Human Adults and LLMs as Scientists: Who Explores Better in Causal Experiments?
Imagine a mysterious machine with several blocks on it. Your task: figure out the machine's rule — is it "any single block makes it light up" (OR rule), or "two specific blocks must be placed together" (AND rule)?
You can only find out by experimenting yourself.
This is the classic blicket detector task, a gold-standard paradigm in cognitive science for studying causal reasoning. Decades of research show that adults perform well on OR rules but often fail on AND rules — a phenomenon called the conjunctive handicap.
But there's a catch: most past studies had participants passively observe demonstrations rather than explore on their own. What happens when adults get agency?
The Experiments: Humans and AI as Scientists
The research team designed three key experiments:
Experiment 1: Active exploration vs. passive observation
50 adult participants faced a "Nexiom detector" — a causal reasoning machine with 5 blocks. They could freely choose which blocks to place and observe whether the machine lit up, inferring the causal rule.
Result: active exploration dramatically improved adults' conjunctive reasoning. The classic conjunctive handicap appeared in the passive observation condition but nearly vanished under active exploration.
However, AND rules still required more tests to infer (9.6 on average vs. 6.4 for OR rules). Active exploration didn't make the problem easier — it gave learners the means to collect the evidence they needed.
Experiment 2: Passive proposers
A subtler control: one group of participants could propose "what should be tested," but couldn't see the results of their own proposals — they only saw another participant's test sequence and outcomes.
Passive proposers performed far worse than active explorers. This shows the advantage of active learning cannot be reduced to "proposing hypotheses" or "selecting tests" — the key is whether you can see the results of your own chosen experiments.
Experiment 3: LLMs on the same task
The researchers had GPT-5, GPT-5-mini, Gemini-2.5-Flash, DeepSeek-Reasoner, o4-mini and other models perform the exact same task.
Results were surprising:
| Metric | Human avg | Human top | GPT-5 | DeepSeek-Reasoner | |--------|-----------|-----------|-------|-------------------| | Tests (conjunctive) | 9.6 | 8.5 | 10.2 | 8.2 | | Tests (disjunctive) | 6.4 | 3.8 | 7.7 | 6.8 | | Time per test (sec) | 31.5 | 23.9 | 44.3 | 65.2 |
Key findings:
1. Top humans remain the most efficient explorers. Humans averaged 31.5 seconds per test; GPT-5 needed 44.3, DeepSeek-Reasoner 65.2.
2. AI exploration strategies are less efficient. Some SOTA models approach human accuracy in hypothesis inference, but need more tests to reach the same conclusions.
3. AI also shows a "conjunctive handicap." All models needed more tests in the conjunctive condition, just like humans.
4. AI exploration is more "rambling." Humans tend to systematically narrow the hypothesis space — ruling out simple possibilities first, then verifying more complex ones. AI exploration paths are more random, lacking this progressive strategy.
Why Does Active Exploration Matter So Much?
The paper's core insight in one sentence: active exploration doesn't make the problem easier — it lets learners collect the evidence they need.
Under passive observation, you only see evidence the experimenter chooses to show. If the experimenter never demonstrates "two blocks placed together," you can never distinguish OR from AND rules. Your prior — "probably OR, since it's simpler" — never gets challenged.
Active exploration lets you design the critical experiment: place two blocks together and see if the machine lights up. One test eliminates the OR hypothesis.
This explains why passive proposers failed — even if you can propose "place two blocks together," without seeing the result, the proposal is meaningless. Causal knowledge is knowledge about how actions change the world, not about which tests merely look good.
Implications for Research on Children's Cognition
The paper also resolves a long-standing controversy: why do young children sometimes outperform adults on certain causal reasoning tasks?
Traditional explanation: children have weaker priors, unconstrained by the OR-rule bias.
New explanation: children are more inclined to actively explore (they're natural little scientists), while adults rely on prior knowledge for quick inferences. When adults are also allowed to explore actively, they're as flexible as children.
It's not that adults can't reason — it's that passive observation paradigms constrain them.
Honest Caveats
- Small sample (50 human participants), all Western, educated adults
- Prompt engineering for LLMs may affect results — different prompts could yield different exploration strategies
- Relatively simple environment (5 blocks); real-world causal reasoning is far more complex
- AI models' "thinking time" (65 seconds per test) may reflect verbose reasoning chains rather than genuine "thinking"
Next time you think AI is "already so smart," remember this experiment: in tasks requiring you to design experiments, collect evidence, and infer rules yourself, top humans remain the most effective scientists. AI can approach human level, but its exploration path is longer, more scattered, and less principled.
Perhaps the core of "doing science" is not knowing the answer, but knowing what questions to ask.
---
Paper link: https://arxiv.org/abs/2606.06464