We still call it talent acquisition, but that’s not what your stack is doing anymore. Your hiring system is an access-control layer that strangers touch all day. When trust breaks, the question isn’t just who did we hire. It’s who did we let in.
Candidates are using AI to compress and polish themselves at scale. Recruiters, buried in volume and judged on speed, are using AI to compress and process candidates at scale. Platforms keep rolling out new AI features, sometimes on purpose, sometimes because an update ships while everyone’s in meetings. Somewhere in the middle, your ATS quietly stopped behaving like a neutral HR database and started functioning like access control.
The shadow AI stack, now in hiring
Last week, I wrote about the shadow AI stack running through organizations whether leaders approved it or not. This is the sequel, but this time the flashlight is pointed at hiring.
The calm story and the reality beneath it
Canada keeps telling itself a calming story about AI. We aren’t seeing the same big, AI-branded white-collar layoffs that dominate headlines elsewhere. Early Canadian data may not show jobs falling off a cliff in roles more exposed to generative AI.
That calm might be real in the numbers. It isn’t real in how people feel.
On paper, AI adoption still looks small. Statistics Canada’s latest analysis puts AI use among Canadian businesses at 12.2 percent – companies using it to produce or deliver, not just write emails. That number doubled in a year. Leaders tend to find that comforting. Double of small is still small, and the rollout plans make it look like most organizations are still watching. But that survey measures what companies report. It doesn’t measure what people do when they’re under pressure, and nobody’s looking.
What actually shapes behaviour
The real force shaping behaviour isn’t your policy manual or your town hall. It’s what people see when they’re scrolling in the dark.
They see headlines about AI reshaping work and asking uncomfortable questions about the future. They watch videos of people who have been out of work for months, describing how they are adjusting and what that is doing to them. In the same feed, they see executives praising AI for efficiency and leaner teams, followed by layoff announcements and rising stock prices. They see the incentives and the outcomes at the same time.
Whatever story you’re telling inside the company is competing with that. If you don’t talk about it, you don’t make it go away. You leave people to figure it out on their own.
The psychological contract under pressure
So when leaders say they are moving slowly, acting cautiously, building governance, they may mean it. What a lot of employees hear is something else.
This is where the psychological contract stops sounding like HR language and starts behaving like infrastructure. Underneath engagement work and employer branding is an unwritten understanding that if you show up and contribute, the organization will not treat you as disposable.
When people stop believing that, they don’t forget how to work. They quietly reprice the relationship. Some leave. Many stay and narrow their contribution to what is required. More importantly, they stop volunteering information. They stop experimenting in the open. They stop assuming anyone is steering. They don’t stop using the tools. They just stop letting you see them. That isn’t rebellion. It’s self-protection.
How shadow AI actually spreads
That’s how shadow AI takes hold. It doesn’t start as defiance. It starts as trying to get the work done.
A recruiter picks up an unapproved sourcing tool. An HR lead connects workflows with whatever is available. A team attaches AI to systems that were never built for it just to keep up with demand. Each step makes sense on its own. Taken together, they become an unmonitored network of data flows, dependencies, and weak spots that nobody has signed off on.
Now apply that directly to hiring.
Hiring has shifted: selection, admission, and access control
Hiring used to be selection. Now it is admission. Admission behaves like access control.
Your applicant tracking system isn’t a neutral HR database. It is an access point with a candidate experience attached. It holds identity data, work histories, contact details, assessments, internal notes, and decision trails. It resembles a passport office more than a filing system, and it is designed to be touched by the outside world continuously.
Candidates interact with it. Agencies interact with it. Vendors interact with it. Screening tools, scheduling tools, chatbots, and every new AI helper interact with it. Every integration is another hand on the door handle. Every hand on the handle is a risk you picked up, whether you intended to or not.
The illusion of “we haven’t turned AI on”
This is where many organizations draw a distinction that no longer holds. They say AI has not been turned on in hiring. That may be technically correct. It is not operationally meaningful.
Candidates are already using AI. Recruiters are already using AI. Platforms continue to release AI features, sometimes intentionally, sometimes as part of routine updates. Hiring becomes a system where candidate automation and employer automation meet. The process becomes smoother. The signal does not necessarily become clearer.
When “what’s real” becomes the question
The question starts to shift underneath the process.
Real people are there. Real people supported by AI are there. Synthetic profiles and generated materials are beginning to appear as well. High-quality output is easier to produce, regardless of the underlying capability.
At scale, hiring starts to resemble working a door where every ID appears professionally produced and the expectation is to keep the line moving. They don’t need to look exceptional. They just need to look normal.
When that happens, the question isn’t just who did we hire. It’s who did we let in.
A real-world failure
In 2025, security researchers identified a vulnerability in McDonald’s AI-powered hiring platform, McHire. The system was operating with default administrative credentials, lacked multi-factor authentication, and exposed applicant data through an internal API that allowed record-by-record access.
The issue affected tens of millions of applicants before it was addressed. It was not a complex failure. It reflected basic access control gaps in a system designed for high throughput.
Where attackers actually start
When systems connect, entry points shift.
If you were looking for a way into an organization, you would not start at the firewall. You would start where access is already built into the process. Hiring systems are designed to be interacted with externally. Around them, additional tools accumulate – sourcing extensions, scheduling assistants, automation layers. You do not need to force entry where entry is already expected.
The bill is coming either way
If AI delivers the return organizations expect, decision-making continues to move upstream into systems while human judgement becomes less central but remains accountable. If it does not deliver, or never scales beyond isolated use cases, the organizational effects of pursuing it remain.
Workforces still adjust. Trust still shifts. Visibility still changes.
Either way, there is a cost. It does not always appear first in system performance. It shows up in behaviour.
Trust as infrastructure
This is where the psychological contract connects directly to the system.
When trust is intact, people surface risk early. They ask before placing data into tools. They flag behaviour that looks unusual. They run experiments where others can see them. When trust is limited, the work does not stop. It moves out of view.
The more you moralize tool use, the further it moves out of sight. Processes are followed on the surface and worked around underneath. The system appears stable because it has stopped receiving signal.
The myth of “human in the loop”
“Human in the loop” is often described as a control. Its effectiveness depends on where decisions are actually made.
If systems determine visibility, screening, and ranking, and the human role is to review output within a time constraint, the primary judgement has already been applied. The human remains responsible for the outcome, regardless of how much influence they had over the decision.
Naming the reality
By the time people are working around processes and keeping tool use out of view, the system is already operating under reduced trust. Hiring begins to resemble controlled entry inside a system that is under strain rather than one operating with full alignment.
So what’s left to do
The response is not to prohibit AI or to assume it has not been adopted.
It is to recognize the ATS as part of the operational perimeter. It is to fund judgement in the same way speed has been funded. It is to assign clear ownership for decisions, including the authority to slow the process when needed.
Shared accountability has a way of removing accountability altogether.
The access problem underneath
The access problem isn’t just who gets into the building. It’s whether the people already inside trust you enough to tell you where the doors really are.