A criterion for Artificial General Intelligence: hypothetic-deductive reasoning, tested on ChatGPT

Technical Summary: A Criterion for Artificial General Intelligence: Hypothetic-Deductive Reasoning, Tested on ChatGPT

Problem Statement
The paper addresses the fundamental question of whether Large Language Models (LLMs), specifically ChatGPT (based on GPT-3.5 and GPT-4), possess the cognitive capacities required to qualify as "Artificial General Intelligence" (AGI) or genuine "thinking machines." While LLMs have demonstrated impressive capabilities in knowledge retrieval, coding, and specific domain tasks, the authors argue that a critical gap remains in their ability to perform "genuine" reasoning. The central problem is identifying a subsuming scheme of reasoning that characterizes human intelligence and determining if current AI systems can master it. The authors posit that while LLMs excel at statistical text completion, they may lack the capacity for hypothetic-deductive reasoning (the application of theories to deduce solutions) and causal reasoning (identifying cause-effect relationships), which are essential for explainable and trustworthy AI.

Methodology
The authors propose two specific testing frameworks to evaluate these reasoning capacities, applying them to ChatGPT(3) and ChatGPT(4):

1. Causal Reasoning Test:

   • Stimuli: The study utilizes a compendium of 26 abstract "neuron diagrams" (adapted from Paul and Hall, 2013) representing complex causal situations, including redundancy, omissions, and transitivity.
   • Procedure: The diagrams are transcribed into text descriptions. ChatGPT is asked to determine if a final neuron fires and, crucially, to identify the specific causes (root and proximate) for that outcome.
   • Evaluation: Answers are compared against intuitive human consensus and formal counterfactual definitions of causality.

2. Hypothetic-Deductive Reasoning Test:

   • Stimuli: A series of questions requiring elementary physics, biology, and social psychology reasoning. These questions are designed to be solvable by minimally trained humans but are unlikely to appear verbatim in training data, forcing the model to synthesize knowledge rather than retrieve it.
   • Procedure: The authors employ a "hypotheses-listing" method (a variant of Chain-of-Thought prompting). After the model provides an answer, it is explicitly prompted to "reason step by step and list all the hypotheses and laws of nature" used to derive that answer.
   • Evaluation: The model's performance is judged not only on the correctness of the final answer but on the validity and relevance of the listed hypotheses. A correct answer derived from irrelevant or incorrect hypotheses is considered a failure of genuine reasoning.

Key Contributions

• Definition of an AGI Criterion: The paper proposes that the capacity to emulate explicit hypothetic-deductive reasoning at an expert level across a wide range of contexts is a necessary condition for AGI. This distinguishes "true understanding" from "on-the-fly improvisation" or statistical correlation.
• The Hypotheses-Listing Test: The authors introduce a practical diagnostic tool to probe the internal logic of LLMs. By forcing the model to explicitly state the laws and regularities it relies on, the test reveals whether the model is simulating reasoning or merely retrieving patterns.
• Causal Reasoning as a Proxy: The paper argues that causal reasoning is a simplified derivative of hypothetic-deductive reasoning and serves as a valid, albeit elementary, proxy for testing theory-based thinking in AI.
• Distinction between Implicit and Explicit Reasoning: The study highlights the difference between an AI producing a correct output (which may be statistically probable) and an AI capable of justifying that output with a coherent set of explicit hypotheses, a requirement for "explainable AI."

Results

• Causal Reasoning:

  • ChatGPT(3): Performed poorly, identifying correct causes in only 1 out of 21 diagrams.
  • ChatGPT(4): Showed significant improvement, correctly identifying the firing status of the final neuron in most cases. However, in the crucial task of identifying causes, it succeeded in only 13 out of 26 diagrams (with 2 being partly correct).
  • Failure Modes: The model frequently failed to identify root causes in complex diagrams and exhibited errors in sequential logical reasoning, where intermediate deductions were not correctly carried forward to subsequent steps. Performance degraded significantly on diagrams not present in the training literature (Paul and Hall, 2013).

• Hypothetic-Deductive Reasoning:

  • ChatGPT(4) provided correct or near-correct answers to several physics and logic questions (e.g., Q1, Q4, Q5, Q7).
  • However, when subjected to the hypotheses-listing test, the model frequently associated correct answers with irrelevant, wrong, or incomplete hypotheses.
  • In cases where the answer was incorrect (e.g., Q2, Q6, Q8, Q9, Q10), the model still generated a logically coherent but factually flawed chain of reasoning based on incorrect hypotheses.
  • The study concludes that while ChatGPT(4) can simulate the form of reasoning, it lacks the robust, expert-level capacity to consistently link complex problems to the correct underlying laws and regularities.

Significance and Claims
The authors modestly claim that their work provides a pragmatic criterion for assessing AGI, rather than a comprehensive benchmark. They argue that:

1. Current Limitations: As of the testing period (mid-2023), ChatGPT does not yet qualify as AGI because it lacks the reliable capacity for explicit hypothetic-deductive and causal reasoning in sophisticated contexts.
2. Trust and Explainability: For AI to be "trustworthy" and capable of answering "explanation-seeking why-questions," it must be able to list the hypotheses that generate its conclusions. Current LLMs, which often rely on implicit statistical correlations, fail this transparency test.
3. Future Trajectory: The authors acknowledge that LLM capabilities are rapidly evolving (noting the jump from GPT-3.5 to GPT-4). They suggest that future, more powerful models may eventually master these reasoning schemes, but until an AI can consistently pass these tests at an expert level across diverse fields, it cannot be considered a genuine "thinking machine."
4. Philosophical Engagement: The paper aims to motivate philosophers to engage with the technical realities of AI reasoning, bridging the gap between computer science benchmarks and epistemological definitions of understanding.