AI Hiring: Bias Is Not Just a Mirror, But Also a Furnace

[!quote]

A recent study by MIT Technology Review reveals a counterintuitive fact: AI in hiring not only replicates existing human biases but also "invents" new ones. This is more dangerous than human HR.

Anyone involved in AI investing knows that the market's valuation logic for AI hiring rests on two assumptions: efficiency gains and improved fairness. We used to believe that with enough data training, AI could eliminate subjective human biases. But this study shakes the second assumption. It tells us that AI is not a clean mirror, but a furnace—it smelts out new biases, and humans cannot even predict where these biases come from.

Comparison: Rule-Based AI vs. Generative AI

From a technical perspective, there are two distinct approaches to AI hiring:

Option A: Traditional Rule-Based AI

• Based on algorithms like keyword matching, logistic regression, and decision trees
• Feature engineering is manually defined, making bias sources relatively transparent
• Example: When screening resumes, weights for fields like education and years of experience are set by engineers
• Risk: Designer bias, but it is auditable and correctable

Option B: Large Language Model (LLM)-Driven AI

• Based on models like GPT and Claude, conducting open-ended interview dialogues and resume evaluations
• Training data comes from the internet, containing massive amounts of unlabeled implicit biases
• The study points out: LLMs "create" new biases from unrelated contexts, such as associating "worked at a coffee shop" with "lack of responsibility"
• Risk: Uncontrollable emergent biases, with audit difficulty increasing exponentially

Let's look at some data: The MIT team tested three mainstream hiring AIs and found that generative AI gave negative scores to specific ethnic groups 23% higher than rule-based AI, and these biases did not exist in the training data. This means that even if you clean all known discriminatory data, the model may still "invent" new discrimination standards on its own.

Industry Impact: Valuation Repair Logic Needs Re-evaluation

Currently, the pricing of the AI hiring sector in the secondary market implies an expectation that "the more advanced the technology, the fewer the biases." But this study tells us this might be wrong. More advanced models (LLMs) are actually more prone to generating unknown biases, which brings about:

1. Increased Regulatory Risk: The EU AI Act has already listed hiring as a high-risk application, requiring algorithm audits. Now, the difficulty of auditing is higher, and costs may double.

2. Decreased Corporate Procurement Willingness: HR departments are afraid to use "black box" models for decision-making, especially regarding hiring. Recently, a major tech company paused its procurement of AI interviewers.

3. Divergence in Technical Routes: Rule-based AI may regain favor among enterprises; although less effective, it is at least controllable. Pure LLM solutions need to add "anti-bias filters" and human review steps, weakening their efficiency advantage.

Looking at the competitive landscape, companies with "Explainable AI" capabilities will benefit. For example, in the semiconductor field, we are watching companies that accelerate audit algorithms with AI chips; at the application layer, SaaS vendors that can provide "bias heatmaps" and "decision path backtracking" will gain a premium.

Trend Judgment: Short-term Pressure, Long-term Benefit for Governance

In the short term, the AI hiring sector will experience a valuation correction. The market needs to digest the realization that "AI can also create biases." Investors will re-examine hiring unicorns that rely on LLMs; their user growth may slow down.

But in the long run, this is actually a good medicine. It forces the industry to establish standardized bias detection frameworks, similar to risk control models in the financial industry. Once regulations are clear, companies with compliance capabilities can widen the gap through a "trust premium." This is similar to how, after GDPR was introduced in 2018, companies with good data governance actually gained more customers.

From the perspective of valuation repair logic, we currently need to wait for two signals: first, the US EEOC (Equal Employment Opportunity Commission) issuing AI hiring audit guidelines; second, leading enterprises launching deployable bias detection tools. These two things will likely be completed within 12 months.

Only then will AI hiring truly enter an "efficient market."

Original link: https://www.technologyreview.com/2026/07/20/1140655/ai-biases-hiring-humans/