The cheaper AI option comes with a hidden cost: full accountability

Open-weight models hand Canadian employers the savings and all governance responsibility – at the same time

The cheaper AI option comes with a hidden cost: full accountability

A California federal lawsuit filed in July 2026 by 26 Meta Platforms employees alleging the company's AI-powered performance metrics disproportionately selected workers on medical, parental, and family leave for redundancy sent a signal far beyond Silicon Valley: when you deploy AI to make workforce decisions, you own the consequences.

The appeal of open-weight AI is straightforward. An open-weight model is an AI model whose core components are publicly released, allowing anyone to use it to run on their own systems and modify it for their specific needs, according to the Stanford University Human-Centered AI Institute. Models such as Meta's Llama series and Mistral allow organizations to download the trained model weights, customize them for internal workflows, and run them on their own infrastructure – without subscription fees or vendor dependency. As AI costs climb –with increased use and cost of AI tokens with  complete vendor-developed AI agents – and pressure on people budgets intensifies, more organizations are exploring this path.

According to a study by Futurum Research in September 2026, 42 per cent of enterprises were running open-weight AI models in production, which was an increase from 31 per cent in July. In a separate survey in October 2026, Futurum found that two-thirds of enterprises use open-weight models in some form, up from 46 per cent two months earlier.

What many are underestimating is how completely the liability shifts with the model. With a vendor-managed platform, governance responsibilities are at least shared. With a self-deployed model, they are not – and as HRD Canada has tracked in its coverage of how employers are deploying AI tools faster than governance frameworks can keep pace, the oversight gap is already widening.

Accountability follows the organization, not the tool

Canada's legal framework offers little shelter to employers hoping the source of the tool will dilute their exposure, according to Jeff Adams, a labour and employment lawyer at SpringLaw in Toronto.

"Even if it’s a vendor-created AI agent, they just don’t care,” says Adams, referring to decision-maker such as courts and tribunals. “The employer is ultimately responsible."

The reasoning is rooted in the structure of Canadian human rights and employment law, where intent carries limited weight in civil and administrative proceedings, says Adams. Whether a discriminatory outcome originated in a purchased tool or one configured and deployed in-house, the accountability standard is the same. Tribunals have already signalled they are not interested in arguments that an AI system bears any share of that responsibility, he says.

That position is reinforced by the design of Ontario's AI disclosure requirement, now in force under the Working for Workers Four Act, 2024. The rule requires employers with 25 or more employees to disclose the use of AI in hiring – but as HRD Canada has examined in its reporting on whether Canadian organizations are genuinely compliant with Ontario's AI hiring disclosure obligations, compliance is a legal starting point, not evidence of risk management.

Adams says he is also watching the litigation environment shift in real time. Employees are arriving with AI-generated demand letters that have not been reviewed by legal counsel. "I'm already seeing it," he says. "I'm getting demand letters pre-written from employee clients... but they're not vetting them." People, he says, are accepting AI output without the trained judgment to identify when it is wrong.

“From an employment law and human rights perspective, open-weight models created a bit of additional exposure, but the biggest concern I have is that there isn't a human in the mix,” says Adams.

What a Canadian investigation would look like

If a Canadian employer found itself in a situation analogous to the Meta lawsuit – a customized model producing workforce decisions that disproportionately affected employees on disability or family leave – the investigation would draw in multiple regulators at once, according to Adams. He outlines the likely pattern: a human rights complaint focused on the relevant protected ground; privacy regulators becoming involved where personal data was used improperly in model training or decision-making; and wrongful dismissal claims if the AI-assisted process produced unlawful outcomes.

Investigators would focus on what data was used, what variables were considered, and, critically, what human review mechanisms were in place. "If there were none, you're dead in the water," Adams says. The key legal questions would centre on whether the process created adverse impacts on a protected class, whether the employer can justify the practice, and whether reasonable steps were taken to identify and prevent discriminatory effects, he adds.

Building a governance model that holds

The governance question extends beyond legal liability, says Diana Valler, Chief Human Resources Officer (CHRO) and HR General Manager at H.I.S. Americas in Toronto.

"The governance needs to change automatically," Valler says. "It's a joint shared responsibility between human resources – because we have our governance to take care of from the policy and from the user/employee user perspective – and IT, which is extremely important for the same reason with security." She identifies siloed ownership as a risk in its own right: HR may overlook security implications while IT misses the human and policy dimension. Her practical advice: validate the platform, run a pilot before committing, and involve the people who will use the system in the evaluation. "Do not use AI unless the company has a policy and proper guidelines that you can follow," she says.

The cultural dimension adds urgency. According to the Ipsos AI Monitor 2026, AI and the Future of the Workplace – a 32-country survey of 23,532 adults conducted between March and April 2026 – 67 per cent of Canadians said AI makes them nervous, one of the highest rates globally. Introducing a customized open-weight AI model without adequate change management is likely to deepen that mistrust rather than close it, according to Valler.

As HRD Canada has documented in its analysis of the most critical pitfalls to avoid when deploying AI in HR, the governance gap in Canadian organizations is already a live concern. Open-weight models compress that gap into a question of direct employer accountability.

Human accountability for decisions and actions

For Adams, the single most important safeguard is “mandatory human accountability for every materially significant decision," he says – not checkbox review, but a trained decision-maker who understands independently why a recommendation is being made and bears personal accountability for the outcome. "Even the incorrect decisions that AI writes are written convincingly and persuasively, and that's most dangerous when people aren't trained and suspicious."

“The safest model is where AI functions as a tool informing trained human decision makers,” says Adams. “This is all because Canadian privacy, human rights, and the civil rules – and the courts and tribunals all fundamentally assume – that the accountability ultimately rests with the employer and not the AI.”

Ultimately, an organization’s leadership team, along with those who are going to be using the tool, should fully understand what they need in an AI tool, says Valler. “Do a big business analysis, gathering all the data and understanding if this is truly the solution that is going to be meeting our needs, because that's very, very crucial,” she says.

LATEST NEWS