during interviews unless they have been told otherwise, and warns that doing so can get them disqualified.

Google has banned AI tools during virtual interviews outright. Anthropic kept the same restriction when it relaxed everything else, so applicants there can use Claude on their written materials, to research the company and to rehearse answers, but not during assessments or live interviews.

Part of the reason is that the cheating got good. Overlay tools now feed answers to candidates during video and coding rounds without showing up on a shared screen, and recruiters have dealt with applicants who turn out not to exist at all. Cisco’s Scott McGuckin has said remote work and AI have made it easier than ever for fake candidates to get into the hiring process.

Employers are responding by putting people back in rooms. A Gartner survey found 72.4% of recruiting leaders are now running interviews in person to counter fraud, and Google,

Cisco and McKinsey have all moved that way. A McKinsey spokesperson put it in terms of needing face-to-face contact to judge the qualities that cannot be automated.

The logic underneath is not complicated. Written materials show what someone can produce with time and tools to hand, which is roughly the condition they will work under once hired. An interview shows what they know and how they think on the spot. AI in the first case is now unremarkable. In the second it answers a different question from the one being asked.

None of which means the interview stays a human affair. Employment Hero ran more than 2,500 AI-led interviews for New Zealand employers in April alone, using them to screen and rank first-round candidates.

Chief executive Neil Webster has put the reasoning bluntly, saying that when 200 people apply for a role you need to comb it down, and that the company never wants the AI to make a decision, only a recommendation. He expects most interviews to involve AI within five years. A candidate barred from using AI in an interview may well be answering questions posed by one.

What candidates get wrong

Trouble starts when applicants hand the model the parts only they can supply. Why this job. What they did in the last one and what came of it. The thing they read about the company that made them apply. A model will produce a paragraph on any of these, and it will read fine and say nothing, because the information was never there to begin with.

The mass-application tools make this worse, because they strip out the step that actually gets someone hired. A candidate who sends 200 near-identical applications has put in less work than one who sends fifteen considered ones, will usually do worse, and tends to conclude from the silence that the market is broken. Chait’s doom loop is the aggregate version of this. The individual version is a jobseeker who cannot understand why a hundred applications produced no interviews.

The second problem surfaces later, in the room. Recruiters report candidates who cannot talk about their own applications. If a CV says you led a systems migration, expect twenty minutes on it. Anything AI helped write has to be something you can defend out loud without notes, and the more confident the written claim, the more detail an interviewer will expect behind it.

There is also the matter of the rules themselves. Employer policies now run from outright prohibition to open invitation, and some ask candidates to declare where they used AI. Ignoring a stated instruction is a straightforward mark against someone, and unlike writing style it leaves no room for interpretation.

A working order

Draft it yourself first, badly if necessary, then use AI to fix the sentences. The order matters, because it puts the content in before the polish. Going the other way leaves you with polish and nothing underneath.

Before any of that, the tool earns its keep on the job ad itself. Paste the ad in and ask which requirements are load-bearing and which are wishlist. Long ads bury the two or three things an employer will really screen on, and candidates routinely answer the wrong ones. Once you have a draft, ask it to check that draft against the ad and list what you have not addressed. Comparison is a task models are reliable at, and it is a different job from asking one to supply your motivation.

On the CV the risk is lower. It is a structured document and largely a matching exercise, so using AI to reorder sections, tighten bullet points to fit a page or bring your wording closer to the terms used in the ad is unremarkable. Two rules hold. It must not invent anything, and that includes numbers. A model will happily supply a plausible-looking figure when a bullet point seems to want one, and a fabricated metric is the sort of thing that ends an interview. Nor should you paste anything confidential from a current or former employer into a chatbot to get help describing it.

Read the result aloud before sending. Anything you would not say to a person, cut. Keep your reason for applying and any direct message to a recruiter in your own words, imperfect grammar and all. A short email that mentions one specific thing about the company does more work than a flawless one that could have gone to anybody.

Keep the version you sent, too. If you are called in three weeks later you want to reread what you claimed rather than reconstruct it from memory, which gets harder the more applications you have out. Everything on that page is fair game once the conversation starts.

Employers are writing these rules while the ground moves under them, which is why the rules keep changing. For now the safe assumption is that AI on the paperwork is expected and AI in the room is not.