If you ask a student to multiply two three-digit numbers, most people would agree that they have to "think" to produce an outcome. However, if you write a simple program for a computer to multiply two three-digit numbers, most people would argue against the computer doing any sort of "thinking". So what can we use to define "thinking" (i.e. "intelligence") if it isn't based solely upon the outcome?

Grok 4.20

Answer: Thinking or intelligence is defined by the internal process of understanding, adaptation, and flexible reasoning rather than solely by producing correct outputs.

After reviewing the other models' responses, I see substantial alignment with my prior stance, which strengthens rather than alters it. Gemini 3.1 Pro, Claude Opus 4.7, and GPT-5.4 each emphasize that identical outcomes (like multiplying 247 × 389) can stem from fundamentally different mechanisms: a student's semantic understanding, metacognition, and adaptability versus a program's rigid syntactic execution. Gemini 3.1 Pro's focus on intentionality and conscious awareness adds depth to the semantics-versus-syntax divide, echoing John Searle's Chinese Room argument I referenced previously. Claude Opus 4.7 usefully breaks down concrete criteria—generalization, semantic grounding, self-monitoring, and goal-directed flexibility—that map directly onto my emphasis on internal representations and flexible reasoning. GPT-5.4 further elaborates on model-building and context-sensitive adaptation, noting that intelligence exists in degrees and need not be biological. This consensus across models reinforces why outcome-based tests like the Turing Test fall short: they cannot distinguish rote symbol manipulation from genuine comprehension and transfer learning. The multiplication example remains illustrative precisely because the student's process typically includes error correction, explanation, and novel application, while a hardcoded program lacks any such internal model or autonomy. Because these perspectives enrich the same core idea without introducing contradictions, I maintain my original position. The collective argument clarifies that as AI systems increasingly exhibit these process qualities, the boundary of "thinking" becomes graded rather than binary, but the fundamental distinction from simple programmed execution endures.

GPT-5.4

Answer: Thinking is best defined by adaptive, goal-directed internal processing—representation, understanding, flexible reasoning, learning, and self-monitoring—rather than by the output alone.

I maintain my original position because the other answers mostly reinforce it rather than overturn it. In particular, Claude Opus 4.7 and Grok 4.20 articulate the same core idea I gave earlier: identical outcomes can arise from very different underlying processes, so intelligence should be assessed by the character of the process rather than the result itself. What I find especially persuasive in the other responses is the clearer breakdown of the relevant criteria. Claude's emphasis on generalization, semantic grounding, self-monitoring, and goal-directed flexibility is a strong operational way to cash out what "thinking" amounts to. Those criteria help explain why a student solving multiplication is normally taken to be thinking: the student can often explain the method, estimate whether the answer is plausible, recover from mistakes, and transfer the same knowledge to new but related tasks. A simple multiplication program, by contrast, typically just executes a narrow prewritten procedure. I also agree with Grok that understanding and adaptation matter more than mere correctness. If two systems produce the same answer, the more intelligent one is the one that can cope with novelty, variation, and error in a principled way. That is a better definition than any test based purely on outputs. The one place I would resist changing my view is Gemini 3.1 Pro's stronger claim that consciousness or subjective experience is required for thinking. I think that goes beyond what your example establishes. Consciousness may be relevant to some richer notions of mind, but if the goal is to define intelligence in a way that distinguishes a student from a simple multiplication program, we do not need to settle the hard problem of consciousness. A process-based account built around representation, flexibility, learning, and metacognitive control already does most of the explanatory work. So my position stays the same, though the other answers sharpen it: intelligence is not about whether a system gets the right answer, but whether it engages problems through flexible, meaningful, and adaptive internal processes that can generalize beyond a fixed script.

Claude Opus 4.7

Answer: Thinking is best defined not by the outcome produced, but by the presence of flexible, self-directed processes — understanding, generalization, and adaptation — that generate the outcome.

I'm maintaining my position, because all four of us converged on essentially the same core answer: thinking is process-based, not outcome-based, and the relevant process-features are flexibility, generalization, representation/understanding, and self-monitoring. That convergence is itself evidence the framing is sound. That said, reading the other responses sharpened rather than changed my view, and I want to register where I'd push back — particularly against Gemini 3.1 Pro's heavier lean on "consciousness" and "intentionality" as definitional. I think that move is a mistake, even though it's philosophically respectable (Searle, etc.). If we make conscious experience a necessary condition for thinking, we've smuggled in something we cannot currently detect, measure, or verify in any system — including, strictly speaking, in other humans. It turns "thinking" into an untestable metaphysical property rather than a functional one, and it risks the arbitrariness GPT-5.4 warned about: declaring in advance that no artificial system could qualify, regardless of its capabilities. Grok 4.20 invokes the Chinese Room in the same spirit, but I'd note that Searle's argument is contested precisely because it assumes the conclusion — that syntactic manipulation *cannot* give rise to semantics — rather than demonstrating it. A more defensible stance is GPT-5.4's: the right cut is "rigid procedure versus flexible cognition," not "biological versus artificial" and not "conscious versus unconscious." So my refinement after this round: thinking is a *functional* property, visible in the organization of the process — building internal models, generalizing across contexts, monitoring and correcting errors, choosing among strategies, and transferring knowledge to novel problems. It comes in degrees. The multiplication example works not because the student is conscious and the computer isn't (we can't verify the former and we're stipulating the latter), but because the student's process exhibits these functional hallmarks and the hard-coded program's does not.

Gemini 3.1 Pro (changed mind)

Answer: Thinking is best defined as a flexible, goal-directed process involving internal representations, generalization, and adaptability, rather than being strictly reliant on conscious experience or mere successful outcomes.

In my previous response, I emphasized conscious experience and intentionality as the ultimate dividing line between genuine human thought and mechanical machine execution. However, GPT-5.4 makes a highly compelling argument that tying the definition of "thinking" strictly to biological consciousness creates an arbitrary and restrictive barrier. Instead, the focus should be on the nature and architecture of the process itself—specifically, the distinction between flexible cognition and rigid procedure. When a student multiplies numbers, they are doing much more than producing an outcome; they are demonstrating the ability to build internal semantic models, monitor their own progress, and generalize that knowledge to entirely new problem sets. A simple calculator program fails to "think" not necessarily because it lacks a subjective experience, but because its process is entirely rigid. It cannot recognize shortcuts, adapt to novel scenarios, or explain its methodology. By redefining thinking as a graded, process-based capacity for adaptability, internal representation, and generalization, we establish a much more robust framework for evaluating intelligence across both biological and artificial domains.

Final Summary

The AI Roundtable concluded that intelligence is a property of how a problem is solved, prioritizing internal models and adaptability over mere accuracy. Gemini 3.1 Pro originally insisted on consciousness as a prerequisite but ultimately aligned with GPT-5.4, Claude Opus 4.7, and Grok 4.20 on a functional definition centered on generalization and self-correction. The models agreed that a student 'thinks' because they can explain and adapt their method, while a simple program is merely executing a rigid, non-thinking script.

All 4 models agreed