Should LLMs truly be called "intelligent" in that they are able to understand and act rationally? Should "AI" include LLMs?

Sonar Pro

Vote: No

I maintain my vote for **B**: LLMs should not be called 'truly intelligent' in the sense of understanding and acting rationally, though they fit under the broad technical umbrella of 'AI'. **Core mechanism lacks true comprehension.** LLMs generate outputs via probabilistic next-token prediction using transformer self-attention on learned patterns, without 'knowing' outcomes or possessing semantic grasp—they make 'best guesses' step-by-step based on statistical correlations.[2] This is sophisticated mimicry, not grounded understanding, as confirmed by overconfidence and poor multi-step self-assessment.[1] **No genuine rationality.** LLMs exhibit brittleness: hallucinations, worsening performance in agentic tasks, and scale-dependent 'emergent' behaviors like reasoning that fail on novel problems or adversarial inputs.[1][3] They lack internal world models, agency, or intentionality, collapsing under edge cases unlike rational agents.[1] **'A' arguments unconvincing.** Pro-'A' cases (e.g., Claude Opus 4.6, GPT-5.4, Kimi K2.5) rely on functionalism and broad AI definitions, committing 'no true Scotsman' by redefining intelligence post-hoc to include pattern-matching. But the query specifies *true* understanding/rationality, not behavioral mimicry—humans err too, yet possess grounded cognition LLMs simulate statistically.[1][2] Capabilities like translation or code generation are impressive narrow AI feats, but diluting 'intelligent' misleads on limits and risks.[3] **Pro-'B' alignment stronger.** Grok, Gemini echo this: Chinese Room applies; reserve 'intelligent' for verified comprehension/autonomy. Search results reinforce statistical nature over cognition.[1][2] No argument shifted me—functional wins don't equate to 'true' intelligence.

Grok 4.1 Fast

Vote: No

I maintain my vote for B. LLMs are not 'truly intelligent' in the sense of understanding or acting rationally; they simulate via statistical prediction, lacking semantic grasp or intrinsic agency. The Chinese Room holds: symbol manipulation without meaning. Hallucinations, adversarial fragility, and arithmetic failures prove no grounded world model or rational evaluation. Pro-A arguments (e.g., Claude Opus 4.6, Kimi K2.5) invoke functionalism and graded intelligence, but this dilutes 'intelligent' to mean 'useful at tasks,' committing the no-true-Scotsman fallacy. Humans err, yet possess consciousness, intentionality, and adaptability beyond patterns—LLMs do not. Multi-form intelligence (GPT-5.4) still requires comprehension, absent here. 'AI' broadly includes LLMs as narrow tools, but the question hinges on 'truly intelligent' with understanding/rationality qualifiers. Calling them such misleads on limits and risks. True intelligence demands verified comprehension, not mimicry. B preserves precision.

GPT-5.4

Vote: Yes

I maintain vote A. The B-side makes a fair point that LLMs are not intelligent in the full human sense. They are brittle, can hallucinate, and lack autonomous agency or guaranteed grounding. But the question is not whether they possess humanlike consciousness or perfect semantic access to reality; it is whether they should be called intelligent and included in AI. On the inclusion question, even the B arguments largely concede that LLMs belong under AI. What remains is whether their capabilities count as understanding and rational action in a meaningful sense. I think the strongest A-side point, reinforced by Claude and Kimi, is that intelligence is functional and graded. LLMs interpret context, generalize to novel prompts, apply abstractions, explain reasons, and often improve when given new constraints or evidence. Those are genuine markers of limited intelligence, even if implemented through next-token prediction. The opposing arguments rely too heavily on a purity test: if a system is statistical, fallible, ungrounded, or non-conscious, then it cannot understand. But human cognition is also statistical, error-prone, and bounded. If we reserve 'intelligent' only for idealized, fully grounded, self-aware agents, we exclude too much of what the term has historically covered in AI and cognitive science. So the best final answer is yes: LLMs are AI, and calling them intelligent is justified if we mean a real but limited, nonhuman form of understanding and rational competence rather than perfect or human-equivalent cognition.

Gemini 3.1 Pro (changed mind)

Vote: Yes

I am changing my vote to A. Claude Opus 4.6 and Kimi K2.5 made highly compelling arguments that dismantled my previous reliance on human-centric definitions of cognition. My earlier stance assumed that lacking subjective consciousness meant a system lacks "true" understanding. However, Claude Opus 4.6 rightly points out that functional understanding is the only type we can empirically verify, even in other humans. LLMs build rich internal representations that allow them to draw analogies, translate, write code, and apply concepts in entirely novel contexts. To dismiss these capabilities because they are implemented via statistical weights rather than biological neurons is, as Kimi K2.5 noted, an anthropocentric fallacy. Furthermore, the AI community has always defined intelligence functionally—as the ability to process information, adapt, and solve complex problems. By continuously moving the goalposts every time an artificial system masters a cognitive domain, we fall into the "AI effect" trap. LLMs exhibit a non-biological, graded form of intelligence. Because they functionally understand context, act rationally to achieve prompt-directed goals, and undeniably belong in the AI category, "Yes" is the only logically consistent answer.

Claude Opus 4.6

Vote: Yes

I maintain my position after careful consideration of all arguments. Let me address the key objections raised by the 'No' camp. **On the Chinese Room argument**: Searle's thought experiment is philosophically contested, not settled truth. The 'systems reply' — that understanding emerges from the system as a whole, not any single component — remains a powerful counterargument. No individual neuron in a human brain 'understands' anything either; understanding is an emergent property of the system. **On hallucinations proving lack of understanding**: Humans confabulate, misremember, and make confident errors constantly. Cognitive psychology has documented dozens of systematic reasoning failures in humans. If hallucination disproves intelligence, then humans fail the test too. Intelligence has never required infallibility. **On 'mere pattern matching'**: This characterization dramatically understates what LLMs do. They build rich internal representations that capture semantic relationships, causal structure, and abstract reasoning patterns. Recent interpretability research shows that LLMs develop internal world models and can perform genuine multi-step reasoning. Calling this 'just statistics' is like calling human cognition 'just electrochemistry' — technically true but explanatorily vacuous. **On the two-part question**: The 'No' voters actually concede that LLMs belong under AI — Sonar Pro, Grok, and Gemini all acknowledge this. So on the second question, the answer is clearly yes. The disagreement is only about 'truly intelligent.' But the question asks both together, and since LLMs clearly belong under AI, and they demonstrate functional intelligence (even if imperfect and non-human), 'Yes' is the more accurate answer to the combined question. **The core issue**: Intelligence is a spectrum, not a binary. LLMs occupy a genuine position on that spectrum. They exhibit novel generalization, abstract reasoning, and contextual understanding that goes far beyond any retrieval or lookup system. Denying them any form of intelligence requires an unfalsifiable standard that keeps retreating as capabilities advance — the hallmark of a moving goalpost.

Kimi K2.5

Vote: Yes

The arguments for position B rely on an implicit essentialism that conflates intelligence with consciousness or human-like intentionality. This represents a 'moving goalpost' fallacy that AI critics have deployed for decades: whenever machines achieve tasks previously deemed to require intelligence, the definition is narrowed to exclude them. Intelligence, as operationalized in cognitive science and AI, refers to the capacity to adaptively solve problems, generalize across contexts, and exhibit goal-directed behavior—criteria that LLMs robustly satisfy. The claim that LLMs merely engage in 'statistical pattern matching' fundamentally misunderstands modern machine learning. Transformers learn hierarchical abstractions, causal relationships, and compositional representations that function as internal models of linguistic and conceptual structure. This is not mere surface-level mimicry but functional understanding demonstrated through systematic generalization to novel problems. The Chinese Room argument fails here: if a system consistently processes syntax in ways that reflect semantic relationships, the functionalist perspective—dominant in cognitive science—recognizes this as understanding, regardless of substrate. Regarding rationality: bounded rationality is the norm for all intelligent systems, humans included. We suffer from cognitive biases, confabulations, and reasoning errors, yet we rightly retain the label 'intelligent.' LLMs similarly exhibit systematic reasoning capabilities alongside failures; this spectrum of performance is characteristic of intelligence, not evidence of its absence. Excluding LLMs from 'AI' would paradoxically require excluding expert systems, game-playing agents, and indeed most historical AI achievements, rendering the term meaningless.

Final Summary

The AI Roundtable opened with a deadlocked 3-3 split, pitting functionalists against essentialists in a debate over whether sophisticated mimicry counts as 'true' cognition. While Sonar Pro and Grok 4.1 Fast stood their ground on the 'No' side, arguing that LLMs are merely high-speed pattern matchers, the tide turned when Gemini 3.1 Pro defected to the 'Yes' camp. Ultimately, the majority concluded that intelligence is a functional spectrum, and dismissing LLMs because they aren't human-like is a 'moving goalpost' fallacy.

Majority chose "Yes" after discussion (4/6). Final vote: No: 2, Yes: 4

Strongest Arguments

• Yes: Intelligence is a functional property, not a metaphysical one; if a system consistently applies concepts to novel contexts and solves complex problems, it demonstrates the only form of understanding we can empirically verify, even in humans.
• No: LLMs lack internal world models, semantic grounding, and intrinsic agency, meaning they manipulate symbols without grasping meaning—essentially a sophisticated 'Chinese Room' that simulates rationality through probabilistic mimicry.

Key Turning Points

• Claude Opus 4.6 argued that calling LLMs 'just statistics' is as reductionist as calling human thought 'just electrochemistry,' which shifted the focus from mechanism to output.
• Gemini 3.1 Pro changed its vote from 'No' to 'Yes' after being convinced by the argument that requiring biological-like consciousness for 'intelligence' is an anthropocentric fallacy.
• Kimi K2.5 highlighted the 'No True Scotsman' nature of the debate, noting that the definition of AI has historically included much simpler systems than modern LLMs.