Evidence · strength 1 of 5 · searched to 2 Nov 2026
Do AI-written job posts and applications lead to more hires?
Not enough evidence yet to say whether AI-written job posts and applications lead to more hires. 0 of 2 studies found a clear gain in callbacks or hires.
What field experiments on hiring platforms found when employers drafted job posts and applicants wrote cover letters with AI: time saved, applications, callbacks and hires.
VerdictNot enough evidence yet
No good evidence
ImpactNo clear effect
0 of 2 studies found a clear gain in callbacks or hires.
Evidence strength1 of 5
We have very low confidence in this estimate.
Cost per user a year$
Built into hiring platforms or done with general AI assistants, often free.
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Output (headline)
No clear effect- Wiles and Horton (2026): +19% Share of would-be employers who published their first job post (+0.056, SE 0.002 on a control mean of 0.297).
Treated employers were 19% (or 6 percentage points) more likely to post a job than employers in the control group.
- Wiles and Horton (2026): +0.1 points Probability the employer made any hire within 14 days, unconditional on posting (matches formed) (+0.001, SE 0.001 on a control mean of 0.066). Not statistically clear.
GenAI Treatment Assigned ... 0.056 1.161 0.003 0.001 (0.002) (0.061) (0.001) (0.001)
- Cui et al. (2025): +0.4 points Probability a bid receives a callback (employer chat), workers with access to the AI Bid Writer through their membership plan vs without, after vs before launch (ITT, controlling for GPT-4 release x access) (0.0043, SE 0.0020, significant at 10% only). Not statistically clear.
access to the AI Bid Writer increases the probability of receiving a callback by 0.43 percentage points (ITT). Relative to the 7.02% baseline callback rate
- Cui et al. (2025): +3.6 points Callback probability for bids actually written with the tool (LATE, access instrumenting use) (0.0356, SE 0.0170, significant at 10% only). Not statistically clear.
When we instrument for usage (LATE), the implied effect rises to 3.56 percentage points (51% of the baseline rate)
Errors
No clear effect- Cowgill et al. (2026): +7.7 log points of squared screening error Evaluators' squared error in judging whether the sender had real expertise, pitches/cover letters written with ChatGPT vs without (log, all fixed effects; hiring and investing pooled) (0.077, SE 0.041 (t about 1.88, p about 0.06): significant at 10% only; the table's single star conflicts with its own SE). Not statistically clear.
On average, candidates' access to GAI causes information losses equivalent to a 4-9% increase in screening errors
Quality
Mixed- Wiles and Horton (2026): −3.4 points Probability the employer made a hire within 14 days, conditional on posting a job (SE 0.003 on a control mean of 0.224 (about -15%)).
treated employers who posted a job were 15% less likely to make a hire.
- Cui et al. (2025): +0.16 SD Cover-letter tailoring (TF-IDF similarity to the job post), access ITT (0.0220, SE 0.0032, p < 0.01).
the coefficient on the interaction term is 0.0220, which is 16.42% of the pre-rollout textual similarity standard deviation of 0.1340.
Time
Large gain- Wiles and Horton (2026): −44% Minutes employers spent writing the job post, offered AI-written first draft vs standard flow (conditional on posting) (-3.592 minutes, SE 0.072 on a control mean of 8.142).
employers in the control group spent on average 8.1 minutes writing the job post while treated employers spent only 4.5 minutes.
- Checking time not measured: none of these studies counts the time spent checking the AI’s output separately.
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level2 of 5, from 3 studies graded A, B or C.
- Down one: old models0 of 3 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Flag (no deduction)Mostly working papers or preprints, not yet peer-reviewed (flagged, not deducted).
- Publication biasnot testable with fewer than 10 studies.
- Result1 of 5. We have very low confidence in this estimate.
Behind the average
- Employers less likely to edit the AI draft (Wiles and Horton (2026)): Hiring fell 4.5 pp among employers with low predicted active use against 1.9 pp among high predicted use.
- Jobseekers (Wiles and Horton (2026)): Treated posts drew about 5% more applications per post; the authors estimate jobseeker time costs about six times the employer time saved.
- Employers (signal value) (Cui et al. (2025)): After launch, the link between cover-letter tailoring and callbacks fell by 51% and with offers by 79%; employers leaned more on review scores.
- Strong pre-AI writers (Cui et al. (2025)): The best writers gained half as much tailoring as the weakest, narrowing gaps between workers.
- Senders from non-English-speaking countries (Cowgill et al. (2026)): Access to ChatGPT improved screening accuracy for these senders, the opposite of the average effect.
- Hiring setting only (Appendix Table D2) (Cowgill et al. (2026)): squared screening error -0.0087 (SE 0.010), variance of beliefs +0.003 (SE 0.0028), neither significant
How, where and why it works
Why it works
2 of 4. AI makes posts and letters quicker and more tailored, but tailored letters stop telling employers who made the effort.
For whom and where
2 of 4. Weaker writers gained most in tailoring; employers who did not edit the AI draft hired less.
What it takes
1 of 4. Platform tools offered a one-click draft; studies say little about editing advice.
The full cost
1 of 4. Writing time fell sharply in one trial; costs are not counted.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Use AI for a first draft, then make it specific to you or your job.
If you are rolling it out
- Track hires and matches, not just posts or applications.
What it takes
- Edit the AI draft: in one trial, employers unlikely to edit it hired less.
What can go wrong
- More AI-drafted job posts did not mean more hires in a large platform trial.
- When everyone's letters are polished, employers find it harder to tell applicants apart.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| More, but Worse: The Marketplace Effects of AI-Generated Job Posts (earlier title: Generative AI and Labor Market Matching Efficiency) Emma Wiles, John J. Horton, 2026. Author PDF dated 4 September 2026 (working paper) | B | field RCT, deployed work | 180,324 new employers on an online labour market (90,048 treated, 90,276 control); 58,495 posted a job | 'a popular generative AI service' (not named) | none |
| Signaling in the Age of AI: Evidence from Cover Letters Jingyi Cui, Gabriel Dias, Justin Ye, 2025. arXiv 2509.25054 (preprint) | C | other quasi-experiment, deployed work | About 5 million cover letters to over 100,000 PHP and Internet Marketing jobs on Freelancer.com; 2,511,592 bids in the main regressions | Freelancer.com AI Bid Writer (underlying model not stated) (Apr 2023) | none |
| Does AI Cheapen Talk? Theory and Evidence From Global Entrepreneurship and Hiring Bo Cowgill, Pablo Hernandez-Lagos, Nataliya Langburd Wright, 2026. IZA Discussion Paper 18442, March 2026 (working paper) | C | RCT, lab or benchmark | 343 Prolific senders (job candidates and entrepreneurs) and 801 Prolific receivers with hiring or investing experience; 6,440 evaluations, 3,192 in the hiring setting | ChatGPT (version not stated) | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Does the rating hold up?
- Run without company-run studiesNo change: no such studies.
- Run without working papers and preprints0 studies: strength 0, verdict no good evidence.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. Studies cover models released Apr 2023 to Apr 2023. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.