AI Agents Inherit Human Biases in Causal Reasoning—Here’s How to Fix It
Language model (LM) agents are increasingly being deployed as autonomous decision-makers, tasked with gathering information and inferring causal relationships in their environments. But a new study reveals a troubling trend: these AI systems exhibit the same cognitive biases as humans when reasoning about cause and effect—particularly a strong preference for simpler "disjunctive" (OR) causal explanations over more complex "conjunctive" (AND) ones, even when evidence suggests otherwise.
The research, published in a preprint paper titled Language Agents Mirror Human Causal Reasoning Biases, adapts the classic "Blicket Test" from developmental psychology to evaluate how LM agents explore and infer causal structures. The findings suggest that LMs, trained on vast amounts of human-generated text, have internalized deep-seated reasoning heuristics that can lead to systematic errors in scientific and business decision-making.
The Blicket Test: A Benchmark for Causal Reasoning
The Blicket Test is a well-established paradigm where an agent must determine which objects ("blickets") activate a machine based on observed interactions. The machine follows either a disjunctive rule (turning on if any blicket is present) or a conjunctive rule (turning on only if all blickets are present). The study converted this into a text-based game where LM agents could place objects on the machine, observe outcomes, and then answer questions about which objects were blickets.
Key Findings
- Disjunctive Bias: Across multiple model families (including GPT-4o, DeepSeek, and Gemma), LMs consistently performed better in disjunctive scenarios than conjunctive ones—even when the evidence equally supported both rules. This bias persisted regardless of model size or prompting strategy.
- Exploration Inefficiency: LM agents struggled to efficiently narrow down hypotheses during exploration. Unlike an optimal "Oracle" baseline, they often failed to eliminate enough possibilities to reach the correct conclusion, especially in more complex environments (e.g., with 8 objects instead of 4).
- Human-Like Reasoning: Strikingly, LMs mirrored the reasoning patterns of human adults more than children. Like adults, they defaulted to disjunctive explanations even when conjunctive evidence was stronger. Children, by contrast, are more flexible and "scientific" in their approach.
Why This Matters for Business
Autonomous AI agents are increasingly used in high-stakes domains like drug discovery, financial forecasting, and operational decision-making. If these systems inherit human biases, they may:
- Overlook complex causal relationships (e.g., assuming a single factor drives outcomes when multiple are needed).
- Make suboptimal exploration choices (e.g., failing to test critical combinations of variables).
- Reinforce existing biases in data-driven decision-making.
A Fix: Hypothesis Sampling
The researchers proposed a novel hypothesis sampling method to mitigate these biases. By explicitly prompting LMs to generate and eliminate hypotheses during inference, performance improved significantly—reducing the disjunctive bias and bringing LM reasoning closer to optimal scientific inquiry.
The Bottom Line
As AI agents take on more autonomous roles, understanding and correcting their inherited biases will be crucial. This study highlights both the risks of human-like reasoning in machines and a promising path forward. For businesses deploying AI, the lesson is clear: don’t assume your AI thinks like a scientist—yet. But with the right techniques, it might learn to.