Can I use AI to write a job application?

Yes, provided the experience in the application is yours and the help goes to the sentences. A writing tool that tidied what jobseekers had written raised their chances of being hired in a randomised experiment. A tool that writes the whole pitch for every applicant has since made the pitch worth less to the employer reading it, which changes what the application is for. Nobody has yet measured what happens to a letter written wholesale by a model, so the line between help and substitution rests on reasoning, and this page marks where.

A tool that fixed the sentences raised the hiring rate#

Wiles, Munyikwa and Horton randomly allocated 480,948 new jobseekers on a large online labour marketplace. Half had the description in their profile checked by a writing-assistance service that flagged spelling and grammar and gave advice on clarity, tone and style. The analysis sample is the 194,700 whose profiles were approved and not empty. Those offered the help were 8 per cent more likely to be hired in their first month (95 per cent confidence interval 3 to 13 per cent), from a base rate of about 3 per cent. Employer ratings of the work afterwards did not fall.

The authors read the result as clarity rather than deception: better writing lets an employer see the same facts more accurately. The working paper named Emma van Inwegen as first author when it circulated in January 2023. The Management Science version (volume 71, issue 12, December 2025) names Emma Wiles, and the title is unchanged.

When every pitch is polished, polish stops being evidence#

Galdin and Silbert looked at what happens when the writing is done by the model. Their data are about 2.7 million applications from 212,000 workers to about 61,000 fixed-price coding jobs on Freelancer.com between January 2021 and July 2024. They used a language model to score how closely each application was tailored to its job posting, then asked how much an employer paid for tailoring.

Before ChatGPT was released on 30 November 2022, a one standard deviation increase in tailoring was worth a $25.67 lower bid to the employer. In the later sample the same increase was worth $14.85. The abstract says employers paid for customised applications before language models "but not after", and the introduction says the pattern weakens significantly or disappears completely, depending on the specification.

The paper then models a market in which written applications carry no signal at all. In that counterfactual the top fifth of workers by ability are hired 19 per cent less often and the bottom fifth 14 per cent more often. That is a model output, built on the authors' assumptions, and it describes a possible endpoint rather than something observed.

The penalty falls on people who are seen using it#

Reif, Larrick and Soll ran four preregistered experiments with more than 4,400 participants. People described as using AI for a task were rated lazier, less competent and less diligent than people described as getting the same help from elsewhere. The penalty was largest among evaluators who rarely used AI themselves, and hiring managers who did not use it preferred candidates who reported no use. It disappeared when the tool's usefulness for the specific task was made explicit. The targets were described in vignettes rather than observed, so this is evidence about perception and says nothing about anyone's actual work. The authors expect the penalty to shift as use becomes ordinary. It bears on whether to say you used AI.

One employer's written rule#

Employers differ. Anthropic, on a page last updated on 10 July 2025, asks candidates to write the first draft of a resume or application answer themselves and then use Claude to refine it. The experiences must be the candidate's own, and using the tool to create experiences a candidate has not had is not allowed. Take-home assessments and live interviews are done without AI unless the company says otherwise, while using it to research the company and practise answers is encouraged. That is one firm's position. It is useful because it separates three things the studies mix together: help with expression, invented content, and permission to use a tool in an assessment.

Why a line between sentences and experience is a reasonable reading#

The tool in the first study corrected prose and could not invent a career. The pitch in the second study was cheap to write because a model supplied the content, and that cheapness is what weakened it as a signal. The one employer rule quoted here forbids invented experience and welcomes refinement. All three are consistent with a split between help with expression and help with substance. None of them tested the split directly. It is an inference from three sources that differ in setting, population and date, and a reader should treat it as one.

It also fits the older test used elsewhere in this research: could you describe, out loud, exactly what you did and what the tool did? The out-loud test was written for students and transfers to applications without change. For the employer's side of the same problem, see hiring signals when effort is free.

What none of this measures#

• No study of fully generated letters in conventional hiring. Both experiments and the Freelancer data come from online labour platforms for freelance work. A graduate scheme or a salaried role may behave differently.
• The Wiles tool is not a chatbot. It checked and advised on text the jobseeker had written. The result says little about handing a model a blank page.
• The Galdin and Silbert paper is a preprint. It was submitted to arXiv on 11 November 2025 and lists no journal. It covers English-language, fixed-price coding jobs between $30 and $250, and only about 39 per cent of the later applications were scored.
• The penalty studies describe people. They do not show what a recruiter does with a real application.
• Nobody here has published how often recruiters can tell. The question of detection is separate, and the evidence there is against reliability in both directions.

Write the claims yourself and let the tool tidy the sentences#

• Start with your own list of what you did. Jobs, results, numbers, the one thing you would raise in an interview. Doing this first means there is something for the tool to be checked against.
• Ask for help with clarity, tone and cuts. That is the use with a measured upside.
• Read every claim back as a question. If an interviewer asked about it, could you talk for five minutes without the document? Remove anything you could not.
• Check the employer's instructions. Application forms, assessment briefs and interview invitations can say what is allowed. The rule that applies is the one on that page.
• Use the tool freely for preparation. Researching the company, rehearsing answers and drafting questions for the panel are the uses employers are most open about.

Key research and primary sources

• Wiles, E., Munyikwa, Z. and Horton, J. J. (2025). Algorithmic Writing Assistance on Jobseekers' Resumes Increases Hires. Management Science, 71(12), 10144 to 10164.
• Galdin, A. and Silbert, J. (2025). Making Talk Cheap: Generative AI and Labor Market Signaling. arXiv:2511.08785.
• Reif, J. A., Larrick, R. P. and Soll, J. B. (2025). Evidence of a social evaluation penalty for using AI. PNAS, 122(19).
• Anthropic (2025). How to collaborate with Claude during our hiring process. Candidate guidance, updated 10 July 2025.

Related SuperSkills research#

On the employer's side of the exchange, hiring signals when effort is free and whether AI should reject an application before a human reads it. On declaring use, telling your boss and being honest about using AI. On writing generally, human at the start and where the advantage goes when everyone has AI.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The Galdin and Silbert figures were read in the arXiv text and checked against its tables, and the journal reference for Wiles, Munyikwa and Horton was confirmed from Crossref. A reputation-score result in the Galdin paper was left out because its dollar figures could not be matched to a table label. A Dutch press report of a Tilburg study on AI cover letters was not used because no primary paper could be found. The line between help with sentences and help with experience is reasoning, and is labelled as reasoning.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.