The premise behind AI hiring tools was straightforward: automate the screening bottleneck, reduce human bias and surface better candidates faster. New research suggests the technology may be delivering on none of those promises.
The “Slopification” Problem in AI-Powered Hiring
Researchers Matthias Holweg, director of the Oxford Artificial Intelligence Program, and Thomas Davenport, a professor at Babson College, published findings in the Harvard Business Review this month warning that AI embedded across business processes is producing what they call “knowledge decay.” In hiring, where AI now touches everything from job description drafting to resume screening to interview scheduling, this decay has a specific shape.
When AI tools are embedded at every stage of hiring, Holweg and Davenport found that the combined effect is a more opaque process. They call the root mechanism “slopification”: the tendency of individuals to use AI to produce polished-looking, low-quality outputs and for others downstream in the same process chain to stop checking those outputs with appropriate scrutiny.
The cascading result, which the researchers compare to “a risky AI-based game of telephone,” is knowledge entropy. Every time a candidate’s profile, a job description or a screening result passes through an AI tool, it drifts further from what was originally true. “The overall impact of AI augmenting each step is that it has sunk trust in the process to all-time lows for both job seekers and recruiters,” the authors wrote.
Three Failure Modes That Matter for HR Leaders
Holweg and Davenport identify three distinct failure modes that compound across a single hiring cycle. Knowledge verification is the first: double-checking AI outputs thoroughly often erases the efficiency gains AI was supposed to deliver, which is why many HR teams end up scheduling in-person, AI-free interviews just to sort out who is actually qualified. In that case AI adds a step to the process.
Knowledge validation is the second failure mode. Human experts now face the task of justifying both the quality of work submitted by candidates and whether actual human intellectual effort produced it at all. This creates friction that does not exist in traditional hiring workflows.
The third failure mode, knowledge entropy, underlies both. Each pass through an AI tool introduces drift. A resume written with AI assistance, screened by AI tools, scored by an AI system and summarized for a hiring manager has passed through multiple layers of transformation. What the hiring manager sees may have very little relation to the actual candidate behind it.
AI Can Pass Its Bias to the People Using It
When AI screening tools are set to recognize the signals from past successful hires, they replicate the familiarity patterns of whoever was already being hired, which amplifies bias at scale.
A 2025 University of Washington study found that 528 people who worked with simulated AI tools to pick candidates for 16 jobs tended to adopt the racial biases of the AI’s recommendations, meaning the problem does not end when a recruiter takes over from the algorithm. The bias transfers.
What This Means for the HR Leader
For CHROs and talent acquisition leaders, these two research threads arrive at the same operational conclusion: AI tools embedded in hiring require active governance. A vendor certification of compliance is not sufficient. The AI tool changes the quality of inputs the hiring manager sees, which means the tool is part of the hiring decision even when the manager believes they are making that decision independently.
The Workday discrimination case (Mobley v. Workday), in which a federal judge has refused to dismiss California employment law claims and allowed ADA claims to proceed, shows that the legal risk from AI hiring tools falls on employers as well as vendors. HR leaders cannot rely on vendor accountability as a substitute for internal audit. (For a full breakdown of that ruling’s implications, see our earlier analysis of the new liability standard for AI screening vendors.)
The knowledge decay research adds a different dimension: even when AI tools are legally compliant, they may be degrading the information quality that hiring decisions are based on. A process that produces polished outputs but less accurate assessments is a more expensive version of unassisted human judgment with a more convincing appearance.
How to Evaluate Your AI Hiring Stack Now
Holweg and Davenport recommend starting with input structure rather than tool restrictions. Instead of open-format resumes, which AI tools can polish beyond recognition, they suggest requiring structured questionnaires with specific, verifiable inputs such as projects led, budgets managed and team sizes supervised. These are harder to fabricate with AI assistance and easier to assess against actual job requirements.
The practical upshot for talent acquisition teams is that AI tools need to be evaluated on what the hiring outcomes look like over time. If the system is tracking past success signals and using them to screen future candidates, it is reproducing the past. The efficiency case for AI hiring is real, but the research emerging from Oxford, Babson and the University of Washington suggests that HR leaders who deploy these tools without active governance over inputs, interpretation and outcome auditing are compounding risk rather than reducing it.
Sources: Harvard Business Review; University of Washington; HR Executive; HR Executive
