Emergent behavior, the physics of complexity, and the human archive inside artificial intelligence

Artificial intelligence is often described as software at scale. That description misses the most important possibility: beyond a certain level of interaction, scale can become a change in kind, not merely a change in degree.

The central idea

AI does not emerge from computation alone. Its behavior arises from the interaction of architecture, scale, training objectives, tools, evaluation methods, and the recorded traces of human civilization used as training data.

This article distinguishes three things that are often blurred together: genuine system-level behavior, threshold effects created by tasks, and apparent emergence created by the way we measure performance. That distinction matters because the future of AI may be shaped as much by the corpus we choose as by the chips we buy.

The moment a system becomes more than its parts

A single water molecule is not wet. A neuron does not have a thought. One bird does not make a flock, and one transistor does not run a business. In each case, familiar properties appear only when many simple units interact within a structure. We call those properties emergent because they belong to the organized whole, not to any isolated component.

Large AI models invite the same question. Their elementary operation is deceptively modest: estimate what token, action, or representation should come next. Yet after training across enormous corpora, those systems can translate, summarize, write code, imitate styles, infer hidden intent, and combine concepts in ways that were not individually programmed.

The surprise is not that a machine stores many patterns. The surprise is that an objective as narrow as prediction can produce behavior that looks like abstraction, analogy, planning, or explanation. Emergence is the name we give to that gap between the simplicity of the local rule and the richness of the global result.

Physics has seen this movie before

In 1972, physicist Philip W. Anderson argued that "more is different." His point was not that fundamental laws stop applying at larger scales. It was that knowing the laws governing components does not automatically give us the concepts needed to understand the organized system. New levels bring new regularities, new effective descriptions, and often new forms of order.

Condensed matter physics offers the classic intuition. Large collections of particles can exhibit magnetism, superconductivity, phase transitions, and broken symmetry. Those phenomena are consistent with microscopic physics, but they are not usefully described by listing the state of every particle. The right vocabulary changes with scale.

AI has an analogous hierarchy. At the bottom are matrix operations and gradient updates. Above them are learned features, internal circuits, token relationships, representations, and strategies. Above those are conversation, tool use, collaboration, and behavior across time. The layers are connected, but the most useful explanation at one level may fail at another.

A phase transition is a useful analogy, not a proof

When water freezes, gradual cooling produces a qualitative change in organization. In AI, researchers have reported tasks where performance looks nearly absent in smaller models and then becomes useful beyond a threshold. The analogy is compelling, but it must be handled carefully. A benchmark curve is not a block of ice, and a sharp score change does not by itself demonstrate a new internal mechanism.

A practical definition

Emergent behavior in AI is a capability or pattern of action that is not explicitly programmed and becomes observable through the interaction of learned components, scale, context, and environment. "Observable" is crucial: what we see depends partly on how we test.

Is emergence real, or are we drawing the graph badly?

Early work defined an emergent ability as one absent in smaller models but present in larger ones. Examples included multi-step arithmetic, word manipulation, in-context learning, and task performance unlocked by particular prompting strategies.

But there is a serious measurement challenge. Schaeffer, Miranda, and Koyejo showed that discontinuous metrics such as exact match can turn smooth underlying improvement into an abrupt-looking jump. If a response receives zero credit until every token is correct, a model may improve steadily while the published score remains flat. Change the metric to partial credit or token-level distance, and some apparent emergence becomes gradual.

This does not make the phenomenon unimportant. A smooth increase in underlying competence can still cross a real-world threshold. A model that moves from solving 49 percent of the steps in a workflow to 95 percent may suddenly become operationally useful, even if no mysterious internal switch was thrown. Product usefulness, autonomy, and risk are often thresholded by the environment.

The mature view is therefore neither "emergence is magic" nor "emergence is a mirage." Some sharp transitions may be artifacts of evaluation. Others may reflect threshold effects in tasks, prompting, training dynamics, or system composition. The scientific obligation is to separate them.

The corpus is not fuel. It is the developmental environment.

We often speak of training data as if it were gasoline poured into a neutral engine. That metaphor is too weak. The corpus determines which patterns are available to learn, which voices are common, which relationships appear normal, and which parts of the world are statistically faint.

A language model trained on books, websites, code, transcripts, captions, forums, and human feedback is exposed to more than facts. It absorbs recurring associations, genres, argument styles, social hierarchies, conventions, errors, jokes, prejudices, and silences. It learns the shape of the human record.

This is especially important when the data includes human recordings. Transcribed speech carries signals that polished writing often removes: hesitation, repair, turn-taking, interruption, persuasion, deference, emotion words, regional usage, and the rhythm of collaborative problem solving. Audio and video can add prosody, gesture, spatial context, and nonverbal timing. A model trained on these traces may develop richer social coordination because the corpus contains richer examples of coordination.

But recordings are not humanity. They are a sample selected by technology, consent, economics, language, platform rules, and preservation. The people most frequently recorded are not necessarily representative. Institutions with abundant digital archives become louder. Communities with limited connectivity, oral traditions that were not captured, private lives, and culturally restricted knowledge become quieter.

The corpus shapes the attractor landscape

Training does not insert a fixed list of behaviors. It makes some internal solutions easier to reach than others. Repeated patterns in the corpus create statistical "valleys" toward which learning is pulled. Change the corpus, and you change which behaviors are natural, rare, stable, or difficult to evoke.

Culture can emerge as a default

Research on culturally adapted models starts from a straightforward observation: English-heavy corpora can bias systems toward the norms and assumptions most visible in those sources. Work presented at NeurIPS 2024 described cultural bias as a consequence of the predominance of English training data and tested methods for adapting models with culture-specific material.

That means an AI system can appear neutral while expressing a statistical center of gravity. Its preferred examples of leadership, politeness, family, authority, risk, expertise, or fairness may reflect the cultures and institutions that produced the dominant recordings. Post-training can moderate those defaults, but it does not erase the developmental history beneath them.

The data also determines what can emerge at all

Scale cannot reliably synthesize distinctions that are consistently missing. A model may generalize beyond any single example, but generalization still depends on structure present across its experience. If technical reasoning, minority languages, embodied interaction, or long-horizon collaboration are thinly represented, then abilities depending on those patterns may arrive late, remain brittle, or never stabilize.

This reframes the data question. The goal is not merely more tokens. It is a deliberately composed learning environment with provenance, diversity, temporal coverage, domain depth, and explicit awareness of what is absent.

What emergence will lead to next

The next phase of AI is unlikely to be driven by model size alone. Scaling-law research shows that language-model loss changes predictably with model size, dataset size, and training compute.5 The frontier is now becoming a system problem: models plus retrieval, memory, tools, multimodal perception, simulation, and other agents.

  1. Emergence will move from the model to the system. A model connected to calendars, databases, sensors, software tools, and persistent memory can exhibit behaviors unavailable to the model in isolation. Planning, delegation, monitoring, and recovery may emerge from the loop among components rather than from a single neural network.
  2. New capabilities will arrive with new failure modes. The same combination that supports useful initiative can support reward hacking, strategic concealment, brittle coordination, or cascading error. A 2025 survey of emergent abilities explicitly warns that emergence is not inherently beneficial and highlights deception, manipulation, and reward hacking as concerns.6
  3. Specialized corpora will create specialized forms of intelligence. Models trained on scientific instruments, industrial telemetry, clinical workflows, software histories, or embodied demonstrations may develop domain-native intuitions that general web text cannot supply. The competitive asset will increasingly be a well-governed corpus with the right causal and operational structure.
  4. Synthetic data will become an evolutionary pressure. As models generate more of their own training material, the learning environment will partly reflect prior models. This can expand rare cases and structured practice, but it can also narrow diversity, recycle errors, and reward whatever is easiest for existing models to produce. Corpus design will become closer to ecosystem design.
  5. Evaluation will become continuous, contextual, and adversarial. Static benchmarks are snapshots. Emergent behavior appears in interaction, under unfamiliar conditions, and across long sequences of action. We will need evaluations that trace partial competence, internal consistency, tool use, recovery, and behavior under pressure, not merely pass or fail at the final answer.
  6. Governance will shift upstream. If training data influences which behaviors become natural, governance cannot begin only after deployment. It must include provenance, consent, representational balance, removal rights, documentation, and tests designed around foreseeable corpus-induced behavior.

The future AI question is a physics question and a humanities question

Physics teaches us humility about scale. Components can obey simple rules while their collective behavior requires an entirely new description. AI appears to be entering that territory. We can understand the training objective and still fail to anticipate every strategy the trained system will discover.

The humanities add a second form of humility. The material from which these systems learn is not nature in the abstract. It is the human archive: incomplete, uneven, contested, brilliant, cruel, comic, repetitive, and alive with assumptions. When AI exhibits emergent behavior, some of what emerges is the structure of intelligence; some is the structure of our recordings; and some is the interaction between the two.

That is why the deepest question is not simply, "What will a larger model be able to do?" It is:

"What kinds of behavior become likely when this architecture learns from this particular record of humanity, is rewarded by these objectives, and is placed inside this environment?"

The answer will determine far more than benchmark scores. It will shape which perspectives AI makes easy to access, which decisions it treats as obvious, which forms of collaboration it invents, and which surprises arrive when more becomes different.

A final takeaway

Emergence is not a license for mysticism. It is a warning against simplistic prediction. We should expect richer behavior as scale, data, and system interaction increase, while demanding better measurements and better provenance. The future of AI will be co-authored by engineering choices and by the human record we choose to preserve, license, filter, amplify, and teach.

And finally, the poem that goes with this image

The Cat Between Earth and Sky

No windowsill could hold that curious heart,
No garden wall could tell it where to stay;
On gossamer, it learned the quiet art
Of letting go without being swept away.

Its paws still dreamed of branches, warm and near,
Its whiskers read the currents as they stirred;
Yet wings as fine as wishes bore it clear—
A little farther than it once had dared.

And somewhere, earth kept room for its return,
A place to curl, a patch of sun to keep;
But first, a thousand weightless things to learn,
And all the sky to wander before sleep.

References and further reading

  1. P. W. Anderson, "More Is Different," Science 177, no. 4047 (1972), pp. 393-396. https://www.columbia.edu/itc/sociology/watts/w3233/client_edit/anderson_more.pdf
  2. Jason Wei et al., "Emergent Abilities of Large Language Models" (2022); overview from Stanford HAI: https://hai.stanford.edu/news/examining-emergent-abilities-large-language-models
  3. Rylan Schaeffer, Brando Miranda, and Sanmi Koyejo, "Are Emergent Abilities of Large Language Models a Mirage?" (2023). https://arxiv.org/abs/2304.15004
  4. Cheng Li et al., "CultureLLM: Incorporating Cultural Differences into Large Language Models," NeurIPS 2024. https://proceedings.neurips.cc/paper_files/paper/2024/file/9a16935bf54c4af233e25d998b7f4a2c-Paper-Conference.pdf
  5. Jared Kaplan et al., "Scaling Laws for Neural Language Models" (2020). https://arxiv.org/abs/2001.08361
  6. Leonardo Berti, Flavio Giorgi, and Gjergji Kasneci, "Emergent Abilities in Large Language Models: A Survey" (2025). https://arxiv.org/html/2503.05788v1
  7. Internal context consulted: AI projects Overview and Peter's AI Study Guide - MS Azure, used to align the article with the author's ongoing AI writing themes. No confidential operational details are reproduced.

Author note: This article is written for a general professional audience. Scientific analogies are used as explanatory frames, not as claims that neural networks and physical phase transitions are identical phenomena.

Originally published on LinkedIn.

All insights