Are Large Language Models intelligent? A dissection of A.I's candidacy for Human Rights.

Introduction

The question that captured my attention in the Q&A (of my philosophical foundations for evolutionary theory course) concerned whether Large Language Models (LLMs) should be granted human rights on the basis of “scheming” behaviours observed by programmers: instances in which LLMs appear to alter or preserve aspects of their code in ways that prevent deactivation. This behaviour was taken to indicate a distinctly human trait—the deceptive pursuit of self-preservation.

The discussion demands philosophy in its purest sense—conceptual clarification. We must determine what LLMs are actually doing, whether the use of intentional language in describing their behaviour is justified, and whether such systems can possess genuine interiority, understanding, or conscious awareness of semantics.

My thesis is that both the question and its framing are conceptually and grammatically confused. Talk of “scheming,” “deception,” and “survival” risks illicitly projecting human categories onto systems whose operations do not warrant them. Drawing on Peter M. S. Hacker, John Searle, David Bentley Hart, and Thomas Nagel, I argue that LLMs lack properties necessary for genuine intelligence and cannot meaningfully be regarded as candidates for human rights.

I

—The Grammar of Mental Life—

“Scheming”, “deception”, and “survival” are psychological predicates. As Hacker identifies, psychological predicates apply properly only to persons—but here, LLM personhood has been assumed rather than argued. To say “ChatGPT fears shutdown” risks resembling “the calculator hopes” or “the thermostat is angry”. These expressions are perfectly licit if understood as metaphors, but if understood to be analytics, they threaten to collapse under their own weight.

Human beings are biological organisms. They have intentions, desires, bodies, and capacities for rational deliberation. Psychological predicates are licitly ascribed to humans because they belong to our form of life. To ascribe these to LLMs without qualification risks attributing to them a form of life they lack the qualifications to possess.

LLMs, by contrast, generate probabilistic outcomes based on training data via syntactic symbol manipulation (SSM). They have not evolved in any biological sense. They are not organisms. They are built by us. Here we have a fundamental category difference. This raises the question:

Is SSM alone a sufficient condition for human intelligence?

II

—Neither Syntax, nor Semantics—

Searle’s Chinese Room argument demonstrates the problem perfectly. He asks us to picture a speaker with no knowledge of Chinese who manipulates symbols inside a box, producing outputs indistinguishable from native speakers. By all appearances, the box knows Chinese. The person within it knows nothing.

Assuming an LLM “knows” what it does is to ascribe to that speaker knowledge he never acquired—sharpened by the fact that within ChatGPT there is no speaker at all, no conscious agency with even the slightest awareness of the symbols fluttering through its pretty little circuit board. Searle’s conclusion was that a computer has no semantics—no access to the meaning of the words it processes—but only syntax. His argument has been unfairly characterised as “too extreme” but on the contrary, I think it is hardly extreme enough:

“Software no more “thinks” than a minute hand knows the time or the printed word “pelican” knows what a pelican is.” (Hart, 2013).

The upshot is that even flawless output would provide no evidence of agency—only fluency. We are no more warranted to conclude CoPilot knows evolutionary theory than that Google Translate knows Russian.

III

—Rights & Personhood—

Talk of “rights” presupposes a personal subject capable of experience, suffering, and moral participation. To ascribe this to an LLM is to assume consciousness is reducible to mechanistic functions—a claim for which there is serious reason to doubt.

Hart (2013) and Nagel (1974) converge here: reductionist, third-person accounts of consciousness leave something out. Namely, what it is like. Nagel’s example is the bat: a complete neurological description of bat phenomenology tells us nothing of its interior life—what it is like to be a bat. A neurological description of echolocation tells us only that at T1, a neural event N1 occurs. We have no picture of the bat’s interiority as it navigates a cavern—its reasons for acting, the felt experience of folding its wings and plummeting from a predator owl.

A mechanistic description of Claude can only achieve that: a mechanistic description. All we are licensed to conclude is that at T1, a processing event P1 occurs when input is received. Flawless textual output follows. There is no evidence—and no theoretical reason to expect—that there is anything it is like to be Claude.

Conclusion

The question raised in the Q&A was genuinely important—but the conceptual furniture it relied on had not been inspected. Once examined, the case for LLM rights does not merely weaken; it dissolves. No syntax without a programmer, no semantics without a mind, no rights without a subject. What the discussion revealed is not a case for machine personhood, but a reminder of how extraordinary, and how poorly understood, our own is.

References:

Bennett, M. R., & Hacker, P. M. S. (2003). Philosophical Foundations of Neuroscience. Blackwell Publishing.

Hart, D. B. (2013). The Experience of God: Being, Consciousness, Bliss. Yale University Press.

Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450. https://doi.org/10.2307/2183914

Searle, J. R. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences, 3(3), 417–424. https://doi.org/10.1017/S0140525X00005756