The Five A's of AI - Chapter 10

The confusion begins with terminology itself, compounded by massive financial speculation. When computer scientists discuss artificial intelligence, they typically mean narrow AI systems that excel at specific tasks. When philosophers and futurists engage with artificial intelligence, they envision artificial general intelligence (AGI) possessing human-like adaptability and understanding across all domains. When vendors market artificial intelligence, backed by the generative AI funding that reached approximately $45 billion in 2024, nearly doubling from $24 billion in 2023 (Mintz Legal Advisory, 2025), they mean whatever sounds most impressive to potential customers.

This linguistic chaos, amplified by extraordinary investment flows, creates dangerous misunderstandings about current capabilities and future possibilities. Artificial general intelligence (AGI), sometimes called human-level intelligence AI, is a type of artificial intelligence that would match or surpass human capabilities across virtually all cognitive tasks. Some researchers argue that state‑of‑the‑art large language models already exhibit early signs of AGI‑level capability, while others maintain that genuine AGI has not yet been achieved.

The stakes of this confusion extend beyond academic debate. Artificial intelligence investment is expected to approach $200 billion globally by 2025, according to Goldman Sachs research, representing enormous capital flowing towards a goal whose definition remains contested. Worldwide spending on artificial intelligence, including AI-enabled applications, infrastructure, and related information technology and business services, will more than double by 2028 when it is expected to reach $632 billion, according to International Data Corporation (IDC) forecasts. This investment surge occurs whilst fundamental questions about artificial intelligence remain unresolved.

Consider the remarkable capabilities of today's most sophisticated AI systems and compare them with what many believe they achieve. Large language models like GPT produce remarkably human-like text yet possess no understanding of meaning. They manipulate symbolic patterns with extraordinary sophistication but lack any internal representation of the concepts those symbols represent. Chess engines defeat world champions but cannot apply strategic thinking to business decisions. Medical diagnosis systems outperform doctors in narrow domains yet cannot engage in basic reasoning outside their training.

Each achievement represents remarkable progress within constraints. Each also demonstrates how far we remain from general intelligence. Understanding this gap between marketing hyperbole and technical reality proves essential for organisational planning, regulatory preparation, and societal adaptation to AI's actual rather than imagined capabilities.

Defining artificial general intelligence proves surprisingly challenging. We recognise human intelligence intuitively yet struggle to specify its essential characteristics. This definitional challenge extends beyond philosophical curiosity. It shapes research directions, investment decisions, and societal preparation for potential AGI emergence.

True AGI would interpret new information, learn, and adapt to scenarios without human oversight, all while demonstrating human-level performance, reasoning, and common sense. This encompasses capabilities that current AI systems fundamentally lack, despite their impressive domain-specific achievements.

Generalisation across domains represents AGI's most fundamental requirement. Humans apply intelligence fluidly between contexts. A physicist learning cooking, a chef grasping basic physics, both navigating social situations whilst appreciating art and solving novel problems. This transfer of cognitive capability distinguishes general from narrow intelligence.

Current AI systems fail catastrophically when asked to generalise beyond training domains. A system mastering medical diagnosis cannot play chess without complete reprogramming. A language model generating coherent text about quantum physics possesses no understanding of scientific methodology. This brittleness persists regardless of sophistication within specific domains. Current AI models like those used in autonomous vehicles require enormous datasets and computational power just to handle driving in specific conditions, let alone achieve general intelligence.

The challenge extends beyond technical limitations to fundamental questions about knowledge representation. Humans somehow encode abstract concepts that apply across disparate domains. The notion of "balance" applies to chemistry, accounting, design, and personal relationships. We recognise patterns like cause-and-effect, competition, and cooperation across physics, economics, biology, and social interactions. This cross-domain pattern recognition enables humans to learn rapidly in new fields by applying relevant knowledge from familiar ones.

Current AI architectures show no evidence of developing such flexible knowledge representations. Each system remains trapped within its training domain, unable to recognise that strategic principles from chess might apply to business planning, or that statistical reasoning from finance might illuminate medical diagnosis. The generalisation problem represents more than an engineering challenge; it reflects our incomplete understanding of how human intelligence achieves domain transfer.

The distinction between pattern matching and genuine understanding remains central to discussions of artificial intelligence. Current AI systems manipulate symbolic patterns without comprehension, operating through statistical associations rather than semantic understanding. Whether understanding requires consciousness, or whether sophisticated pattern matching might constitute a form of intelligence, represents an ongoing philosophical question with practical implications for AI development.

We see implications beyond literal statements. Current language models manipulate symbolic patterns without comprehension. 

Recent research provides sobering perspective on this limitation. Apple researchers investigating the mathematical reasoning capabilities of large language models found that these systems rely on sophisticated pattern matching rather than genuine logical reasoning. 

When they introduced GSM-Symbolic, a benchmark that generates diverse mathematical questions through symbolic templates, they discovered that adding irrelevant but seemingly related information leads to a performance drop of up to 65%. This reveals that even our most advanced AI systems don't reason in the way vendors suggest; they pattern-match at extraordinary scale.

The implications prove profound. Large language models and other AI systems have already learned, from their training, the ability to deceive via techniques such as manipulation, sycophancy, and cheating the safety test. Yet this apparent sophistication masks fundamental limitations. The systems exhibit behaviours associated with understanding whilst possessing no internal model of meaning.

Understanding involves more than correct responses; it requires comprehension of relationships, context, and implications. When humans read about photosynthesis, we grasp not just the chemical equation but its significance for life on Earth, its relationship to climate, and its applications in technology. We understand why changing one variable affects others and can predict consequences in related systems. Current AI systems demonstrate no evidence of such deep comprehension, instead relying on statistical associations between text patterns learned during training.

Creative problem-solving through insight distinguishes intelligence from computation. Humans faced with novel challenges don't just interpolate from experience. We re-conceptualise problems. We discover non-obvious connections. We generate genuinely new solutions. We have "aha!" moments where scattered information suddenly crystallises into understanding. Current AI systems optimise within problem spaces. They cannot redefine the spaces themselves. They climb hills efficiently. They cannot recognise when different mountains offer better peaks.

The creative limitation manifests most clearly in breakthrough scientific thinking. Consider Darwin's insight that natural selection could explain biological diversity, Einstein's realisation that space and time are relative, or Watson and Crick's recognition that DNA forms a double helix. Each breakthrough required more than processing existing information; it demanded reframing entire conceptual frameworks. The scientists didn't solve problems within established paradigms but created new paradigms for understanding reality.

Modern AI systems show no evidence of such paradigm-shifting capability. They can optimise designs within known parameters, suggest improvements based on existing examples, and even combine elements in novel ways. But they cannot step outside established frameworks to create fundamentally new approaches. Their creativity remains combinatorial rather than revolutionary, rearranging existing elements rather than discovering new principles.

This limitation extends to everyday problem-solving. Humans excel at recognising when conventional approaches won't work and developing entirely different strategies. A chef facing missing ingredients doesn't just substitute similar items but might completely change the cooking method. An engineer encountering unexpected constraints doesn't just optimise existing designs but might adopt entirely different principles. This adaptive creativity remains beyond current AI capabilities.

The relationship between consciousness and intelligence remains hotly debated. Some researchers, including microprocessor inventor Federico Faggin, propose that consciousness represents a fundamental aspect of reality rather than an emergent property of complex systems. This perspective suggests that artificial systems, however sophisticated, may never achieve genuine understanding without consciousness. Others argue that functional equivalence might suffice, regardless of subjective experience.

Some argue consciousness emerges from sufficient complexity. They believe AGI will necessarily be conscious. Others contend philosophical zombies could exhibit all external signs of intelligence without inner experience. The debate matters practically. Conscious AI would demand different ethical considerations than unconscious systems, however intelligent.

Artificial consciousness, also known as machine consciousness, synthetic consciousness, or digital consciousness, is the consciousness hypothesised to be possible in artificial intelligence. The 2020s have witnessed exceptional focus on this question. Since sentience involves the ability to experience ethically positive or negative (i.e., valenced) mental states, it may justify welfare concerns and legal protection, as with animals. The practical implications prove profound. Conscious AI would demand different ethical considerations than unconscious systems, however intelligent their behaviour.

Qualia, or phenomenological consciousness, is an inherently first-person phenomenon. Because of that, and the lack of an empirical definition of sentience, directly measuring it may be impossible. This measurement challenge creates profound difficulties for recognising consciousness in artificial systems. Brain-Computer Interface (BCI) technology, which facilitates direct communication between the brain and external devices, emerges as one approach for understanding consciousness, demonstrating significant promise in research contexts.