The Accountability Gap is Symmetric
AI accountability work has named one side of the gap. The other side is the same gap.
There is a real and useful conversation happening right now about AI accountability inside organizations. It is the right conversation. Who decides. Who owns. Who governs. What gets logged. What gets reviewed. What gets escalated when something does not go as expected.
The conversation has an edge, and the edge is the organizational firewall. Decisions made by automated systems inside an organization sit inside the accountability frame. Decisions made by automated systems that affect people outside the organization sit outside it.
That boundary is starting to fail. The failure is structural, not incidental. And the shape of the failure is symmetric on both sides, which is the part the current conversation has not yet absorbed.
Start with what responsible-AI work on the organizational side has correctly named.
Deployment is moving faster than accountability design. Agents are being deployed to take actions on behalf of users (schedule the meeting, send the email, approve the expense, place the order) at the same time the question of who owns the outcome is being deferred. The action and the decision blur. A human ends up responsible for what the system did without having been in the loop when it did it. Most of the time, this works. The cases where it does not are the cases that are starting to surface in the press.
The second observation is harder to see from inside an organization that does not believe it has adopted AI: most organizations have already adopted AI. Holly Hartman has been making this point clearly. The systems already running inside the enterprise (Workday, Salesforce, Microsoft 365, Zoom, the embedded scoring inside whatever applicant tracking platform sits underneath the recruiting team) are already AI systems in everything but the contract line item. The decision to adopt has been made. The accountability question has not been answered, partly because leadership does not yet recognize that the adoption has occurred.
Hartman's framing is the clearest single line in the current conversation:
Every AI decision is a workforce decision.
Read carefully, that line already implies the part of the conversation that is missing.
The part of the conversation that is missing is the other side of the same transaction.
Every candidate applying to a posted role at a mid-market or enterprise employer in 2026 is being processed by AI systems they did not choose, did not see disclosed, and cannot inspect. The applicant tracking system that scored them inside Workday, Greenhouse, or Lever. The resume parser that decided which sections of their experience to read and which to discard. The ranking inside LinkedIn Recruiter that determined whether a recruiter ever saw their profile. The video-assessment scoring that ranked their interview response before a human watched it. The candidate did not select these vendors. The candidate cannot see the criteria. The candidate cannot tell whether they were screened by a human reading their materials or by a model that decided in milliseconds.
No candidate is told. Rejection messages, when they arrive, do not name the vendor that produced the decision, the criteria the model weighted, or the threshold the candidate fell below. The decision was made. It cannot be traced. It cannot be appealed. The candidate cannot find out who made it, what it was based on, or whether the basis would survive review if anyone reviewed it. The mechanism that produced the outcome is invisible to the person the outcome happened to.
The structural shape on both sides is identical.
On the organizational side, the accountability gap is this: a decision was made by an automated system, no named human owns the outcome, and affected parties cannot trace or contest it.
On the candidate side, the accountability gap is this: a decision was made by an automated system, no named human owns the outcome, and affected parties cannot trace or contest it.
The systems are different. The vendors are different. The decisions are different. Different parties stand inside the gap. The same mechanism produced it.
The accountability gap is symmetric.
Naming the symmetry is not a rhetorical flourish. It changes what counts as inside the conversation.
Responsible-AI work, as it is currently practiced, takes the organizational firewall as the edge of the accountability frame. When the question is who inside the organization owns the outcome of an automated decision, the people most affected by that decision (the people standing on the other side of the firewall) are not in the room. Their absence is not adversarial. It is structural. The frame did not include them.
Hartman's framing already names the principle that breaks the firewall: responsible AI requires that the people affected by automated decisions be able to understand them, question them, and appeal them. Accountability is a property of the whole transaction, not of one side of it. The symmetry argument this piece develops follows directly from that requirement.
Her closing observation makes it concrete:
Every HR technology platform processing candidate data deserves a critical eye.
That is the move. Not a parallel conversation about candidate-side AI as a separate domain. The same conversation, with the frame redrawn to include both sides.
The fix is not symmetric. The surfaces are different. What disclosure looks like for an internal AI decision is not what disclosure looks like for an automated rejection at the top of a hiring funnel. What contestability looks like inside an enterprise governance process is not what it looks like for a candidate who has no relationship with the organization that filtered them.
But the disciplines that make accountability real are the same on both sides.
What does accountability architecture look like on the candidate side. Four disciplines, all already named in org-side responsible-AI work, all extending across the firewall without losing their shape.
Disclosure. Candidates should know when automated decision-making is part of their evaluation. Not the vendor's commercial terms, not the model architecture, not the proprietary weights. Those are not the relevant facts. The relevant fact is that an automated system played a material role in deciding what happened to the application. The right time to disclose this is at the point of application, not in the rejection message six weeks later, and not nowhere.
Contestability. If a decision was automated, there should be a path to appeal it to a named human with the authority to override it. The org-side conversation has already named this requirement for high-stakes internal decisions. Hiring is a high-stakes external decision. The mechanism does not need to be elaborate. It needs to exist.
Auditability. Organizations should be able to answer, on demand, who made a hiring decision, on what basis, and whether the basis is defensible. "The model scored them low" is not an audit answer. It is the absence of one. The internal AI governance work that organizations are starting to invest in produces audit trails for decisions about employees. The same discipline applied to decisions about candidates would produce audit trails the candidate could reference if they asked.
Named ownership. Every automated decision in hiring needs a named human owner. Someone whose name attaches to the outcome. Someone who could be asked to defend it. The org-side governance conversation has named this requirement for internal AI decisions. Decisions about whom to interview, and decisions about whom to reject before any human reads the application, are not exempt because they happen to land on people the organization has not yet hired.
None of these disciplines are radical. They are the floor that org-side responsible-AI work has already named. Extending them across the organizational firewall is what the symmetry observation requires, once it is taken seriously.
Hartman's line is the clearest summary of what the org-side conversation has converged on: every AI decision is a workforce decision.
The extension the symmetry requires: every AI decision in hiring is also a decision that lands on a person who currently has no path to understand it, contest it, or learn who made it.
Taking both sides as inside the accountability frame is not a different category of work. It is the same work (disclosure, contestability, auditability, named ownership) applied across the boundary that the current conversation stops at.
The accountability gap is symmetric. The discipline that closes it has to be symmetric too.
That is the work that is still to be done.
