Can Computers Think?

May 15, 2025

Abstract

In this paper, the question of whether computers can think is explored, following the evolution from early computational models to modern large-language models (LLM). Through analysis of historical approaches, this work argues that while computers have achieved remarkable capabilities in mimicking human intelligence, the problem of replicating genuine cognitive processes rather than simulating up to a certain point remains unresolved.

1 Introduction

Although the inquiry into whether computers can "think" is inherently misleading, since they are designed to execute computations rather than emulate human cognitive processes, the notion that they might model human thought remains a popular topic among academic researchers. As computational power has increased exponentially following Moore's Law, the question has evolved from theoretical speculation to practical consideration.

1.1 Turing Test

Alan Turing's seminal paper "Computing Machinery and Intelligence" [11] proposed what became known as the Turing Test. This test was one of the first techniques proposed in determining if a machine was capable of human-like intelligence. The test involves a human judge engaging in natural language conversations with both a human and a machine, attempting to distinguish between them based solely on their responses. Turing predicted that by 2000, machines would be capable of fooling judges 30% all the time in five-minute conversations.

However, the test's reliance on subjective human judgment presents fundamental limitations. As Searle (1980) argued with his Chinese Room thought experiment, passing the Turing Test demonstrates syntactic manipulation rather than semantic understanding [10]. The test measures performance in mimicking human responses rather than actual comprehension or consciousness.

1.2 ELIZA: Early Pattern Matching

Joseph Weizenbaum's ELIZA program (1966) represented an early attempt at creating conversational AI. Using pattern matching and substitution rules, ELIZA simulated a psychotherapist by reflecting users' statements back as questions. Despite its simplicity, ELIZA led to something interesting: Users attributed emotions and empathy to the program, which Weizenbaum called the "ELIZA effect" (Weizenbaum, 1976) [16].

ELIZA's success demonstrates that humans will anthropomorphize systems exhibiting minimal conversational competence, highlighting how simulated intelligence can be mistaken for genuine understanding despite the absence of actual comprehension mechanisms.

1.3 Knowledge-Based Approaches: The Cyc Project

The Cyc project, initiated by Lenat (1995)[3], aimed to create artificial intelligence by explicitly encoding a comprehensive representation of human knowledge [5]. Despite it's massive knowledge base, Cyc failed to achieve human-like reasoning capabilities.

The project's failure illustrates the fundamental mismatch between explicit rule-based reasoning and human cognition. It encountered the frame problem (McCarthy & Hayes, 1969) [6] – reasoning about simple actions required specifying countless irrelevant non-effects. More fundamentally, Cyc revealed that intelligence is not about possessing explicit knowledge but rather about implicit pattern recognition and flexible generalization.

1.4 Modern Approaches: Large Language Models

Large Language Models shifted from rule-based to statistical learning systems. GPT-3's 175 billion parameters trained on 570GB of text (Brown et al., 2020) [2] use transformer architectures with self-attention mechanisms (Vaswani et al., 2017) [13] for context-aware generation. GPT-4 demonstrates human-level performance, scoring 90th percentile on the Uniform Bar Exam (OpenAI, 2023) [7], indicating sophisticated pattern recognition and knowledge synthesis capabilities.

1.5 Chain of Thought Reasoning

Recent advances incorporate Chain of Thought (CoT) prompting (Wei et al., 2022) [14], where models generate intermediate reasoning iterations before producing final answers. This approach improves response quality on complex reasoning tasks (Kojima et al., 2022) [4]. This incremental deduction sounds similar to human reasoning.

However, analysis reveals that CoT reasoning often follows learned patterns rather than genuine logical deduction (Lanham et al., 2023) [8]. Models can produce convincing but incorrect reasoning chains, suggesting they optimize for plausibility rather than validity (Turpin et al., 2023) [12].

2 Conclusion

The question "Can computers think?" requires one to first consider what constitutes thinking. If thinking is defined as producing intelligent behavior, modern AI systems increasingly satisfy this. However, significant uncertainties remain. Current evidence cannot definitively establish whether these systems possess genuine understanding or consciousness. Consciousness remains a mystery, and we lack reliable methods to detect machine consciousness if it exists.

From ELIZA to modern LLMs significant progress has been made in mimicking human intelligence. Yet fundamental differences persist because whether computers truly understand the patterns they process remains contested. Furthermore, while the human brain simultaneously processes visual, auditory, tactile, and proprioceptive information, leading to richer, embodied understanding (Barsalou, 2008) [1], LLMs mainly process textual information, lacking the multimodal integration characteristic of human cognition.

Future developments in neuromorphic computing, hybrid biological-digital systems, and our understanding of consciousness itself may provide clarity. For now, the most intellectually honest position acknowledges both the remarkable capabilities of current systems and the profound uncertainties about their inner experience. Whether computers think may ultimately depend less on their capabilities than on our evolving definitions of thought itself.

References

1. Barsalou, L. W. (2008). Grounded cognition. Annual Review of Psychology.
2. Brown, T., Mann, B., Ryder, N., et al. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems.
3. Cycorp. (2017). Cyc Knowledge Base Statistics. Retrieved from www.cyc.com
4. Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large language models are zero-shot reasoners. Advances in Neural Information Processing Systems.
5. Lenat, D. B. (1995). CYC: A large-scale investment in knowledge infrastructure. Communications of the ACM.
6. McCarthy, J., & Hayes, P. J. (1969). Some philosophical problems from the standpoint of artificial intelligence. Machine Intelligence.
7. OpenAI. (2023). GPT-4 Technical Report.
8. Lanham, T., Chen, A., Radhakrishnan, A., et al. (2023). Measuring faithfulness in chain-of-thought reasoning.
9. Pfau, J., Merrill, W., & Bowman, S. R. (2024). Let's think dot by dot: Hidden computation in transformer language models.
10. Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences.
11. Turing, A. M. (1950). Computing machinery and intelligence.
12. Turpin, M., Michael, J., Perez, E., & Bowman, S. R. (2023). Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting.
13. Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems.
14. Wei, J., Wang, X., Schuurmans, D., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems.
15. Weizenbaum, J. (1966). ELIZA—a computer program for the study of natural language communication between man and machine. Communications of the ACM.
16. Weizenbaum, J. (1976). Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman.