AI Writes the Application. AI Reads It. Where Are You?

The rejection arrives after two layers of software

The answer is not that AI assistance is automatically dishonest. Using a tool to clarify your own experience, organise evidence, translate a draft or support accessibility can remain compatible with honest representation. The boundary shifts when a tool invents experience, supplies an answer during an assessment that prohibits outside assistance, or produces work the candidate cannot explain. On the employer’s side, automated screening does not prove that no human is involved. It does mean that the information exchange needs deliberate checks.

The practical boundary is preparation versus performance

Policies differ by employer, assessment and jurisdiction, so the instruction attached to the specific stage of recruitment matters more than a universal rule. Microsoft’s candidate guidance permits responsible AI use during preparation when it reflects a candidate’s true capabilities. During assessments and interviews, it says candidates should demonstrate their own skills without outside assistance unless that help is explicitly permitted.

A useful working question is: did the tool improve the presentation of evidence, or replace the evidence itself? Correcting grammar in a personal draft is different from generating a story about a project you never led. Turning notes about your own work into a clearer structure is different from asking a model to solve a timed technical problem. That is practical guidance, not a universal legal test; the employer’s stated policy still controls the particular process.

If an assessment bans outside assistance, ask what that includes before starting. Does the rule cover translation, spelling tools, accessibility software, notes or only generative answers? A candidate can say: “I want to follow the process correctly. Which forms of support are permitted?” If an accommodation may be needed, raise it before the assessment where possible. The available route depends on the employer, the role, the jurisdiction and the facts. Permission for translation, notes and spelling tools must be confirmed; the examples above do not establish a general entitlement to use them.

The numbers are real, but narrower than the headlines

A study first published on March 9, 2026 examined reported assistance during one operational pre-hire assessment. Study 1 included 5,675 applicants in Q3 2024. In its no-warning control condition, Table 2 reports 3% for GenAI and 19% for the combined “algorithm sources” category. The latter adds search-engine and GenAI resources; it does not mean 19% of applicants used both together. This is assessment-specific evidence, not a measure of AI-written CVs.

Study 2 included 3,356 applicants in Q3 2025. For the 3,021 respondents in Table 5’s usage analysis, reported GenAI use was 23.2% in the control condition and 10.6% in the warning condition. These were successively collected groups, not a simultaneous randomized comparison. The figures describe self-reports, not independently observed use.

The findings concern a particular assessment, not CVs, cover letters or preparation across the hiring market. Authors affiliated with Talogy, the assessment provider, contributed to the paper. Self-report and non-concurrent groups limit interpretation; the percentages should not become a universal estimate of applicant AI use.

What employers are actually automating

The employer-side process is not one machine performing one decisive act. Microsoft’s own Careers disclosures describe resume parsing, inferred skills, job recommendations, recruiter-facing match information and attributes inferred from a profile. The company says some recruiter highlights are additional information rather than inputs into match recommendations. These are descriptions of Microsoft’s systems, not evidence of how widely equivalent tools are used or how accurate they are elsewhere.

HireVue’s September 2024 explainability statement describes specified AI-scored interview models as evaluating the words used in a response. It also describes a configuration in which a candidate who does not consent receives the same questions without AI assisting the review. That is a provider’s description of particular assessment configurations. It does not establish a universal alternative, meaningful human review in every deployment or a general right to opt out.

The information problem is symmetrical. Candidate-side tools can make materials faster and more uniform. Employer-side tools can parse, infer, rank or score before a conversation. That does not prove that every CV is machine-read or every rejection is automated. It does create a risk that polished signals become easier to produce while their connection to underlying capability becomes harder to inspect.

Detection is a poor substitute for evidence

An employer looking for certainty may reach for an AI-writing detector. The research does not justify treating a detector score as definitive proof of authorship. A study published in December 2023 by the Association for Computational Linguistics showed that specially designed prompting could generate text capable of evading high-performing detectors across several language models and writing tasks. That is experimental evidence about detector vulnerability, not proof that every tool fails in every context.

A separate study published online in July 2023 in Patterns found that tested detectors misclassified a substantial share of writing by non-native English authors in an academic-writing dataset. It is evidence about those detectors and that dataset, not a quantified employment-discrimination rate or a finding about every current system. It is enough to make one point plainly: unusual fluency, unusual simplicity or a detector score should not become a shortcut for calling someone dishonest.

A practical way to verify capability is to ask for a short explanation of the reasoning behind submitted work, a walkthrough of a relevant decision, a clearly scoped work sample or a conversation about trade-offs. This is an accountability approach, not an evidence-proven guarantee that follow-up is always better. It should also be proportionate: asking someone to reproduce confidential work or complete an unpaid project of unreasonable size simply moves the burden elsewhere.

What the law says—and what it does not say

Disclosure duties differ across the jurisdictions discussed here. The following US, New York City and EU examples are not a survey of every country’s law. In the United States, the Equal Employment Opportunity Commission says existing federal employment-discrimination laws apply when AI is used in recruiting, screening or hiring, including resume screening, recorded-video evaluation and recruitment chatbots. The agency also says existing law may require reasonable accommodation in some circumstances. Coverage, duties and remedies depend on the employer, the protected issue, the jurisdiction and the facts.

New York City’s Local Law 144 is more specific for covered automated employment decision tools. The city says an employer or employment agency generally may not use a covered tool unless it has undergone a bias audit within one year, audit information is publicly available and required notices have been provided. The scope depends on whether the particular tool qualifies under the law and its rules; it is not a universal model for every employer.

In the European Union, Annex III of the AI Act lists certain recruitment and selection systems—including systems used to place targeted job advertisements, analyse or filter applications and evaluate candidates—as high-risk, subject to scope and exceptions. The European Commission’s current timeline says the high-risk rules for employment and other specified Annex III uses apply from December 2, 2027 after the 2026 simplification changes. That application date should not be confused with the classification itself or with every other AI Act obligation, some of which already apply.

A candidate may ask: Is AI used to parse, rank, recommend or score applications? At what stage does a human review the material? What assistance is permitted in each assessment, and what support or alternative process is available? These are practical questions, not universal rights to receive a complete answer or a non-AI pathway. They make the process legible before a misunderstanding becomes a rejection.

Return accountability to the exchange

The central danger is not simply that machines are present. It is that both sides can produce convincing signals without enough underlying evidence. Applicants can make every application sound polished; employers can make every profile look measurable. The result may be efficient while becoming less informative. That is an editorial assessment of the documented capabilities and bounded studies, not a measured claim about the whole labour market.

For candidates, the durable test is whether you can explain the reasoning, evidence and work behind what you submitted without exposing confidential material. For employers, the corresponding test is whether you can explain what the system evaluates, what humans still decide, what support exists and how capability is checked beyond fluent prose. Neither test removes uncertainty. Both make the uncertainty visible—and leave responsibility with people rather than with a score that nobody can properly interpret.