Ottawa allows AI to help decide who advances in federal hiring, but is there enough oversight?
Since March 2025, Public Service Commission guidance has explicitly allowed artificial intelligence to help determine which candidates advance in the federal hiring process.
Under the guidance, first published in March 2025, the commission—which oversees merit-based federal hiring and reports independently to Parliament—says AI can assist hiring managers with assessing, scoring, sorting and ranking candidates, all of which are stages that can determine who advances in a competition and who does not.
It says a set of safeguards—including bias testing, human oversight, candidate notification, accommodations, and recourse—is meant to keep those risks in check.
But experts and labour representatives interviewed by The Hill Times say some of those safeguards may be weakest exactly where AI-related harms are hardest for candidates—or government itself—to detect.
The guidance comes as the Public Service Commission's (PSC) own 2025 survey shows federal hiring managers were already using AI at multiple stages of staffing, while government records show AI-powered assessment tools have been used in federal hiring processes.
At the same time, federal hiring has slowed sharply, with new permanent hires falling 40 per cent in 2025-26, to 4,630.
AI moves from guidance to practice
The Treasury Board Secretariat (TBS) paid $118,650 between October 2023 and March 2024 for Knockri, an AI-powered platform used to conduct and assess candidate interviews for administrative services and executive staffing processes, according to government records tabled in the House of Commons. TBS said it completed both an algorithmic impact assessment, which evaluates the risks of an automated decision system, and a privacy impact assessment before using the platform.
TBS told The Hill Times on Aug. 11 that Knockri was used voluntarily in a one-time pilot across three selection processes in 2024, with about 230 candidates assessed using the platform and fewer than 40 assessed without it.
When the system found that a candidate did not meet an assessment criterion, an HR adviser reviewed the candidate’s audio response to validate the result. TBS said none of those reviews changed a candidate’s result, there were no complaints or challenges, and the secretariat is not currently using Knockri or any other AI tool to assess candidates.
An earlier Department of National Defence hiring process shows how consequential that kind of assessment can become. In a 2020 campaign for entry-level executive positions, 422 of 471 applicants completed a Knockri assessment. The platform analyzed the content of candidates’ recorded responses against predetermined criteria, and 74 candidates passed that stage and moved on to Plum, another AI-enabled assessment tool that evaluated personality, problem-solving ability, social intelligence and job fit.
DND's report said its four-member assessment board retained ultimate decision-making authority and that samples of candidate responses were manually reviewed as part of quality assurance. But the process also shows how AI-generated scores can influence who advances even when humans retain final authority.
DND told The Hill Times on Aug. 11 that it has not used Knockri or Plum in staffing processes since that pilot.
The commission’s guidance is meant to govern uses like these. For staffing uses covered by the Treasury Board's Directive on Automated Decision-Making, departments must complete an algorithmic impact assessment (AIA) before the system is deployed.
The assessment determines the system's level of potential impact and, in turn, which requirements apply. The rules also require measures around testing, transparency, human involvement and recourse, with some requirements becoming more stringent as the assessed impact increases.
The PSC guidance separately requires managers to identify and mitigate biases and barriers in assessments, tell candidates when and how AI will be used, provide information about accommodations, and be able to explain how an AI tool contributed to an assessment or decision.
The commission told The Hill Times its Personnel Psychology Centre contributed expertise on issues including validity, adverse impact, bias mitigation, transparency, and human oversight, while departments and agencies are responsible for complying with the applicable safeguards.
Experts and labour representatives interviewed by The Hill Times, however, questioned how well those protections work when bias, errors, or inappropriate reliance on an automated assessment are difficult to detect. One federal public servant’s account of a 2024 staffing process shows how uncertainty can arise when hiring evaluators try to apply rules around AI in practice.
‘We were using another AI to evaluate someone’s alleged AI’
The difficulties of applying rules around AI in individual staffing processes were already apparent before the PSC first published its guidance in March 2025.
A federal public servant who participated as an evaluator in a 2024 external national staffing process for a small federal organization told The Hill Times, on condition of anonymity, that candidates had been explicitly instructed not to use generative AI in their application responses. Evaluators were asked to flag suspected AI use to human resources officials, but were not given a common standard for identifying it.
Some responses contained obvious signs that material had been copied from a chatbot, the evaluator said, including chatbot disclaimers or instructional text inadvertently left in the response.
The harder cases were far more subjective. At one point, someone suggested that unusually polished, logical, and well-organized writing could itself indicate AI use, an approach other evaluators objected to because of the bias it could introduce.
“They were trying to be zero tolerant, but they didn't know what they had no tolerance for,” the evaluator said.
The evaluator said that, for less obvious cases, staff turned to publicly available third-party AI-detection websites, pasting candidates' complete responses into the tools and receiving percentage estimates of whether passages were AI-generated. The detector did not explain why it classified particular passages that way, he said, and did not resolve the underlying uncertainty.
“We were using another AI to evaluate someone's alleged AI,” he said.
He also raised privacy concerns about pasting candidates' complete responses into third-party websites without knowing what happened to the information once it was submitted.
The PSC's current guidance cautions that AI-detection tools may be unreliable and says using them to determine whether a candidate used AI is itself an automated assessment subject to the Treasury Board's Directive on Automated Decision-Making.
Where bias is hardest to measure
One potential weakness is conventional bias testing, according to Jutta Treviranus, director of the Inclusive Design Research Centre at OCAD University and chair of the technical committee that developed Accessibility Standards Canada’s national standard on accessible and equitable AI systems.
Treviranus said conventional bias testing generally assumes there is an identifiable group against which outcomes can be compared, an approach that can work for characteristics such as gender or language. Disability is different, she said, because there is no single identifiable pattern.
“People with disabilities tend to be an ‘N of one,’” Treviranus said. “The only common thing amongst people with disabilities is difference from the average.”
AI hiring systems that try to match candidates against dominant patterns of past success can disadvantage people with atypical employment histories or different ways of communicating and working, she said. Because those harms can differ from one disabled candidate to another, Treviranus said they can be dismissed as anecdotal or statistically insignificant, making discrimination difficult to identify and prove.
Candidate notification, another safeguard in the PSC framework, raises a separate question about what applicants can actually learn when AI affects their assessment.
Lorin MacDonald, a disability-rights lawyer, addressed that issue in a March opinion piece for Canadian Lawyer about an Ontario requirement for employers with 25 or more staff to disclose whether they use AI to screen applicants.
“Disclosure is where accountability begins,” she wrote. “It cannot be where it ends.”
MacDonald argued that knowing AI was used does not necessarily tell a rejected applicant why they were screened out, whether the system had been tested for discriminatory effects, or who is accountable for the result.
Oversight under scrutiny
A similar concern applies to another safeguard in the PSC framework: human oversight.
Sean O'Reilly, president of the Professional Institute of the Public Service of Canada (PIPSC), said retaining a human decision-maker does not necessarily amount to meaningful oversight, particularly as departments face pressure to do more with fewer resources.
“They'll just accept what the AI has given them,” O'Reilly said, adding that he worries managers may not go back and reconsider candidates an AI system has screened out.
O'Reilly said PIPSC also lacks enough information about how AI is being used in staffing, including which AI systems are being used and how the rules governing them are being applied and enforced. Those gaps make it difficult for the union to judge how effective the government's safeguards will be, he said.
“You can say there's bias testing, but my wonder is: what does that actually mean in practice?” O'Reilly said.
He said bargaining agents should be consulted before AI is introduced into hiring or promotion processes and that, as far as he was aware, those guardrails had not been discussed with PIPSC.
“I feel that we have more questions than answers at this point,” he said.
O'Reilly also called for federal AI legislation, an independent regulator, and a permanent government-labour committee to oversee AI implementation across the public sector.
Those questions about how AI use can be scrutinized extend to one of the government's main transparency requirements.
For systems covered by the Directive on Automated Decision-Making, departments must publish the results of their algorithmic impact assessments. Teresa Scassa, Canada Research Chair in Information Law and Policy at the University of Ottawa, said that requirement has not always been followed consistently.
“This requirement has not been rigorously met,” Scassa wrote in an Aug. 5 email to The Hill Times.
Scassa said she had seen roughly 20 to 30 published assessments when she checked a couple of months earlier, compared with 42 when she looked again in August. Of those, she said she was able to find only one related to employment, an assessment for a federal hiring pilot.
She also pointed to uncertainty around what happens when someone wants to contest an automated decision. Although the directive requires mechanisms for challenge or appeal, Scassa said it is unclear what those mechanisms will involve, or how far they will go.
The PSC told The Hill Times it does not centrally track the use of AI in staffing processes. Its 2025 survey found 5.8 per cent of hiring managers reported using AI tools such as ChatGPT, Gemini, or Copilot, most often to prepare assessment tools, merit criteria, and job posts.
When asked whether it had received any investigation requests from candidates alleging unfair treatment arising from an employer’s use of AI in an appointment process, the commission said it had received none.
Editor’s Note: This story has been updated on Aug. 14, 2026, at 3:11 p.m. to clarify that the Public Service Commission’s guide on artificial intelligence in hiring was first published in March 2025, and that the provisions on assessing candidates were included in that original version. The PSC said a June 2026 update made no substantive changes.
The Hill Times