Who’s responsible when things go wrong? Employers need clear AI policies before disciplining workers for AI errors: lawyer
When AI goes wrong, who gets blamed for it? Many employees feel like it ends up being them, according to a recent study.
More than four in 10 (43 per cent) of employees say that when AI goes wrong, they take the blame. Adding to that, 41 per cent of CHROs believe many employees may not feel safe in challenging or overriding AI outputs.
The IBM report highlights the AI accountability gaps having an effect on employees, with 36 per cent of CHROs saying unclear accountability complicates AI deployment, based on a survey of 1,500 CHROs and 8,800 employees globally.
So, how and when should an employer hold an employee accountable if an AI tool goes wrong?
The answer starts with taking a closer look at the word "blame,” according to Vanessa Lapointe, a partner at Canadian law firm Borden Ladner Gervais LLP.
"The word 'blame' needs to be examined carefully. We should be careful about equating 'being blamed' with being unfairly disciplined.”
An ordinary work tool
Lapointe says that while AI might be a vastly different tool than many workplaces have worked with, employers still must remember it is an ordinary work tool and should be treated as such.
"Employers need to treat potential disciplinary or performance incidents the same way they would if an employee used any other work tool inappropriately. There's nuance in what 'using inappropriately' means — it's not the same as, say, using a company vehicle inappropriately — but the fundamental principles are the same," she says.
Just like with any other tool, if employees are trained on the tool and have a policy but don't follow it, they can face discipline, and it's the same with AI, Lapointe says.
"Take an example where there's a policy stating that a specific tool must be used internally. If an employee disregards that and uses a different tool, and that tool produces a hallucination — or even if it's the company-approved tool and it produces a hallucination — and the employee doesn't apply proper judgment, doesn't review the information, and simply copies and pastes it to a client or internal partner, resulting in a critical mistake or damage to the company, then discipline is certainly appropriate.”
The same logic applies if an employee feeds confidential information into an AI tool.
"They knew the rule, they broke the rule, and so there is a consequence," Lapointe says.
Setting the rules and policies first
That's why employers should ensure the proper framework is in place first, from approved tools to training, policies and rules, because disciplining outside of that is likely more difficult, says Lapointe.
That means having a good AI policy that covers approved tools, what information can be entered into AI systems, a required human review before AI-generated content is acted on, and what documentation and procedures should happen when things go wrong.
According to the IBM report, employees are more likely to feel comfortable pushing back on AI when HR has a say in how it's used. If the CHRO shares responsibility for which decisions remain human-led, 76 per cent of employees feel safe about questioning or overriding AI recommendations, compared with 43 per cent if HR is only advisory.
A fair investigation into AI use
When it comes to disciplining an employee if an AI tool goes wrong, Lapointe says the distinction comes down to whether employees used judgment and followed policies.
"Did the employee follow the policies, were they trained and aware, did they exercise judgment, double-check the information, and there was still a mistake? That's one thing. But if the employee didn't use appropriate judgment, didn't verify the information, and didn't follow the specific policies and procedures, then discipline is entirely appropriate," she adds.
Rather than setting rigid rules where a specific mistake automatically leads to suspension or termination, Lapointe says the focus should be on whether the employee's use of AI and their behaviour around it amount to a punishable offence.
Employers must also follow the same procedure they would for any disciplinary action: investigate, look into the facts, give the employee the opportunity to share their version of events, and make sure the discipline is proportionate, she adds.
"Think of it in terms of green, yellow and red flags. Say an employee's behaviour qualifies as a yellow flag. If it's their first, they're contrite, they apologize, they commit to training, they demonstrate a willingness to address the issue, and the mistake didn't cause extensive damage — that's one scenario.
“But if it's the same type of yellow flag, and it's the 10th time, and the employee lies, attempts to conceal it, blames others, takes no accountability, refuses training, and tries to repeat the same behaviour while under investigation — more serious disciplinary action may be warranted, potentially including termination.”
But that approach cuts both ways, as employers also cannot automatically assume that every AI-related error is solely the employee's fault, she adds.
"Disciplining an employee when it is fair to do so is entirely appropriate — but jumping to the conclusion that every AI mistake is automatically the employee's fault is not. Conducting a fair process is always the way to go."
