Why general artificial intelligence will not be realized

The paper interrogates whether artificial general intelligence (AGI) is achievable and argues it is not. It situates AGI within the broader AI landscape by distinguishing it from artificial narrow intelligence (ANI) and from strong AI. The author reviews historical definitions (e.g., Minsky’s weak AI, Searle’s strong AI) and frames the core research question around whether progress in ANI implies steps toward AGI. He highlights early critiques (Weizenbaum’s distinction between computer power and human reason; Penrose’s non-algorithmic mind) and introduces Hubert Dreyfus’s foundational critique that intelligence is embodied, encultured, and grounded in tacit know-how that cannot be fully formalized. The introduction also discusses later paradigms—neural networks, deep learning, and Big Data—and the cultural and commercial forces fueling overoptimistic claims. The article’s purpose is to evaluate these claims, separate hype from reality, and argue, in principle, against the realizability of AGI on the grounds that computers are not ‘in the world’ (lack embodiment, culture, and capacity to intervene causally). The stakes are practical and ethical: overestimating technology and underestimating human skills risks replacing effective human practices with inferior systems.

• Classical and contemporary critiques of AI: Weizenbaum (1976) distinguishes fast algorithmic “computer power” (akin to ANI) from non-algorithmic human prudence and wisdom; Penrose (1989, 1994) argues human thinking is fundamentally non-algorithmic.
• Dreyfus tradition: Drawing on Heidegger, Merleau-Ponty, Wittgenstein, and Polanyi, Dreyfus (1972; Dreyfus & Dreyfus, 1986) contends expertise is rooted in embodied, tacit knowledge that resists full articulation and programming; expert systems miss this tacit dimension.
• Neural networks and deep learning: Reviewed as a shift from symbolic AI to connectionist models capable of learning patterns without explicit rules; successes in pattern recognition and games (AlphaGo) are acknowledged.
• Big Data discourse: Mayer-Schönberger & Cukier (2014) champion correlation-driven analysis and suggest diminishing centrality of causality; critiques include the ‘fallacy of initial success’ (e.g., Google Flu Trends) and risks of spurious correlations.
• Causality and inference: Mill’s methods (1882) and historical cases (John Snow’s cholera investigation) illustrate causal reasoning by intervention; Pearl & Mackenzie’s “mini-Turing test” (2018) argues human-like intelligence requires causal models and counterfactual reasoning—capabilities current ML systems lack.
• Broader philosophical context: Koyré and Galileo’s mathematization of nature; Russell’s skepticism about causality in physics; Husserl’s warning against scientism and emphasis on the lifeworld grounding of science; contemporary reductionist claims (Harari 2018; Crick 1994) critiqued for neglecting embodiment and context.
• Turing and conversational benchmarks: The Turing test and Loebner Prize chatbots (e.g., Mitsuku) illustrate persistent failures in understanding, context, and coherence.

The article employs philosophical and conceptual analysis supported by historical and contemporary case studies. The author:

• Theoretically grounds the argument in phenomenology and philosophy of science (Husserl, Heidegger, Wittgenstein, Polanyi; Dreyfus) to define intelligence as embodied, encultured, and tacitly skilled.
• Analyzes emblematic AI milestones (Deep Blue, IBM Watson, AlphaGo) to separate ANI successes from claims about AGI, examining scope, domain specificity, and transfer/generalization.
• Uses empirical cautionary tales (Google Flu Trends, the husky-vs-wolf snow detector) to demonstrate pitfalls of correlation without causal understanding.
• Explains causal inference via Mill’s methods and the historical John Snow cholera case to emphasize intervention and control as central to causal knowledge.
• Evaluates conversational evidence (Loebner Prize/Mitsuku transcripts) to show failures in pragmatic understanding and context handling, linking them to non-embodiment. Overall, the approach triangulates philosophical argument, illustrative examples, and critical review of AI outcomes to argue that AGI is, in principle, unrealizable because computers are not situated ‘in the world’ and cannot engage in causal intervention and lifeworld understanding.

• Advances in ANI do not constitute advances toward AGI: Systems like Deep Blue (1997) and AlphaGo (2016) exhibit high performance in constrained domains but lack generality, flexibility, and understanding.
• IBM Watson: Despite winning Jeopardy! in 2011 against top champions (Ken Jennings and Brad Rutter) with access to ~200 million pages of information, subsequent efforts to translate this into medical super-diagnostics underperformed; several Watson Health initiatives were reduced to or repurposed as routine AI assistants, with notable project closures and failures.
• Deep reinforcement learning limitations: AlphaGo was trained on ~150,000 expert games and self-play, and DeepMind showed successes on Atari and StarCraft, yet systems proved brittle to small perturbations, inflexible to environmental changes, and hard to commercialize at scale. Reported DeepMind financial losses underscore the challenge: approximately $154M (2016), $341M (2017), $572M (2018).
• Opacity and spurious correlations: A neural network distinguishing huskies vs. wolves achieved ~90% accuracy by exploiting snow in backgrounds—an example of learning the wrong feature due to non-causal correlations.
• Big Data caution: Google Flu Trends initially appeared successful but subsequently overestimated influenza-like illness rates, at one point reporting roughly double the CDC’s estimates (2013), illustrating the ‘fallacy of initial success’ and the hazards of relying on correlations without causal grounding.
• Causal knowledge requires intervention: Following Mill’s method of difference and historical epidemiology (John Snow’s cholera study), causality is established by controlled differences and manipulation; Snow’s data (deaths per 10,000 households: Southwark & Vauxhall 315; Lambeth 37; rest of London 59) and pump removal align with causal inference via intervention.
• Mini-Turing test (causal reasoning) and full Turing test remain unmet: Pearl & Mackenzie argue current ML operates associationally, lacking causal and counterfactual reasoning. Loebner Prize chatbot exchanges (e.g., Mitsuku) reveal failures in understanding simple pragmatic questions, suggesting lack of world-embedded comprehension.
• Embodiment and lifeworld: Intelligence depends on being ‘in the world’—having a body, growing up in a culture, practical engagement, and tacit know-how. Computers lack this, so AGI is argued to be unrealizable in principle.
• Overestimation of AI and underestimation of human skills: Optimistic narratives (commercial and quasi-religious tech beliefs) lead to harmful substitutions of effective human practices with inferior automated systems.

The paper addresses the central question—whether AGI is achievable—by integrating philosophical analysis with case-based critique. Building on Dreyfus and phenomenology, it argues that human intelligence crucially involves embodiment, cultural practices, and tacit expertise, which cannot be reduced to explicit rules or purely statistical associations. While deep learning can approximate tacit pattern recognition in idealized, closed-world tasks (e.g., Go), it lacks transparent, causal, and context-sensitive understanding required for general intelligence. The examination of IBM Watson, AlphaGo/DeepMind, and Big Data exemplifies the gap between hype and reality: systems excel in narrow, static, and well-specified environments but falter when facing shifting contexts, require causal reasoning, or must integrate lifeworld knowledge. Pearl & Mackenzie’s mini-Turing test frames the necessity of causal models and interventions; current ML remains largely associational, failing this benchmark. The Loebner Prize dialogues further illustrate absence of pragmatic understanding and embodiment. The significance is twofold: theoretically, it challenges reductionist claims that human cognition is merely algorithmic; practically, it warns that deploying overpromised AI can degrade professional practice and scientific inquiry by privileging correlation over causation and undervaluing human judgment and skills.

AGI will not be realized: computers are not ‘in the world’—they lack embodiment, cultural upbringing, and the capacity for causal intervention and lifeworld understanding required for general intelligence. The successes of ANI (chess, Go, quizzes) do not imply progress toward AGI, and Big Data’s emphasis on correlation cannot replace causal knowledge. Empirical case studies (Watson Health setbacks, brittleness in deep RL, Google Flu Trends failures) and philosophical analysis (Dreyfus, Polanyi, Husserl; Pearl & Mackenzie on causality) converge to show that AGI is, in principle, unattainable. Overestimating AI while underestimating human skills risks replacing well-functioning human practices with inferior technologies. The paper reaffirms the continuing relevance of Dreyfus’s critique and cautions against scientism and techno-optimist overreach. No explicit future research agenda is proposed.