How to document skills for recruiters who use AI screening

Employers in Ontario with 25 or more employees have had to state in public job ads whether they use AI to screen, assess or select applicants since January 1, 2026. The EU’s hiring-specific AI rules now start on December 2, 2027, after a delay, and South Korea’s AI Basic Act has applied since January 22, 2026. Which of these reaches a given application depends on where the employer and the job are based, and the disclosure Ontario requires covers whether AI is used, not how it scores anyone.

The tools are less settled than the law. A screening system can parse a file into fields, filter on fixed criteria, or rank candidates with a language model, and the evidence differs for each. Parsing faults are documented by vendors but disputed in detail, exact-match filters are admitted by employers to exclude qualified people, and controlled tests of language-model rankers have found name bias and a preference for AI-written text. None of the studies cited below tests whether a particular CV layout or keyword count changes a hiring outcome.

Parsing, filters and rankers: where each one fails

Parsing comes first. The software extracts text from the uploaded file and sorts it into fields such as name, job title, dates and skills. Workable, a recruiting-software vendor, explained in 2023 that parsers ignore everything except text, so image banners, odd character spacing, headers, footers, tables and columns can lose or scramble details.

Vendor advice on columns is not uniform. A 2026 guide from the vendor RecruitBPM says documentation and testing from Greenhouse, Workday, iCIMS and SAP SuccessFactors show that columns built with a word processor’s native column tool are not automatically a problem, while tables and text boxes remain unreliable. It still calls a single column the safest default. Neither vendor publishes error rates, and this research found no independent measurement of how often a parsing fault decides an outcome.

Filtering applies fixed criteria to the parsed record. A 2021 Harvard Business School and Accenture study, summarized by HR Dive, found that 88% of employers surveyed said qualified high-skill candidates were screened out because they did not match the job description’s criteria exactly, and 94% said the same of middle-skill candidates. The authors recommended replacing negative filters, such as excluding anyone with an employment gap or without a particular degree, with affirmative searches for named skills. The figures are employers’ own reports from 2021 and do not show how often any one system does this.

What controlled tests found: names, short CVs and AI-written text

Ranking is the newest layer, and the most tested. In a University of Washington study presented in October 2024, researchers swapped 120 first names across 554 real resumes and ran three open-source language models over 571 job descriptions in nine occupations. The models favored white-associated names 85% of the time against 9% for Black-associated names, and male-associated names 52% against 11% for female-associated names. They never preferred a resume with a Black male name over an otherwise identical one with a white male name.

The limit is in what was tested. The models were open-source systems, and lead author Kyra Wilson said proprietary tools can only be approximated by researchers, so the rate in commercial software is not known.

A second University of Washington experiment tested the human review that employers often cite as a safeguard. In 1,526 screening scenarios with 528 participants, TechRepublic reported, reviewers chose white and non-white candidates at equal rates without AI advice, but followed an AI recommendation that favored one racial group up to 90% of the time. In the first study, according to the same report, cutting resumes to a name and job title raised biased outcomes by 22.2%.

A June 2026 working paper from the University of Maryland’s Smith School looked at a different problem: models rating text written by the same model more favorably. The Smith School’s summary reports self-preference rates of roughly 67% to 82% when AI-written resumes were compared with human-written ones, across more than 2,200 resumes. In simulated hiring pipelines across 24 occupations, candidates using the same AI system as the employer’s screener were 23% to 60% more likely to be shortlisted than equally qualified candidates with human-written materials. The results come from simulations of a working paper, not from observed hiring, and an applicant cannot know which model an employer uses. The authors found that instructing the evaluator to ignore a resume’s origin, or using several models in a majority vote, cut the effect by more than half.

Ontario: AI disclosure in the job ad since January 1, 2026

Ontario employers with at least 25 employees must say in any publicly advertised posting, and in the application form, whether AI is used to screen, assess or select applicants, Canadian HR Reporter notes. A companion law raised penalties for Employment Standards Act offences from $50,000 to $100,000. Employers with fewer than 25 employees are exempt.

The visible effect is on postings. Indeed Hiring Lab found that the share of Ontario postings with AI-related terms rose from 9% in October 2025 to 28% in May 2026, against 4% to 12% in the rest of Canada and an average of 7% across nine other advanced economies in May 2026. Many of the postings also state that people make the final hiring decision. Outside the technology sector, daily applications per posting were similar whether or not a posting mentioned AI.

The rule requires disclosure and nothing more. Employment lawyer Matthew McCarthy told the same publication that it resolves very little on its own. The Indeed measure also counts AI-related terms, which is a proxy for the statutory disclosure rather than a count of employers using screening software.

EU: Article 22 applies now, the AI Act’s hiring rules from December 2, 2027

In the EU, a rule already in force applies first. Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects. Exceptions exist for contract necessity, authorization by law and explicit consent. Where the contract or consent exceptions are used, the employer must offer at least human intervention, a chance to state a view and a way to contest the decision.

The AI Act adds a second layer later. The Council adopted an amending regulation on June 29, 2026, and it was published in the Official Journal on July 24 and entered into force on July 27, law firm Garrigues reports. It moves the deadline for Annex III high-risk systems, which Garrigues says include uses in education and employment, from August 2, 2026 to December 2, 2027. Article 50 transparency obligations were not deferred and took effect on August 2, 2026.

A separate, general Digital Omnibus that would amend other EU laws, including the GDPR, was still awaiting final approval in July 2026. The description of Article 22 above reflects the text as it stood then.

South Korea: employment judgments counted as high-impact AI since January 22, 2026

South Korea’s Framework Act on the Development of Artificial Intelligence and Establishment of Trust took effect on January 22, 2026, according to law firm Littler. The Act defines high-impact AI as systems that may significantly affect human life, safety or fundamental rights, and Littler reads it as expressly including judgments or evaluations affecting individuals’ rights or obligations, such as employment determinations. Littler says the category may reach resume screening, candidate ranking and AI-based skills assessments.

Operators of high-impact AI must notify users in advance that AI is being used, and are expected to keep risk management plans, safety documentation and human oversight. The Act applies to foreign operators whose systems affect Korean users, and a local representative is required above thresholds of 1 trillion won in global revenue, 10 billion won in domestic sales or 1 million daily users in Korea.

The duties fall on AI business operators, and Littler notes that the Act does not create a broad workplace disclosure regime, so what an applicant is told depends on how the duty is applied to a hiring tool. Littler also reports government commentary suggesting initial enforcement will emphasize guidance, although the Act provides for administrative fines and corrective orders. Detailed requirements are expected in presidential decrees and ministry guidance.

Next fixed date: December 2, 2027 in the EU

The next fixed date is December 2, 2027, when high-risk obligations for Annex III systems, including those used in employment, begin to apply in the EU. Ontario’s disclosure rule is already live. Until the EU date arrives, Article 22 of the GDPR is the existing EU provision on solely automated decisions.

Two items have no fixed date. Korea’s detailed requirements depend on presidential decrees and ministry guidance still to be issued, and the general Digital Omnibus that would amend the GDPR awaited final approval as of July 2026.