Canadian lawyers urge employers to treat AI as a supplement, not a substitute — especially with human rights and privacy obligations on the line
As artificial intelligence becomes embedded in many workplace tasks, HR professionals are navigating both the gains of efficiency and the legal issues — especially when it comes to hiring practices and performance management.
In both, the underlying message from employment lawyers speaking to Canadian HR Reporter is consistent — AI is a tool, not a decision-maker.
"This is not intended to discourage employers," says Arielle Sie-Mah, an associate at Gowling, "but really it’s about thinking through how to implement AI intentionally where it can improve efficiency and processes, while still striking that balance with managing the associated legal risks."
Recruitment: bias by algorithm
When employers use AI to write job postings, scan resumes or rank candidates, the efficiency gains are obvious, but there are definite risks when it comes to bias entering the process.
In many cases, AI recruitment tools are scoring resumes on a scale of one to 100 against criteria the employer has defined.
“The employer is going to need to input into whatever AI system they're using what they're looking for in the ideal candidate. It's up to the employer to make sure that the descriptive wording being used is accurate, fair and reasonable, and that the result they ultimately get from the AI system matches that,” says Michael Grodinsky, a partner at BLG in Montreal.
But tools trained on historical hiring data can reflect and replicate biased patterns from the past.
“The overarching theme when it comes to using AI is just being extremely cautious about not unintentionally tripping over the human rights obligations that are intimately there,” says Sie-Mah.
“You can't blindly rely on past practices and past experiences, because some of them could be skewed and not really examined in terms of what were the reasons things looked the way they did in the past, and what has changed now.”
Know AI's limitations
Shefali Rajaputra, lawyer at Gowling in Montreal, agrees that employers need to be aware of AI’s limitations when it comes to potential accommodations for individuals or perpetuating hiring biases.
“Ultimately, the responsibility for the recruitment process still lies with the employer to comply with their obligations under human rights legislation, privacy, and employment standards legislation.
"If an AI process doesn't take into account those nuances under human rights legislation, it has the effect of perpetuating potentially discriminatory practices."
Another challenge? Job candidates are catching on.
"Many applicants are, at this point, very used to dealing with AI-based systems and they understand how those systems will filter their own resumes," says Grodinsky, which means employers relying solely on AI output may be selecting for people who are good at gaming the algorithm, but not necessarily the best fit for the role.
"[Employers] have to perform, in many cases, their own redundancy test — to make sure that whatever's been output by that AI program is in fact accurate and reliable in terms of screening," he says.
Ontario's disclosure requirement
Ontario is currently the only province with a statutory requirement specifically around AI use in recruitment. Under amendments to the Employment Standards Act, employers with 25 or more employees must disclose in publicly posted job ads if AI is being used in the hiring process.
Sie-Mah says that disclosure requirement opens the door to scrutiny.
"In my view, that disclosure essentially raises a flag and opens a door to candidates, workplace discussion boards, and even plaintiff counsel to come and ask more questions — ‘What tool did you use, what were the metrics used to assess that applicant, why was this candidate rejected?’”
She draws a direct line to how hiring decisions have always worked. Managers were expected to document their thinking, explain non-discriminatory criteria and maintain notes in case of a complaint.
"Instead of the manager, now your AI tool needs to be able to establish that," she says. "I always like to look at things from the perspective of: ‘How is this going to be viewed by a court or a tribunal when it ends up there, and is that going to be a good look for you as an employer?’”
Chatbots and candidate communications
As AI-powered chatbots handle initial candidate queries and interview scheduling, another layer of risk emerges. What are these tools saying to people, and could there be legal liability?
Sie-Mah draws a distinction between a basic chatbot — one that simply acknowledges receipt of a query — and one that answers questions based on what an applicant has asked. "There could be legal ramifications to that, and they could be held accountable for it," she says.
"Employees or third parties receiving communication through a chatbot are going to see that as an extension of the organization it's speaking on behalf of," says Sie-Mah. "Organizations using those chatbots should be very careful about what content is being produced by them."
Grodinsky frames it in terms of organizational values: a chatbot is a frontline representative, and the same standard applies whether that representative is human or AI.
“You can have a racist, sexist, or discriminatory human being picking up the phone and calling people, just as you could perhaps have the same experience with an AI bot. But the onus is on the company or the employer to make sure those types of things don't happen and that reasonable safeguards are put in place to prevent it.”
Rajaputra adds that automated communications need to be monitored carefully not just for human rights and employment law compliance, but for privacy obligations as well — particularly in provinces like Alberta, BC and Quebec where privacy legislation governs the personal information of prospective employees.
"Despite there not being an express statutory obligation, the laws currently in place do suggest that that kind of consent and notice is something employers should be doing," she says.
Offer letters: basic template not enough
All three lawyers emphasized that any AI-generated document that goes to a candidate must be reviewed by a human before it goes out the door — and nowhere is that more important than in the offer letter.
"Employment agreements are like prenups when you're getting married," says Sie-Mah. "They are tools that will help you resolve issues or solve for issues even before they come up."
The termination clause, in particular, is not something an AI template should be trusted to get right: "That defines the interpretive profile should things work out differently," she says.
Rajaputra agrees: "If there's any use of AI-generated documents, those should always be reviewed by a human to ensure that the important clauses you would like to see in an offer letter or hiring document are contained in it."
Third-party considerations
And using a third-party vendor carries the same considerations, according to Sie-Mah.
"The biggest thing employers need to be thinking about is asking the right questions of vendors before just signing contracts," she says. "How does the AI really function? What metrics are being used to assess candidacy? What data was used to train the AI to pick and choose?"
Third-party vendors handling the hiring process also bring legal obligations: "The liability does not shift by using a vendor.”
When it comes to third-party vendors, Grodinsky says liability is shared — but that doesn't let employers off the hook. The onus is on the employer to understand what they've purchased, ensure the vendor provides proper training and onboarding, and designate an internal person with sufficient knowledge to operate the tool responsibly.
"It's not just about buying a product and using it," he says. “It's the company's responsibility to make sure that person — whoever the internal stakeholder is — has sufficient training to operate it in a way that minimizes the employer's liability."
Performance management: context is everything
A second major area of AI adoption in HR is performance management — and here, the risks are arguably more nuanced because AI may be asked to evaluate human behaviour and output, not just sort structured data.
Used carefully, AI can make performance management more consistent and less burdensome; AI-generated first drafts of performance reviews can, in some cases, be superior to what managers might produce under time pressure, says Grodinsky.
"But just because the AI does it for you doesn't mean the result is going to match the context,” he says, so HR should be reviewing the content for accuracy. “You wouldn't want an overly harsh performance review for somebody who is an acceptable or above-average performer. You want to make sure, for retention purposes, that the context doesn't lead to somebody souring on their job or their employer.”
That problem runs deep. Quantitative metrics — keystrokes, email response times, deadlines met or missed — tell only part of the story. Sie-Mah notes that AI systems tracking those metrics miss the client-facing, creative and collaborative dimensions of a role entirely.
More troublingly, poor metrics may reflect an accommodation issue rather than a performance issue.
"Certain things like attendance or productivity might have quantitative metrics that suggest low productivity or poor performance, when there could be underlying accommodation issues that the employer does need to inquire into and address," says Rajaputra.
'Nuanced and complex'
The accommodation assessment piece is nuanced and complex, adds Sie-Mah.
"It's not a one-size-fits-all solution. You can't just feed certain prompts to AI tools, because it is entirely fact-specific and driven by the context of each case."
The stakes escalate sharply when AI moves from monitoring to recommending action. She says she sees serious problems with AI tools that don't just track performance data but generate conclusions from it — such as telling the employer that termination for cause is warranted.
"I think that cuts against the employer's obligations of good faith and procedural fairness embedded within the employment relationship," Sie-Mah says. "If an employer relies on an AI-generated performance report, they should still be able to show an underlying process undertaken in good faith."
AI can be a great tool for organizing and compiling information, or developing a process for conducting performance reviews, says Rajaputra.
“But so much of performance management relies on the specific context and circumstances of that workplace, of the employee, and their job duties. So it's important — as you say — to have that human element, because AI cannot fully understand the nuances that might contribute to fluctuations or variations in an employee's performance,” she says.
“Certain things like attendance or productivity might have quantitative metrics that suggest low productivity or poor performance, when there could be underlying accommodation issues.”
The evidentiary trap
When it comes to performance management, you're really asking the AI system to look at people's output, compare people against each other, or compare them to benchmarks the company has set internally — whether it's for productivity, responsiveness to customers or suppliers, attendance, work from office versus work from home, work accidents, says Grodinsky.
“These are different things. You're asking the AI to render a more sophisticated work product in that case.”
In the United States, courts have begun requiring disclosure of the AI prompts used to generate performance assessments. Canada is not there yet — but it’s an important consideration, according to Sie-Mah.
"If you have managers with access to AI tools and one of them types in 'Write me a performance review justifying termination for this employee,' and then there's litigation underway with respect to that termination,” she says. "Prompts are forms and they’re potentially admissible as evidence. That is a really bad look."
Sie-Mah says the way to mitigate the prompt-as-evidence risk is to train managers who have access to AI tools on what they should not be doing when using them.
Disciplinary letters and progressive discipline
When it comes to using AI for discipline concerns, Grodinsky says a disciplinary letter is "probably on the lower end" of a sophisticated AI work product: “You can put in a set of facts or parameters and arrive at an accurate warning letter, suspension letter or termination letter.”
But Rajaputra cautions against over-relying on AI when applying progressive discipline, citing “nuances and specific contextual factors” in employment law that AI can't necessarily account for: "In imposing disciplinary action, it's about relying on AI as a supplementary tool — not as the decision-maker."
Sie-Mah takes the same view, saying AI should never be used independently to apply any kind of disciplinary or corrective action.
AI note-taking: don't skip the review
One area that receives less attention but carries meaningful risk is AI-generated meeting notes. In disciplinary and performance conversations especially, what gets documented matters enormously — and AI transcription tools can introduce errors.
"Anything generated by AI will still need a human to review, confirm and finalize," says Sie-Mah. "AI is certainly going to reduce the time humans spend populating their notes, but it's still not going to take away that final step of confirming — ‘Yes, this is the discussion I had, accurately recorded.’"
The stakes of skipping that review are significant.
"There could be things in there that were not agreed to, things that are completely opposite of what was discussed, and potentially comments that could be perceived as offside of some statutory obligations," she says.
Best practices: audit, train, assign accountability
Across both recruitment and performance management, the lawyers converge on a few concrete recommendations.
First, appoint an internal AI accountability lead — someone Grodinsky calls "an AI quarterback" — who understands both the technology and how each department uses it. "If a question arises about how such a system renders such a result, there is an internal understanding of how that process works," he says.
Second, run regular internal audits. " Taking the recruitment screening process as an example — when we feed data into the system, is it giving us the result we think it will?" says Grodinsky.
Third, train the people who use AI tools on what those tools are and aren't capable of, and specifically on the kinds of inputs that could create legal problems later.
And above all, keep humans in the loop.
"Relying on AI as a supplement, relying on it to get tasks completed — that's good," says Sie-Mah. "But relying on it for judgment and letting things prepared by AI go out without human oversight and confirmation — that's a mistake."
Rajaputra sums up the fundamental principle: "We're dealing with humans, after all, and there could be a variety of reasons that AI can't fully appreciate."
