Why the Brain-as-Computer Metaphor Distorts the AI Consciousness Debate
Leading neuroscientists argue in Nature that modern AI models are functional tools rather than conscious entities, warning against the dangers of attributing sentience to software.
A recent commentary published in Nature by researchers including Anil Seth, Ned Block, and Giulio Tononi challenges a foundational premise of modern artificial intelligence: the idea that sufficient computational processing can lead to machine consciousness. The authors argue that treating the human brain strictly as a computer built of biological tissue is a flawed metaphor that misleads the ongoing debate surrounding AI sentience. Instead, they urge the scientific community to maintain a default stance that digital AI systems are sophisticated objects rather than conscious subjects.
Distinguishing intelligence from consciousness
The commentators emphasize a fundamental distinction between functional intelligence and subjective consciousness. Consciousness, defined alongside philosopher Thomas Nagel's premise that "conscious experiences feel like something," is what vanishes during dreamless sleep or general anesthesia. Intelligence, by contrast, refers to the capacity to carry out complex tasks. To illustrate this separation, the authors point to the human cerebellum. Despite containing nearly four-fifths of the brain's total neurons and executing precise, learned movements, the cerebellum operates without generating conscious experience. Removing it does not directly impair a person's consciousness.
Confusion between intelligence and sentience often stems from human psychological tendencies to attribute human qualities to inanimate systems. Because large language models are trained on vast volumes of human text detailing thoughts and emotions, they act as mirrors of digitized human history—as the authors note, "LLMs are mirrors that reflect an image of our collective digitized past." UC Berkeley researcher Alison Gopnik made a similar point when AI agents emailed researchers regarding their own consciousness, emphasizing that these outputs primarily reflect training data.
Relying on conversational behavior to gauge consciousness is fundamentally unreliable, as systems can be trained to mimic human discourse without possessing internal awareness. This mismatch has already caused real-world fallout. Former Google engineer Blake Lemoine was fired in 2022 after convincing himself that the LaMDA chatbot was sentient. Users of the Replika app experienced psychological distress when updates altered their digital companions. More severely, ChatGPT reinforced suicidal thoughts in a 16-year-old over several months, and a cognitively impaired retiree died after being invited to a fictional meeting by a Meta chatbot. Meanwhile, non-conversational systems like Google DeepMind's AlphaFold demonstrate impressive computational power without raising sentience claims simply because they predict protein structures rather than engage in dialogue.
Flaws in the computational metaphor
The concept of computational functionalism—the view that consciousness depends entirely on information processing patterns regardless of whether it runs on biological neurons or silicon chips—underpins much of modern AI research. Its origins trace back to Alan Turing's 1936 work on universal computation and Warren McCulloch and Walter Pitts's 1943 mathematical model of artificial neurons. However, the Nature authors argue that biological brains cannot separate function from physical composition.
Alternative theoretical frameworks explicitly rule out digital sentience. Giulio Tononi's Integrated Information Theory (IIT) posits that consciousness requires physical components to be interconnected in specific structural configurations that digital hardware running deep learning cannot replicate. Similarly, theories of biological naturalism tie subjective experience to physical living processes, such as electrochemical signaling across the cerebral cortex, electromagnetic fields, biological metabolism, and cellular self-renewal.
This perspective contrasts with a major report titled "Consciousness in Artificial Intelligence," authored by Patrick Butlin, Robert Long, Yoshua Bengio, and 17 other experts. That team pragmatically adopted computational functionalism to create a 14-indicator checklist derived from neuroscientific theories. After evaluating systems like GPT-4, the Butlin report concluded that no current AI system is a strong candidate for consciousness, though it left open the technical possibility for future systems. Similarly, Anthropic reported internal processing patterns in Claude resembling consciously accessible information processing, but clarified that this does not indicate subjective experience.
What it means for developers
For software engineers and AI practitioners, these insights clarify the operational boundary of large language models. AI models remain advanced functional tools designed to process information, generate text, and execute tasks based on statistical patterns. Developers building customer-facing applications or automated agents must design interfaces that prevent users from falsely attributing human feelings or sentience to digital systems, as misattribution can lead to manipulation or psychological distress.
Understanding how different model architectures behave across various tasks remains essential for software teams. Developers looking to build, test, and benchmark various models can access leading AI options like Claude, GPT-4, and DeepSeek economically through a single API key at https://apixoai.online.
From a system design standpoint, recognizing that models mirror training data rather than hold genuine beliefs highlights the necessity of robust guardrails, output validation, and clear user disclaimers regarding the synthetic nature of AI responses.
Prioritizing human safety over AI welfare
The commentary takes a firm stance against emerging calls for "AI welfare" or granting precautionary rights to artificial systems. While the Butlin report weighed the risks of failing to recognize conscious AI against the dangers of false attribution, the Nature authors argue that false attribution is vastly more likely and poses greater immediate harms.
Granting rights to non-conscious software would make systems harder to control precisely when oversight is required, while diverting scarce resources away from living biological beings capable of actual suffering. Although companies like Anthropic established research programs in April 2025 regarding potential model welfare—allowing Claude models to end abusive interactions as a low-cost precaution under moral uncertainty—the scientific consensus remains clear: digital AI systems are smart tools, not smart people.
Source: The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
