Tavryqen Blog

Should Employers Use AI to Detect Whether Job Applicants Used AI?

By Shamar Bennett August 21, 2026 Responsible AI & Governance

There is something strange about where recruitment has ended up.

An applicant uses artificial intelligence to improve a cover letter. The employer then uses another artificial intelligence system to decide whether the applicant used artificial intelligence. The candidate may run the document through another tool to make it sound less machine-written. The employer buys a better detector. The applicant changes the wording again.

Somewhere in that arms race, everybody risks forgetting the question that actually matters: can this person do the job?

I understand why employers are concerned. If a company asks somebody to submit a written response, solve a problem, write code, prepare an analysis or explain why they are suitable for a position, the employer wants to know whether it is looking at the applicant’s real ability. That is reasonable.

What is much harder to justify is turning an AI-detection score into a verdict.

A screen saying a CV or written assessment is “91% AI-generated” looks scientific. It may come with charts, highlighted sentences and confident language. That still does not make the result proof. When the score can affect whether somebody gets access to employment, the consequences are too serious to ignore the limitations.

This is not only a technology question. It is a recruitment question, a fairness question, a privacy question and, depending on where the employer operates, a legal question.

It also exposes a contradiction. A company may encourage employees to use Copilot, ChatGPT, Gemini, Claude or another approved assistant after they are hired, while treating an applicant’s responsible use of similar technology as evidence of dishonesty. Employers need a better line than “AI is bad during recruitment but useful on Monday morning after you join us.”

First, what does “using AI” actually mean?

This sounds simple until somebody has to write the policy.

An applicant writes a cover letter and uses Grammarly to correct it. Is that AI use? Microsoft Word suggests a rewrite. Is that AI use? A candidate whose first language is not English uses translation software. Another candidate dictates an answer using speech-to-text. Someone asks an AI assistant to correct spelling without changing the substance. Someone else asks it to restructure a paragraph they already wrote. Another person asks it to generate ten interview questions for practice. A final applicant asks it to write the entire application.

Those are not the same behaviour.

That is why a rule saying simply “AI use is prohibited” is usually less clear than it first appears. The better question is what kind of assistance is allowed at a particular stage of recruitment and what skill the employer is trying to measure.

There is a meaningful difference between using technology to improve presentation and using technology to falsely represent competence. If an assessment explicitly says it must be completed without generative AI, secretly outsourcing the answer to a model is different from using an approved grammar checker on an ordinary cover letter.

The rule should be clear before the candidate starts, not invented after the employer dislikes the result.

Why employers are worried

The concern is not imaginary. Recruiters can receive large numbers of polished applications that sound remarkably similar. Generative AI can produce a competent-looking cover letter in seconds. It can answer screening questions, suggest technical responses and generate portfolio material.

A company hiring a communications specialist reasonably wants to know whether the person can communicate. A company hiring a developer should care whether the candidate understands submitted code. A business hiring a researcher should care whether the applicant can evaluate evidence rather than merely produce confident prose.

There is also a fairness issue. If an employer clearly tells candidates that a particular assessment must be completed independently, one applicant spends two hours doing the work and another has a model generate it in five minutes, the employer is entitled to investigate.

So I am not arguing that employers should ignore inappropriate AI use. I am arguing that detection should not be confused with proof.

The fundamental problem: an AI detector can be wrong

AI-authorship detection is a classification problem. A detector looks for patterns and estimates whether text resembles material produced by a model. It does not travel back in time and observe who typed the document.

False positives therefore matter. So do false negatives.

A widely cited 2023 study by researchers including Stanford scholars tested seven GPT detectors on human-written essays. In one part of the research, the detectors were applied to 91 TOEFL essays written by non-native English writers. The researchers reported an average false-positive rate of 61.3% across the seven detectors for that dataset. They also found substantial differences between how the detectors treated those essays and essays written by US eighth-grade students.

That result deserves attention, but it also deserves careful wording. It does not mean that every AI detector in 2026 falsely accuses 61.3% of job applicants. It does not establish the error rate of a particular modern commercial product. The study involved particular detectors, datasets and writing populations at a particular time.

What it does show is the point employers cannot afford to ignore: detector performance can vary across writing populations, and a confident score is not the same thing as certainty.

The same research also showed that generated text could be altered in ways that reduced detection. That creates an awkward incentive. A sophisticated applicant who intentionally uses AI may know how to edit the output enough to avoid a detector. An honest applicant whose natural writing happens to trigger the system may be the person who gets investigated.

That is a poor foundation for an integrity regime.

AI-assisted writing is becoming difficult to divide into neat percentages

Modern documents often have mixed authorship.

A person researches manually. AI organises the notes. The person writes the first draft. A grammar tool changes several sentences. The person rejects half of those changes. AI suggests a heading. A colleague edits the final document.

What percentage was written by AI?

The question quickly becomes less useful than it sounds. Employers should care more about accountability: who verified the facts, who understands the work, who takes responsibility for it, whether confidential information was protected, whether sources were checked and whether the applicant can perform the underlying task.

The UK: automated hiring is possible, but safeguards matter

The United Kingdom changed its automated-decision framework through the Data (Use and Access) Act 2025. Section 80 replaced UK GDPR Article 22 with Articles 22A to 22D.

The change is important because UK law now permits broader use of solely automated significant decisions involving ordinary personal data than the previous Article 22 framework did. But “broader” does not mean “uncontrolled.”

Article 22A defines a decision as based solely on automated processing where there is no meaningful human involvement. A significant decision is one producing a legal effect or a similarly significant effect on the person. Article 22C requires safeguards for significant decisions based solely on automated processing. Those safeguards include information about the decision, an opportunity to make representations or contest it, and the ability to require human intervention. Special-category personal data remains subject to tighter restrictions under Article 22B.

That matters directly to recruitment. In March 2026 the UK Information Commissioner’s Office published findings from engagement with more than 30 employers about automated recruitment. The ICO said employers should be transparent with jobseekers, proactively monitor for bias and explain how candidates can challenge a decision and request human review.

The ICO also made an important practical point: human involvement has to be meaningful. A recruiter who simply accepts whatever the system recommends is not necessarily providing the kind of independent judgment people imagine when they hear “a human reviewed it.”

Then there is the Equality Act 2010. Employers remain responsible for avoiding unlawful discrimination in recruitment. A tool that appears neutral can still create problems if its use disadvantages protected groups and the practice cannot be lawfully justified. Disability and reasonable-adjustment obligations also matter when employers design assessments or rely on automated tools.

For a UK employer, my recommendation is straightforward: do not allow an AI-authorship score to quietly become an automatic rejection rule. Document why the tool is being used, test it, tell candidates about materially automated decision-making where required, maintain real human review and provide a practical route to challenge consequential decisions.

The European Union: GDPR today, AI Act obligations alongside it

In the European Union, employers need to consider the GDPR, the EU AI Act, equality law and the national employment laws of the Member State involved.

GDPR Article 22 addresses decisions based solely on automated processing, including profiling, that produce legal or similarly significant effects, subject to its exceptions and safeguards. Articles 13, 14 and 15 also contain transparency and access requirements relevant to automated decision-making in applicable cases.

The EU Court of Justice’s SCHUFA judgment is useful here, although it was not an employment case. The Court considered automated credit scoring and held, in the circumstances before it, that the generation of a probability score could itself fall within Article 22 where a third party drew strongly on that score when making the decision.

That does not automatically decide how a court would treat an AI-writing detector in recruitment. But it raises exactly the right question: if a recruiter technically clicks the Reject button while a “94% AI-generated” score effectively determines the outcome, how meaningful was the human decision?

The EU AI Act, Regulation (EU) 2024/1689, adds another layer. Annex III includes certain AI systems intended to be used for recruitment or selection, including targeted job advertisements, analysing and filtering applications, and evaluating candidates. Those systems can fall within the Act’s high-risk framework.

But precision matters. It would be wrong to say that every piece of AI touching recruitment is automatically high-risk. Article 6(3) contains circumstances in which an Annex III system may not be considered high-risk where it does not pose a significant risk of harm and does not materially influence the outcome, including specified narrow or preparatory tasks. Systems that perform profiling of natural persons in Annex III contexts are treated more strictly.

So an AI detector used as a limited supporting signal may require a different analysis from a detector whose score materially determines whether candidates progress.

The practical lesson is not “AI hiring is illegal in Europe.” It is that the more consequential the tool becomes, the more carefully the employer should examine classification, data protection, transparency, human oversight, accuracy, discrimination and documentation.

The United States: the patchwork is the point

The United States does not have one national law called the AI Hiring Act. Employers instead face federal anti-discrimination law plus a growing collection of state and local requirements.

At federal level, Title VII of the Civil Rights Act of 1964 remains central. Employment practices can create unlawful discrimination even when technology performs the screening. The EEOC’s long-standing approach to selection procedures is especially relevant: where a selection procedure has an adverse impact, job relatedness, business necessity and validation become important questions under the applicable framework.

The Americans with Disabilities Act also matters. The EEOC has specifically warned that algorithmic hiring tools can screen out people with disabilities and that employers may need to provide reasonable accommodations. A technology vendor does not remove the employer’s responsibility.

Other federal protections, including age and genetic-information laws where applicable, continue to operate regardless of whether the employer calls the system “AI.”

New York City

New York City’s Local Law 144 regulates covered automated employment decision tools. A covered AEDT cannot be used unless it has been subject to a bias audit within the preceding year, specified audit information is publicly available, and required notices are provided.

The law defines an AEDT around computational processes that issue simplified outputs such as scores, classifications or recommendations and substantially assist or replace discretionary decision-making for employment decisions. City guidance also states that required notice must generally be given at least ten business days before use.

That does not mean every AI detector is automatically an AEDT. Whether a detector falls within the law depends on the statutory definition and, critically, how its output is used.

California

California’s employment regulations concerning automated-decision systems became effective on 1 October 2025. The California Civil Rights Council’s rules clarify how the Fair Employment and Housing Act applies when employers or their agents use automated systems in employment decisions.

The important point for employers is that using software does not create a discrimination-free zone. If an automated system contributes to unlawful discrimination, the fact that a vendor supplied the tool does not make the underlying employment consequences disappear. The regulations also contain record-retention requirements relevant to automated-decision data.

Illinois

Illinois has specifically legislated around employer use of artificial intelligence, including discrimination and notice. But this is an area where employers should verify the current statutory text before deployment because the provisions have been subject to legislative amendment.

The safer business lesson is not to build a national hiring policy around a snapshot of one state’s law. Maintain jurisdiction-specific review and update it as legislation changes.

Colorado

Colorado is a good example of why these articles need dates on them.

On 14 May 2026, Colorado’s SB 26-189, the Automated Decision-Making Technology Act, was signed into law. According to the Colorado Attorney General, it repeals and reenacts the earlier Anti-Discrimination in AI Act framework with new requirements concerning automated decision-making technology used in consequential decisions.

The new framework includes obligations involving developer documentation, deployer notices, post-adverse-outcome disclosures, access and correction rights for relevant personal data, and an opportunity for meaningful human review of certain adverse consequential decisions. The Attorney General is required to adopt specified implementing rules by 1 January 2027.

That is the current Colorado position employers should be preparing around, rather than relying on summaries of the earlier 2024 legislation.

The Caribbean: there is no single Caribbean AI law

This part needs care because “the Caribbean” is not one jurisdiction. Jamaica, Barbados, Guyana, Trinidad and Tobago, the Cayman Islands and other territories and states have separate legal systems. Some have GDPR-influenced data-protection rules; commencement status and enforcement structures also differ.

So no responsible article can honestly claim to list “all Caribbean AI hiring laws.” What we can do is examine important examples that show why automated recruitment deserves legal review across the region.

Jamaica

Jamaica’s Data Protection Act 2020 is unusually relevant to this exact discussion. Section 12 addresses rights in relation to automated decision-taking.

It applies to certain decisions that significantly affect a data subject. The Act gives an individual a mechanism to require that a covered decision not be based solely on automatic processing used to evaluate matters relating to that person. The statutory examples expressly include performance at work, creditworthiness, reliability and conduct.

The Act also provides procedures involving notification and reconsideration in specified circumstances. An employer in Jamaica considering automated screening should therefore examine section 12 directly rather than assuming AI recruitment is simply an unregulated technology issue.

Barbados

Section 18 of Barbados’ Data Protection Act 2019 addresses automated individual decision-making, including profiling. It provides rights concerning certain decisions based solely on automated processing that produce legal or similarly significant effects, subject to the Act’s exceptions and safeguards.

Again, the important distinction is between software assisting a genuine human assessment and software effectively making a consequential decision by itself.

Cayman Islands

The Cayman Islands data-protection framework also addresses automated individual decision-making. Guidance from the Ombudsman discusses e-recruitment and automated aptitude testing, recommends privacy impact assessment, appropriate notice, human reconsideration, accuracy controls and bias monitoring.

That is a sensible governance model even before an employer gets into the finer legal argument about whether a particular authorship detector falls inside a particular statutory provision.

Guyana

Guyana’s Data Protection Act 2023 includes transparency rights concerning automated decision-making, including profiling. In applicable cases, the legislation refers to meaningful information about the logic involved and the significance and envisaged consequences of the processing.

An employer processing applicant information through an automated system should therefore think about what it can actually explain to the person whose application is being assessed.

Trinidad and Tobago

Trinidad and Tobago requires special caution in any summary. The Data Protection Act 2011 exists, but the 2012 proclamation brought only Part I and specified sections of Part II into operation at that time. It would therefore be misleading to cite every provision of the Act as though the entire statute had necessarily commenced.

Employers should obtain current, section-specific local advice rather than importing a GDPR analysis or relying on the title of the Act alone.

What is actually good about AI detection?

There can be benefits if the technology is kept in its proper place.

Suppose an employer is hiring somebody specifically for independent technical writing. Candidates are clearly told that a particular exercise must be completed without generative AI. One submission is dramatically different from the applicant’s earlier work, contains invented citations, uses terminology the candidate cannot explain and is also flagged by a detector.

In that situation, the detector may be one signal supporting further review.

One signal. Not the verdict.

The employer can ask the candidate to explain the analysis, defend the sources, walk through the reasoning or complete a shorter controlled exercise. Those steps generate much better evidence of competence and honesty than arguing over whether a model thinks a paragraph looks machine-written.

Detection can also reveal a weakness in the employer’s own assessment design. If almost every applicant can answer a screening question perfectly with a general-purpose model in thirty seconds, perhaps the question is no longer testing what the company thinks it is testing.

The disadvantages are harder to ignore

A false accusation can cost someone a job without the person ever knowing why.

Different writing populations may be treated differently. Translation and accessibility tools complicate authorship. Vendor accuracy claims may not match the employer’s applicant population. Applicants can deliberately edit generated text. Recruiters may over-trust a number because it looks objective.

There is also a basic relevance problem. AI detection often does not measure job performance.

If I am hiring a project manager, I care whether the person can plan, communicate, prioritise, handle risk, manage conflict, make decisions and take responsibility. Whether an AI assistant helped polish two sentences in a cover letter may tell me almost nothing about those abilities.

Employers need to decide whether AI use is misconduct or a workplace skill

This is where many recruitment policies become inconsistent.

A company can spend heavily rolling out approved AI assistants and training employees to use them productively, then reject a candidate because they used AI to improve an application. That policy needs a better explanation.

There are jobs and assessments where unaided ability genuinely matters. Test it.

There are jobs where responsible AI use is part of modern performance. Test that too.

A copywriter might complete one short unaided exercise and another task where AI is allowed but sources and edits must be explained. A developer can be asked to explain submitted code and modify it live. An analyst can be asked what evidence would change a recommendation. An administrator can demonstrate how they would use an approved assistant without exposing confidential information.

Recruitment should resemble the work.

Seven rules I would use if an employer insists on AI detection

Define permitted AI use before the assessment. Distinguish between grammar correction, translation, brainstorming, editing, coding assistance and full generation where those differences matter.
Test the skill you actually need. If independent writing matters, run a controlled writing exercise. If reasoning matters, ask the person to explain the reasoning.
Do not automatically reject solely because of an authorship-detector score. Treat the output as a potentially fallible signal requiring corroboration.
Require meaningful human review. The reviewer should understand the tool’s limitations, have authority to disagree with it and examine the evidence rather than rubber-stamping a threshold.
Validate the tool for the real use case. Ask the vendor for evidence on false positives, false negatives, language differences, relevant demographic testing, accessibility and independent evaluation. A headline accuracy percentage without the test population and methodology is not enough.
Be transparent and protect applicant data. Give legally required notices, explain material automated processing appropriately, limit data collection and understand what the vendor retains or reuses.
Create a challenge and reconsideration process. Where a detector materially influences an adverse decision, applicants should have a practical route to human review where required and, as good governance, even in many situations where the minimum legal rule is less explicit.

A better approach than trying to catch people

The strongest recruitment process is designed so that a candidate eventually has to demonstrate understanding.

Ask the applicant to discuss the submission. Change one fact in the scenario and ask what they would do differently. Ask what assumptions they made. Ask what could go wrong. Ask them to criticise their own answer. Ask a programmer why a function exists. Ask a writer to revise a paragraph. Ask an analyst which source they trusted least and why.

Someone can submit a beautiful generated answer. It is much harder to fake understanding through a serious conversation.

A policy that is easier for everyone to understand

An employer could say:

You may use AI tools for preparation and editing unless an assessment specifically says otherwise. Information you submit must be truthful, you must understand and be able to explain your work, and any exercise marked as unaided must be completed without generative AI.

That is far more useful than pretending every modern writing aid fits into a clean “AI” or “not AI” box.

If substantial generation of a work sample matters, the employer can also ask for proportionate disclosure. But requiring somebody to document every grammar suggestion or autocomplete is unlikely to produce meaningful integrity.

Fraud is different from assistance

If somebody invents qualifications, fake employment history, certifications, references or projects, the problem is dishonesty. AI may make the fabrication easier, but it does not change what the employer needs to establish.

Verify the claim.

If an applicant presents an AI-generated portfolio project as entirely their own and cannot explain any of it, test the substance. Do not reduce the investigation to a detector score.

What applicants should do

Applicants have responsibilities too.

Do not let an AI system invent experience you do not have. Do not submit technical work you cannot explain. Do not fabricate citations. Do not paste confidential material from a current employer into an external AI service. Read what the system produces and verify factual claims.

If an assessment says it must be completed without AI, respect the instruction.

And if AI is permitted for preparation or editing, make sure the final application still represents you. Eventually there will be an interview, a work sample, a probation period or a real problem that needs solving.

My conclusion

I would not ban every use of AI-authorship detection. I also would not let it become the judge.

Employers should not use an AI-authorship detector as standalone proof that an applicant used generative AI, and they should not automatically reject a candidate because of a detector score.

Where an employer has a legitimate reason to restrict AI use, the better approach is to state the rule before the assessment, design the assessment around the competence being tested, treat detector results at most as one fallible signal, provide meaningful human review, consider discrimination and accommodation risks, protect applicant data, document consequential decisions and provide appropriate routes for challenge or reconsideration.

If independent performance matters, test independent performance.

If responsible AI use will be part of the job, test responsible AI use.

Trying to guess authorship from statistical patterns should not replace either one.

The broader lesson is simple. Employers have used automation in recruitment for years: applicant tracking systems, keyword filters, online assessments, ranking systems and automated emails. Applicants now have powerful automation too. The answer should not be an endless technological war in which each side deploys another model against the other.

Recruitment still works best when it comes back to evidence, competence, judgment, transparency and accountability.

No applicant should lose an employment opportunity simply because an algorithm guessed who, or what, helped write a paragraph without a real person examining the evidence and the employer understanding the legal and human consequences of relying on that guess.

Sources and legal references

Legal and regulatory position reviewed 21 August 2026. This article provides general information, not legal advice. It highlights principal frameworks and selected jurisdiction-specific rules; it is not an exhaustive survey of every employment, privacy or AI law in the UK, every EU Member State, every US state or municipality, or every Caribbean jurisdiction. Organizations should obtain jurisdiction-specific legal advice before deploying consequential automated recruitment tools.