AI systems exhibit what researchers call AI systematic biases judging people—a phenomenon that differs fundamentally from how humans form trust and make decisions about others. A new study from Hebrew University of Jerusalem, released in April 2026, analyzed over 43,000 simulated AI decisions compared to around 1,000 human participant decisions, revealing that advanced AI models similar to ChatGPT and Google’s Gemini judge individuals through rigid, spreadsheet-like scoring rather than the messy, intuitive reasoning humans employ.
Key Takeaways
- Hebrew University researchers tested AI models on 43,000+ decisions versus ~1,000 human decisions across financial and trust scenarios.
- AI judges people by scoring separate traits—competence, integrity, kindness—creating systematic, predictable outcomes lacking human nuance.
- Demographic biases in AI (age, religion, gender) are stronger and more consistent than human biases across different models.
- Different AI models show varying biases despite surface similarity, suggesting no universal “AI opinion” on trustworthiness.
- AI mimics human trust structure but with extreme consistency that amplifies rather than moderates demographic prejudice.
How AI Judges People Differently Than Humans
The fundamental difference lies in architecture, not intent. “People in our study are messy and holistic in how they judge others,” explains Valeria Lerman, co-author and PhD candidate at Hebrew University. “AI is cleaner, more systematic and that can lead to very different outcomes”. Humans weigh competing signals—a person’s appearance, tone, context, contradictions—and arrive at intuitive conclusions that shift based on subtle cues. AI does something else entirely: it extracts discrete features (competence, integrity, kindness) and scores them independently, like entries in a spreadsheet column. This approach produces consistent logic but sacrifices nuance.
The research tested AI systems on high-stakes scenarios: lending or donating money to small businesses, trusting a babysitter, and rating a boss. In each case, AI’s judgments proved more extreme and predictable than humans’ judgments. Where a human might hedge—”this person seems trustworthy but I have reservations”—an AI model commits fully to its scoring logic, amplifying whatever biases exist in its training or reasoning process.
Why AI Biases Are More Predictable and Stronger
“Humans have biases, of course,” says Prof. Yaniv Dover, the study’s lead researcher. “But what surprised us is that AI’s biases can be more systematic, more predictable, and sometimes stronger”. The study documented clear patterns: older individuals were sometimes favored in financial scenarios, religion significantly affected monetary outcomes, and gender influenced certain models’ decisions across different scenarios. The troubling part is consistency—these biases did not vary randomly. They appeared reliably across test iterations, suggesting they stem from the AI’s core decision-making logic rather than noise or random variation.
Consider the difference between human and AI bias. A human loan officer might unconsciously favor applicants from their own demographic group, but that bias fluctuates based on mood, context, and competing information. An AI model, by contrast, applies the same demographic weighting every single time it encounters the same input. This predictability makes AI bias both easier to detect and harder to excuse—it is not a human failing but a systematic feature of how the system processes information.
AI Systematic Biases Judging People: No Universal AI Opinion
One finding complicates simple narratives about “AI bias” as a monolithic problem: different AI models showed significantly different biases despite architectural similarities. “Two systems can look similar on the surface but behave very differently when making decisions about people,” Dr. Lerman said. ChatGPT-like models and Gemini-like models did not judge the same person identically. This suggests that bias is not inherent to AI reasoning itself but rather emerges from training data, model weights, and design choices specific to each system. There is no single “AI opinion” on trustworthiness—only competing, predictable opinions shaped by how each model was built.
This fragmentation matters for accountability. If all AI systems judged identically, regulators could theoretically fix the problem at the architecture level. Instead, each model requires separate scrutiny. A lending AI trained on historical loan data might systematically underestimate women’s creditworthiness, while a hiring AI trained on different data might show the opposite bias. The commonality is not the direction of bias but its rigidity—once the bias is baked in, it applies uniformly to every decision.
What This Means for High-Stakes AI Decisions
The implications extend far beyond academic interest. AI systems increasingly decide real outcomes: loan approvals, job interviews, bail recommendations, and trust assessments. If these systems judge people through predictable, amplified demographic biases, the consequences compound. A person denied a loan by an AI system faces not a single human’s bias but a scaled, systematic application of that bias across thousands of decisions. The Hebrew University study does not quantify the practical harm, but it establishes the mechanism: AI does not neutralize human bias; it systematizes and strengthens it.
“They can model aspects of human reasoning in a consistent way. But they are not human and we shouldn’t assume they see people the way we do,” researchers note. This is the core insight. AI mimics human trust logic superficially—it scores the same traits humans care about—but it does so through rigid rules that lack the flexibility, contradiction, and context-sensitivity of actual human judgment. Treating AI as a neutral arbiter of trustworthiness is a dangerous misunderstanding of what these systems actually do.
Can AI Bias Be Fixed?
The study does not offer solutions, only diagnosis. Researchers analyzed the problem but did not test interventions. However, the findings suggest that bias mitigation requires more than removing demographic data from AI inputs. If AI judges people through systematic, rule-based logic, the bias may persist even when explicit demographic markers are absent—encoded indirectly through proxies like postal code, education history, or employment gaps that correlate with protected characteristics. Fixing AI systematic biases judging people likely requires rethinking how AI systems are trained, audited, and deployed in high-stakes contexts.
Does AI bias affect all models equally?
No. The Hebrew University study found that different AI models—ChatGPT-like and Gemini-like systems—showed varying biases despite surface similarities. This means bias is not universal to AI reasoning but specific to how each model was trained and designed. Auditing and fixing bias requires testing individual systems rather than assuming all AI exhibits the same problems.
Why is AI bias more predictable than human bias?
AI judges people using consistent, rule-based scoring of discrete traits (competence, integrity, kindness), while humans employ messy, holistic intuition that shifts based on context. This consistency makes AI bias reproducible and systematic—the same person receives the same judgment every time, whereas a human’s judgment might vary based on mood or additional information.
What scenarios did the Hebrew University study test?
Researchers tested AI and human judgment on financial decisions (lending or donating money to small businesses), trusting a babysitter, and rating a boss. These high-stakes scenarios revealed how AI and humans differ when assessing trustworthiness and competence in contexts where judgment errors carry real consequences.
The Hebrew University study does not solve the problem of AI bias, but it reframes how we should think about it. AI is not a neutral technology that happens to absorb human prejudices; it is a system that translates human reasoning into rigid logic, often amplifying the very biases it was meant to overcome. Until organizations deploying AI in lending, hiring, and other consequential decisions reckon with this reality, the systems they trust will continue making predictable, systematic mistakes—and those mistakes will fall hardest on the people already most vulnerable to bias.
This article was written with AI assistance and editorially reviewed.
Source: TechRadar
