The AI Hiring Screen Deserves Scrutiny
Compliance Issue: Colorado's AI Act mandates transparency and accountability from employers using AI in hiring.
AI Transparency: Many candidates aren't aware AI systems influenced their job applications or evaluations.
Interpretation Gap: HR departments often struggle to explain AI scoring models and their implications for hiring.
Vendor Liability: Under recent regulations, employers are liable for adverse outcomes from AI tools, not vendors.
Governance Shift: Organizations must prioritize explainability and documentation in AI hiring processes to ensure compliance.
Somewhere in your hiring pipeline, a candidate was filtered out by a system your team can't fully explain. You may have a vendor agreement, a dashboard, and a score. What most organizations don't have is a clear answer to the question that new AI regulations now effectively require you to answer: why.
That difference between what AI is doing inside enterprise hiring and what candidates and regulators can actually see, is now a compliance issue. Colorado's AI Act, among the first laws of its kind in the United States, took effect in February and requires employers using high-risk AI systems in employment decisions to notify applicants and provide a mechanism for them to correct data or appeal an adverse outcome.
For many HR departments, meeting that requirement means explaining systems they don't fully understand themselves.
Keep Reading—and Keep Leading Smarter
Create a free account to finish this piece and join a community of forward-thinking leaders unlocking tools, playbooks, and insights for thriving in the age of AI.
Have an account? Log In
The market for AI-powered hiring tools has expanded significantly over the last several years. Vendors offer screening platforms, automated interview scoring, and predictive fit models to employers managing application volume at scale.
The pitch is efficiency, but the math on the ground tells a more complicated story. Applications on LinkedIn surged more than 45% year-over-year in 2025, with roughly 11,000 hitting the platform every minute. At the same time, 64% of recruiters reported an increase in look-alike, AI-generated applications during the same period, which increased their screening workload rather than reducing it. The tools designed to filter the volume are, in part, generating it.
Tatiana Teppoeva, a former Microsoft and Boeing data scientist who now advises organizations on AI-assisted hiring risks, sees this dynamic play out consistently across enterprise clients.
Many HR teams understand vendor claims but have limited visibility into how screening outputs are generated or which candidate signals are being weighted. When candidates ask why they were screened out, organizations often struggle to translate a vendor score into a clear, job-relevant explanation.
In her experience, the gap is specific: organizations can describe the score a tool produces more easily than they can explain what it represents in terms of job-relevant capability.
That opacity has a measurable cost on the candidate side. A survey of 1,066 U.S. job seekers conducted by Enhancv in April 2026 found that 68.5% were never told AI played any role in their evaluation, and only 9.7% said an employer had clearly disclosed it.
Nearly a third said they had walked away from a role rather than complete a one-way AI screening, and the pattern wasn't evenly distributed. Almost 80% of the abandoned roles paid under $100,000. The candidates with the fewest options are bearing the most exposure to systems they can't see and didn't consent to.
A Black Box With a Vendor Agreement
The accountability issue isn't entirely about opacity, though opacity is part of it. Many screening platforms operate as black-box or near-black-box systems, where the model's weighting logic isn't visible to the employer, let alone the candidate.
Even when vendors provide explainability features, the explanation a recruiter sees may be a simplified proxy for a more complex underlying process. An HR leader can read the output. They often can't interrogate how it was produced.
This matters more at scale. A company processing tens of thousands of applications a year is making a large number of consequential decisions through a system its own team may not be equipped to audit. The Colorado law's notification and appeal requirements force that question into the open. If a candidate asks why they were rejected, who answers, and with what information?
Teppoeva's work focuses specifically on what she describes as "the interpretation layer", the space between an AI-generated hiring signal and a human decision. The first question she asks organizations that have already deployed a screening tool is why: "Why did this candidate receive a low score despite looking strong to the hiring team? Why did another candidate score high but later end up on a performance improvement plan?"
Most organizations, she says, can't answer those questions. They have the score but not the reasoning behind it.
Organizations frequently rely on vendor-generated outputs without maintaining a clear internal framework for interpreting or documenting those outputs," she says. "This can make it difficult to explain how human judgment was applied or how a final decision was reached.
When a decision is challenged, that gap becomes a liability.
Liability Landed With the Employer
The regulatory timing also exposes a problem with how AI hiring tools have typically been procured. Decisions about which platform to use have often been driven by talent acquisition teams focused on throughput, with legal, compliance, and sometimes HR leadership less involved in evaluating what the tool is actually doing.
Colorado's law shifts that calculus. Under SB 205, liability for adverse outcomes from high-risk AI in employment sits with the employer, not the vendor. The company using the tool is accountable for its results.
Jimmy Hurff, COO and co-founder of Brightmove, an applicant tracking system vendor, says the procurement conversations he sees reflect exactly this problem.
Employers seem most focused with integration of features and with candidate volumes. Our prospective customers are mostly checking for integration features, but rarely focus on the quality of the candidate pipeline and the automation of scoring and assessment.
The questions that matter most under Colorado's framework, explainability, auditability, appeal processes, aren't typically the ones being asked at the table.
Brightmove published a Responsible AI Policy as a direct response to SB 205, and Hurff says the law has forced a broader reckoning across the vendor landscape. He also makes clear where he thinks the regulatory conversation needs to go next, toward a single federal standard rather than a state-by-state compliance patchwork, which he argues would give vendors a clearer mandate to build against.
The documentation problem becomes most consequential when candidates whose communication patterns fall outside what the system was trained on, due to disability, neurodivergence, or language background, are evaluated by tools that can't account for context. Teppoeva describes the core problem.
"A signal can be measured accurately and still interpreted incorrectly," she says.
When a candidate challenges an outcome, most organizations can reproduce the output. They can't explain the decision, and they rarely have documentation showing that what was measured was actually relevant to the role.
The EEOC has been watching this space since at least 2021, when it launched its initiative on AI and algorithmic fairness in hiring. Its 2023 guidance made clear that employers can't outsource liability for discriminatory screening outcomes to a third-party vendor.
Colorado's law goes further, adding transparency and appeals obligations that most enterprise hiring operations weren't built to fulfill.
What HR Leaders Need to Know
What this creates, practically, is a compliance surface that many HR leaders haven't mapped. They may have a vendor agreement. They may have a general sense of what the tool does. But documentation of how the model was trained, what it's optimizing for, and how decisions are logged and retrievable for appeal review is often thinner than the regulatory moment now demands.
Teppoeva frames the shift in terms that HR leaders will recognize from other compliance conversations.
Emerging regulations are shifting the conversation from 'Does the tool work?' to 'Can we explain and defend how it was used?'
Those treating AI screening as a procurement decision rather than an ongoing governance one are most exposed when a candidate or regulator asks a question the tool can't answer.
Some of this will sort itself out through vendor pressure. If enterprise customers start requiring explainability documentation and audit trails as part of procurement, vendors will build them. But that process moves slowly, and in the meantime, candidates in Colorado and elsewhere are being filtered by systems whose logic the employers operating them can't fully explain.
The conversation HR leaders need to be having isn't only about whether AI screening tools work. It's about whether the organizations using them can account for what they do.
