Humble AI in the real-world: the case of algorithmic hiring

Rahul Nair, Inge Vejsbjerg, Elizabeth M. Daly, Christos Varytimidis, Bran Knowles

Abstract

This paper applies the "Humble AI" framework—built on principles of skepticism, curiosity, and commitment—to the high-stakes domain of algorithmic recruitment. Traditional hiring tools often suffer from a "certainty illusion," providing definitive rankings that mask underlying data noise and structural biases. By replacing these rigid point-estimates with "rank sets" that quantify model uncertainty, the authors demonstrate how AI can acknowledge its own limitations. This approach fosters more robust human-in-the-loop decision-making, ensuring that candidates with non-traditional backgrounds are not prematurely excluded by overconfident but statistically fragile models.

Introduction

Algorithmic hiring tools are increasingly utilized to manage massive candidate pools, yet they frequently present a false sense of precision. Most systems provide a definitive ranking (e.g., "Candidate A is the #1 match"), which obscures the inherent noise in labor market data. This "certainty illusion" is particularly dangerous in recruitment, where structural inequities and the complexity of human potential make precise prediction difficult.

When a model expresses 100% certainty about a ranking that is actually a statistical "toss-up," it amplifies bias and risks "misrecognition"—the systematic exclusion of qualified individuals who do not fit a standard profile. The authors argue for a shift toward "Humble AI": systems designed to admit when they lack sufficient information to make a definitive judgment.

Methodology: From Point-Estimates to Rank Sets

The researchers operationalize Humble AI by moving from deterministic scoring to a probabilistic framework known as Rank Sets. The methodology involves four primary stages:

  1. Uncertainty via Perturbation: Recognizing that many hiring models are proprietary "black boxes," the authors employ a technique inspired by LIME. They perform local perturbations on candidate profiles—such as random feature masking—to observe how the scoring function reacts to minor variations in input.
  2. Monte Carlo Simulations: By executing these perturbations repeatedly, the system generates a distribution of scores for each candidate rather than a single, static number.
  3. Probabilistic Ranking: The scores are used to construct a probability matrix where each cell represents the likelihood of a specific candidate occupying a specific rank. From this matrix, three key metrics are derived:
  4. Expected Rank: The weighted average position of a candidate across all simulations.
  5. Rank Entropy: A measure of the "spread" or uncertainty regarding a candidate's position.
  6. Rank Variance: The degree to which a candidate's rank fluctuates across simulations.
  7. UX Design and the "Humble Toggle": To integrate these metrics into a recruiter's workflow, the authors developed a prototype interface. A "Humble AI" toggle allows users to surface candidates with high rank entropy in a separate view, signaling that these individuals require a manual "second look" because the AI's confidence is low.

Experimental Data

The study utilizes two distinct data sources to validate the Humble AI approach:

  • Synthetic Data: A controlled environment where the "true" ranking is known. This allowed the researchers to inject specific levels of noise and measure the accuracy of probabilistic methods versus standard deterministic methods in recovering the ground truth.
  • Real-World HR Data: Anonymized candidate profiles and job specifications from Workable, a major hiring platform. This data covered various sectors, including technical roles (e.g., ReactJS Developer) and administrative roles, testing the feasibility of the approach in live market conditions.

Key Insights and Results

The Fragility of Deterministic Rankings

The research revealed a startling lack of correspondence between deterministic rankings and probabilistic ones. In several real-world job categories, the Jaccard similarity between the two was near zero. This suggests that the "top picks" surfaced by standard AI are often arbitrary artifacts of noise rather than reflections of superior candidate quality.

Robustness of Expected Rank

In synthetic experiments, as noise levels increased, the "Expected Rank" metric proved significantly more robust than single point-estimate scores. While deterministic performance plummeted under noisy conditions, the Humble AI approach maintained a much higher Rank Biased Overlap (RBO), successfully recovering the true best candidates more consistently.

Model Ignorance in the Wild

In real-world trials, the mean entropy across different job titles ranged from 3.0 to 3.7. This high level of entropy indicates substantial "model ignorance," proving that single-score rankings are often statistically indefensible. The research suggests that candidates often exist in clusters of suitability rather than a strict linear hierarchy.

The Attribution Trap

A critical finding from focus groups was the risk of "attribution bias." Recruiters may misinterpret the model's uncertainty as a reflection of the candidate's weakness. If the AI signals it is "unsure" about a person, a recruiter might unfairly label that candidate as "risky" rather than recognizing the model's own lack of sufficient data to make a judgment.

Discussion and Future Work

While the Humble AI framework offers a path toward more ethical and accurate hiring, it introduces new challenges. Surfacing uncertainty increases the cognitive load on recruiters, who are often operating in time-constrained environments. Future research must investigate whether users will meaningfully engage with high-entropy candidates or simply default to the easiest path.

Furthermore, the "ground truth" problem remains a hurdle in recruitment; because we rarely see the performance of candidates who were not hired (the unobserved counterfactual), validating long-term outcomes is difficult. The authors conclude that future AI design should focus on ensuring users attribute uncertainty to the model's limitations, transforming AI from a definitive judge into a humble assistant.

Summary updated

BibTeX
@inproceedings{nair2025humble,
  author = {Rahul Nair and Inge Vejsbjerg and Elizabeth M. Daly and Christos Varytimidis and Bran Knowles},
  title = {{Humble AI in the real-world: the case of algorithmic hiring}},
  booktitle = {CHIWORK (Adjunct)},
  year = {2025},
  doi = {10.1145/3707640.3731919},
}
Figure