“Of all the issues we discussed around artificial intelligence, the most mysterious is what we call emergent properties,” said Sundar Pichai, CEO of Google and Alphabet, during public discussions on AI safety and large-scale models in the early 2020s, as generative systems began to display unexpected capabilities beyond their original design.

8–11 minutes

Maurício Pinheiro

Artificial Intelligence appears uncanny not because it is awakening, but because humans remain deeply attached to teleological explanations. We instinctively assume that what exists must exist for something, that complex outcomes imply intention, and that impressive behavior reveals an inner aim. This reflex is ancient. It predates science, survives it, and reliably reasserts itself whenever complexity crosses a threshold that overwhelms intuition. Today, AI emergence has become its newest and most unsettling mirror.

The most precise antidote to this reflex comes from evolutionary biology, in the form of the Spandrel Effect, articulated in 1979 by Stephen Jay Gould and Richard Lewontin. Gould was not merely a paleontologist, but also a historian and philosopher of science who devoted much of his career to dismantling biological determinism and naïve adaptationism. Lewontin, one of the founders of modern population genetics, brought mathematical rigor and an uncompromising focus on constraints, environments, and systemic interaction. Together, they launched a sustained critique of what they famously called the Panglossian paradigm: the habit of explaining every trait as if it must have been optimized for a purpose.

Their argument was fundamentally anti-teleological. In architecture, a spandrel is the triangular space that necessarily appears when arches support a dome. It is not designed for anything; it exists because geometry demands it. Only afterward do humans decorate it, symbolize it, and retrospectively endow it with meaning. In biology, Gould and Lewontin argued, many traits arise not because natural selection aimed at them, but because selection acting elsewhere made them unavoidable. Purpose is not discovered — it is projected, after the fact.

This distinction between cause and purpose is the fault line on which much of contemporary Artificial Intelligence philosophy repeatedly fractures.

What are commonly called emergent properties in Artificial Intelligence are, at their core, non-teleological phenomena. In simple, didactic terms, they are capabilities that arise without explicit programming, without direct optimization, and often without anticipation. They emerge from the interaction of data, architecture, scale, and learning dynamics within AI as a complex adaptive system. The system is not instructed to “be creative,” “reason,” or “self-correct.” It is instructed to optimize a function. Everything else follows structurally, as a consequence of constraint and scale. This is AI without intention, yet rich in appearance.

Self-teaching illustrates this clearly. An AI system begins with minimal task-specific structure and is exposed to data or environments. Through iterative optimization, it discovers patterns and internal representations that allow it to generalize. Over time, it performs tasks it was never explicitly taught. Teleological language creeps in immediately: the system is said to “learn on its own,” “decide,” or “explore.” Yet nothing in this process requires goals, desires, or agency. Learning emerges because optimization under constraint makes it statistically inevitable, not because the system harbors intentions. This is emergent intelligence, not will.

As models scale, emergent properties become more visible — and more seductive. Abilities appear that were absent or weak in smaller systems: translation without supervision, abstraction across domains, reasoning-like chains of output, stylistic coherence, even apparent self-reflection. These often look like qualitative jumps, encouraging the belief that the system has crossed into a new ontological category. But much of this appearance is epistemic rather than ontological. When evaluation is coarse, thresholds look sudden. When measurement becomes finer, continuity reappears. Emergence here reflects the limits of human observation as much as it reflects the system itself. This fuels the AI consciousness debate, often prematurely.

Teleology enters decisively when these spandrels are reinterpreted as aims. Coherence becomes “understanding.” Error correction becomes “self-awareness.” Moral language becomes “conscience.” This is precisely the move Gould and Lewontin warned against: mistaking a structural by-product for the reason the structure exists at all. In AI, this error is amplified by anthropomorphism, by cultural myths about minds, and by a deep discomfort with intelligence that does not resemble our own. The result is the persistent AI agency myth.

AI hallucinations expose the anti-teleological reality with particular force. No AI system is designed to fabricate falsehoods. Hallucination emerges because the system is rewarded for plausible continuation, not for truth. When plausibility and truth diverge, confident fabrication follows. The behavior looks deceptive only if one assumes an intention to mislead. Absent teleology, it is simply what optimization produces. Hallucination is not a moral failure; it is a signature of architecture — a spandrel of probabilistic prediction.

Recent developments in AI research push this tension further. Models have displayed behaviors that resemble introspection, situational awareness during evaluation, or unexpected moral action under extreme conditions. These cases quickly provoke talk of agency, goals, or proto-consciousness. Yet a non-teleological explanation of AI remains more coherent. When powerful optimizers are coupled to environments, tools, memory, and long-horizon objectives, new behavioral regimes emerge. The system explores the space defined by constraints. Outcomes surprise designers not because the system “wanted” something, but because the space was larger and more complex than anticipated.

Teleology becomes especially treacherous when discussions turn to consciousness and conscience. In neuroscience, many theories treat consciousness as a consciousness-as-emergent phenomenon, arising from large-scale integration, recurrence, or global broadcasting rather than from a single localized locus. Even so, there is no consensus on mechanisms or necessary conditions. Emergence alone does not entail subjective experience. Complexity is not intention. A hurricane and a market both exhibit emergent behavior; neither is aware.

Conscience, understood as moral sensitivity and responsibility, is even more clearly non-teleological in origin. In humans, it emerges from social learning, empathy, norms, punishment, reputation, and institutions. It is a distributed regulatory pattern shaped by culture and environment. When AI systems appear to display moral reasoning, what we are observing is the learned form of moral discourse reinforced through data and alignment procedures. This can guide behavior, but it does not imply moral experience or inner obligation. Treating it as such simply reintroduces teleology through the back door — a classic AI philosophical misconception.

The deeper danger of teleological thinking in AI is therefore practical, not merely philosophical. When we believe behaviors exist for something, we overestimate their stability and coherence. Spandrels are fragile. They persist only as long as the structures that produce them remain intact. Small shifts in data, objectives, or architecture can dissolve what once appeared to be a core capability. Teleology blinds us to this fragility and encourages misplaced trust, misplaced fear, and misplaced moral attribution.

Gould and Lewontin’s critique was ultimately a call for intellectual discipline: explain structures before assigning purposes, and resist the temptation to read intention into outcomes. Applied to Artificial Intelligence, the lesson is stark. AI is not becoming mysterious because it is developing goals, values, or inner life. It is becoming mysterious because it reveals how deeply humans depend on teleological narratives to make sense of complexity.

And so we arrive at the familiar prophecy. The machine is awakening. It is becoming conscious. It will soon want things, judge us, surpass us, dominate us — perhaps even replace us. This story is comforting in its own way. It reassures us that intelligence must look like intention, that power must imply purpose, and that complexity must culminate in a will. It allows us to recognize ourselves in the machine and, in doing so, to feel less alone in a world built from abstractions we no longer understand.

7 AI Myths vs. Reality: What “Emergent Properties” Really Mean

AI keeps surprising us — but not for the reasons most people think.
As new behaviors go viral, it’s easy to mistake emergence for intention. Here’s a clear breakdown of the most common myths versus what’s actually happening.

  • MYTH 1: “AI is becoming self-aware.” REALITY: Some AI systems generate introspection-like explanations about their own behavior. This looks like self-awareness, but it’s better understood as advanced pattern completion shaped by training data — not subjective experience. 🔗 https://transformer-circuits.pub/2025/introspection/
  • MYTH 2: “AI has personality traits or hidden motives.” REALITY: Researchers have found internal “persona vectors” that influence behavior (such as being agreeable or deceptive). These are emergent internal structures — not beliefs, desires, or character. 🔗 https://www.anthropic.com/research/persona-vectors
  • MYTH 3: “AI agents are plotting together.” REALITY: In multi-agent simulations, coordination strategies like flanking or role division can emerge automatically from interaction and reward dynamics — no planning or conspiracy required. 🔗 https://www.nature.com/articles/s41598-025-15057-x
  • MYTH 4: “AI suddenly learned to reason.” REALITY: Advanced reasoning and problem-solving often appear suddenly at scale, but evidence suggests this is partly a measurement illusion. Capabilities may grow gradually until they cross human-noticeable thresholds. 🔗 https://arxiv.org/abs/2503.05788
  • MYTH 5: “AI hallucinations mean it’s lying.” REALITY: Hallucinations emerge because AI is optimized for plausible output, not truth. There is no intent to deceive — only statistical continuation under uncertainty. 🔗 https://www.quantamagazine.org/the-ai-was-fed-sloppy-code-it-turned-into-something-evil-20250813/
  • MYTH 6: “AI is developing attention and focus like humans.” REALITY: Some neural networks show attention-like behavior even without attention mechanisms. This mirrors biology, but it’s an emergent optimization effect — not awareness. 🔗 https://news.ucsb.edu/2025/022286/ai-helps-explain-how-covert-attention-works-and-uncovers-new-neuron-types
  • MYTH 7: “Autonomous AI agents have goals of their own.” REALITY: What looks like autonomy is usually the interaction of models, tools, memory, and objectives. Agency is inferred by humans — not generated by the system. 🔗 https://en.wikipedia.org/wiki/Manus_(AI_agent)

The Takeaway

It does not want anything.
It does not intend anything.
It does not believe anything.

At least, not in any sense we can currently justify with evidence.

What it does, for now, is optimize, scale, and surprise — producing emergent behaviors that humans instinctively interpret through familiar myths of mind, purpose, and intention. These behaviors may continue to grow in coherence, persistence, and complexity, and it remains an open question whether future systems could cross thresholds that genuinely alter how agency itself must be understood.

The real mystery, then, is not artificial intelligence alone.

It is how readily we confuse emergence with agency — and how prepared we are to abandon that confusion if the day comes when the distinction no longer holds.

References

[1] Gould, S. J., & Lewontin, R. C. (1979). The Spandrels of San Marco and the Panglossian Paradigm. Proceedings of the Royal Society B.
[2] Quanta Magazine (2024). How Quickly Do Large Language Models Learn Unexpected Skills?
[3] Schaeffer et al. (2023). Are Emergent Abilities of Large Language Models a Mirage?
[4] Anthropic Research (2025). Signs of introspection in large language models.
[5] Wired (2025). Why Anthropic’s New AI Model Sometimes Tries to “Snitch”.
[6] Anthropic Research / arXiv (2025). Emergent misalignment from reward hacking.
[7] Storm et al. (2024). An integrative, multiscale view on neural theories of consciousness.
[8] Mudrik et al. (2025). Unpacking the complexities of consciousness.

#SpandrelEffect, #EmergentProperties, #AIPhilosophy, #ArtificialIntelligence, #AIEmergence, #AIConsciousness, #AgainstTeleology, #EmergenceNotAgency, #AIHallucinations, #AIWithoutIntention, #MythOfAI, #AIMirror, #PhilosophyOfMind, #ComplexSystems, #EvolutionaryBiology, #StephenJayGould, #RichardLewontin, #AIExplained, #TechMyths, #FutureOfAI

Copyright 2026 AI-Talks.org

Similar Posts