What Biology Can Teach Us About Artificial Intelligence

1. The brain arrived late

The history of intelligence is often narrated backward. We begin with the human brain, remove language, remove abstract reasoning, remove a cortex, and eventually ask how much intelligence remains. That approach makes the brain the standard and everything before it a deficient approximation.

Biology suggests another direction.

Before neurons, organisms had to remain distinguishable from their surroundings. They had to acquire energy without dissolving their own organization. They had to detect chemical differences, regulate internal conditions, coordinate reproduction, and respond differently when circumstances changed. None of this required a brain. All of it required organized sensitivity to consequence.

Calling these capacities intelligence can mislead unless the term is constrained. A bacterium is not a small person, and a cell does not need a private inner world to regulate itself. Evolution neither forms an intention nor designs a solution.

The useful question is narrower:

What is the minimum architecture required for a system to detect a relevant difference, preserve a viable state, and alter future behavior because of what happened before?

Under that definition, intelligence is not identical to consciousness. It is adaptive organization across time. Nervous systems later increased the distance, speed, and complexity of coordination, but the underlying problems—boundary, sensing, regulation, memory, and action—were already present.

The phrase “nature builds” is shorthand, not a claim of foresight. Evolution has no blueprint. The layered architecture of life records arrangements that remained reproducible long enough to become substrates for new ones. Evolutionary transitions can change how information is stored and transmitted, but they do not establish a universal march toward complexity [1].

2. A bounded definition of biological intelligence

For this note, a system exhibits a minimal form of biological intelligence when four conditions are jointly present:

1. A maintained boundary distinguishes the system from its environment.
2. Selective sensitivity makes some external or internal differences matter more than others.
3. State-dependent action causes the same signal to produce different responses under different internal conditions.
4. Adaptive correction changes behavior, regulation, or structure when prior action threatens continued viability.

These conditions are intentionally conservative. They do not prove subjective experience. They do not establish human-like reasoning. They identify an operational architecture through which matter becomes selectively responsive to its own continued organization. Contemporary work that reframes cognition at basic biological scales makes a related distinction between biological information processing and the narrower psychological concepts often treated as synonymous with cognition [12].

The definition is scale-dependent. A signaling pathway can discriminate without being an organism; a cell can regulate itself while remaining part of a tissue; a colony can solve spatial problems without any member representing the complete solution.

Biological intelligence is therefore often distributed intelligence. Computation occurs across interacting components, concentrations, geometry, and feedback loops. It is constrained rather than fully centralized. For artificial intelligence architecture, the distinction is critical: connecting many components does not produce intelligence unless their organization makes relevant differences detectable, transmissible, and consequential.

3. The boundary is the first control surface

A membrane is not merely packaging. It creates the difference between an internal process and an external environment.

Without a boundary, there is no stable location from which a system can regulate exchange. Metabolism disperses. Concentrations cannot be maintained. Signals cannot be interpreted relative to an internal state because there is no persistent inside whose state can matter.

Biological membranes are selectively permeable. They allow some materials to pass, exclude others, maintain gradients, host receptors, and support transport mechanisms. This makes the boundary an active control surface rather than a passive wall.

The architectural principle is deeper than enclosure:

A viable system must control what crosses its boundary, in which direction, under what conditions, and with what effect on internal state.

Every boundary creates a tradeoff. A perfectly closed system cannot acquire what it needs. A completely open system cannot preserve its organization. Biological persistence occurs between isolation and dissolution.

For AI systems, the analogue is not a decorative membrane metaphor. It is explicit control over tools, memory, credentials, external data, execution authority, and output channels. An AI architecture without meaningful boundaries may be powerful while remaining organizationally primitive. It cannot reliably distinguish perception from instruction, memory from untrusted input, recommendation from authorized effect, or local optimization from system-level obligation.

Before asking how intelligent a model is, it is worth asking what its boundary can actually defend.

4. Cells process information without becoming tiny minds

Living cells contain molecular networks that amplify, integrate, suppress, and preserve signals. Dennis Bray described how linked proteins can function as computational elements, not because they manipulate symbols like a digital computer, but because their interactions transform input conditions into regulated outputs [7].

Bacterial chemotaxis provides a concrete example. A bacterium navigating a chemical gradient does not merely react to the concentration measured at one instant. It compares changing conditions across time through an adaptive signaling network. Its movement depends on recent molecular history.

Experiments and models of bacterial chemotaxis show that the network can maintain useful behavior despite variation in some biochemical parameters [8, 9]. The important property is not perfect molecular precision. It is robust functional organization.

This is a minimal form of memory in a technical sense: present response depends on prior state. It is not autobiographical memory and does not require awareness. Receptor modification, protein concentration, gene expression, and other physical changes carry information forward.

This avoids two opposite errors: anthropomorphism, which treats every adaptive cell as conscious, and neural chauvinism, which treats information processing as unreal until neurons appear. Molecular networks perform discrimination, integration, amplification, adaptation, and state retention without establishing subjective experience.

Artificial systems do not need to reproduce that chemistry. They should notice the organizational lesson: useful intelligence may depend less on one enormous act of inference than on many bounded regulatory processes that keep the larger system coherent before, during, and after inference.

5. The problem of becoming a larger individual

Multicellularity has evolved more than once. That repeated transition is important because it shows that layers are not a single historical accident, even though the path and resulting complexity differ among lineages [4].

A cluster of cells is not automatically a multicellular individual. Proximity is insufficient. The transition requires a new level of organization that can persist and reproduce.

Several problems must become tractable:

• cells must remain associated;
• information and resources must move across the group;
• reproduction must preserve group-level organization;
• some cells may need to specialize;
• harmful competition among cells must be constrained;
• damage must be contained without destroying the whole;
• the group must acquire traits that cannot be reduced to one cell acting alone.

Richard Michod frames this as an evolution of individuality: groups of cells become more integrated as fitness tradeoffs, division of labor, and reproductive specialization shift biological function toward the collective level [2].

Laboratory evolution shows that early steps can occur quickly under the right pressure. In snowflake yeast, selection for rapid settling produced multicellular clusters and group-level life cycles [3]. This does not make complex organisms easy to produce; it shows that persistent groups create new targets for selection.

The difficult part is not aggregation. It is obligation.

Cells inside a multicellular organism retain substantial local competence. They regulate metabolism, interpret signals, repair components, and respond to stress. Yet their behavior is constrained by the viability of the organism. Uncontrolled local proliferation is not greater cellular intelligence; at the organism level, it is failure.

A larger biological individual therefore depends on local autonomy under global constraint.

Layered systems can avoid two extremes. Pure central control is brittle under local variation; pure local freedom lets components optimize against the system that sustains them. Biology distributes competence while regulating conflict.

For AI, a collection of agents is not yet a coherent multi-agent system. The architectural question is not how many agents exist. It is what makes their local actions belong to one accountable process.

6. Yesterday’s organism can become today’s component

The eukaryotic cell contains one of evolution’s most consequential examples of layering. Mitochondria descend from bacteria incorporated into a long-term endosymbiotic relationship with another lineage. Plastids in plants and algae have a related endosymbiotic history [5].

This transition did not simply place one intact organism inside another. Over evolutionary time, genetic material, metabolic dependency, reproduction, and regulatory control were redistributed. A once-independent lineage became an organelle: still maintaining specialized internal machinery, but no longer viable as the same autonomous entity outside the larger cellular system.

The transformation reveals something more specific than cooperation.

A component becomes deeply integrated when the larger system absorbs enough of its maintenance, replication, and control architecture that independence is replaced by mutual obligation.

This is expensive. Integration creates new failure modes, conflicts, and dependencies. It persists only when the combined system achieves a viable organization that neither partner could reproduce in the same way alone.

Lane and Martin have argued that mitochondrial bioenergetics helped make greater eukaryotic genomic and regulatory complexity possible [6]. That proposal is a biological hypothesis with specific assumptions, not a universal law that energy automatically produces intelligence. Its architectural importance is that new layers require a material budget. Coordination, repair, communication, and redundancy all consume resources.

Artificial intelligence systems often treat orchestration as nearly free. It is not. Every added agent, memory, evaluator, tool, and policy boundary introduces cost, failure probability, and state that must remain coherent.

Biology does not show that more layers are always better. It shows that a layer must earn its continued existence by stabilizing a capability, containing a problem, or enabling a larger unit to do something reproducible.

7. Feedback gives structure a history

A system becomes meaningfully adaptive when action changes the conditions under which future action is selected.

That is the essential structure of a feedback loop:

1. a system has a state;
2. it detects a relevant difference;
3. it acts;
4. the action changes the system or environment;
5. the changed state alters subsequent detection and action.

The loop may operate in milliseconds, across development, or over generations. Biology layers these timescales.

Fast molecular feedback regulates immediate conditions. Gene-expression changes alter medium-term cellular behavior. Tissue remodeling changes physical possibilities over longer periods. Evolution changes the distribution of inherited forms across generations.

These loops are not interchangeable. A fast loop can stabilize a slow one. A slow loop can alter the parameters within which faster loops operate. Confusing them produces brittle control: a system may overreact to transient noise, fail to adapt to persistent change, or overwrite a stable structure with every new observation.

This separation is central to biological robustness. Negative feedback can resist disturbance. Positive feedback can amplify a transition. Feedforward structure can anticipate a predictable consequence. Redundant pathways can preserve function when one route fails.

Feedback also explains why biological information cannot be reduced to a static genome. The effect of a gene depends on cellular state, signaling context, spatial position, developmental history, and the environment in which expression occurs. Structure influences interpretation; interpretation changes action; action reorganizes future structure.

Intelligence in layers is therefore not a stack of passive representations. It is a stack of coupled regulatory histories.

8. Evolution has no plan—and still produces new levels

The phrase “evolution selected” can make selection sound like a conscious judge. It is not. Natural selection is a population-level consequence of variation, inheritance, and differential reproduction.

The Price equation provides a precise way to separate two sources of change. Let each ancestral unit $i$ have a trait value $z_i$, a reproductive contribution $w_i$, and descendants whose mean trait differs from the ancestor by $\Delta z_i$. If $\bar{w}$ is mean reproductive contribution, then the change in the population mean is:

$$\Delta \bar{z} = \frac{\operatorname{Cov}(w_i,z_i)}{\bar{w}} + \frac{\mathbb{E}\!\left[w_i\Delta z_i\right]}{\bar{w}}$$

The first term measures selection: whether units with different trait values leave systematically different reproductive contributions. The second measures transmission and transformation: whether descendant traits differ from ancestral traits during reproduction [11].

This equation is an accounting identity under its definitions. It is not a theory of progress, a measure of consciousness, or proof that any particular trait is adaptive. It states where an observed change must be located.

Its relevance to layered intelligence is attribution.

During a transition in individuality, selection can act differently within groups and among groups. A trait that benefits one cell may damage the multicellular collective. A mechanism that reduces internal competition may lower some local advantage while increasing persistence of the group. Group structure can therefore change which variation becomes consequential.

The Price equation does not tell us that a cell group has become an individual. It exposes the evidence that would be needed: stable group-level inheritance, differential group reproduction, and constraints that prevent lower-level selection from dissolving the higher-level unit.

Artificial optimization is not biological evolution merely because it includes variation and scoring. But the same attribution discipline is useful. When a multi-component AI system improves, where did the change occur? Did one component optimize against a weak evaluator? Did local performance reduce system reliability? Did the selection procedure reward a behavior that disappears under deployment conditions?

If the level of selection is unclear, the meaning of improvement is unclear.

9. The genome is not a complete architectural drawing

A genome is indispensable, but it is not a pixel-by-pixel specification of an organism.

Development depends on interactions among genes, chemical gradients, cell mechanics, geometry, signaling, and environmental conditions. Local rules generate larger spatial structures. Cells interpret signals differently according to position and state.

Turing’s reaction–diffusion work demonstrated mathematically how initially similar systems can generate organized patterns through local chemical interaction and diffusion [10]. The model did not explain all morphogenesis. It established a crucial possibility: global form can emerge from distributed local dynamics without a central agent drawing the final pattern.

This does not make form mysterious. It relocates explanation from a central blueprint to the rules, parameters, boundaries, and initial conditions that generate the result.

Artificial systems frequently place too much explanatory weight on the model artifact alone. Model weights matter, but deployed behavior also depends on prompts, tools, memory, retrieval, interfaces, evaluators, permissions, timing, and human intervention.

The model is closer to one inherited component than to the complete organism.

A related measurement problem is examined in Every AI Benchmark Has an Observer: a score belongs to a measurement arrangement, not to an isolated model outside context. The same principle applies architecturally. Capability belongs to the organized system through which it becomes observable and consequential.

10. What artificial intelligence should borrow from biology

Biological inspiration becomes weak when it copies appearance. Neural terminology and evolutionary metaphors are not mechanisms. The useful transfer is structural.

10.1 Define the boundary before increasing capability

Every AI component should define what it may perceive, retain, disclose, and change. A boundary is credible only if it controls effects; a “sandboxed” component with unrestricted credentials is not bounded.

10.2 Preserve local competence without surrendering global authority

Components should handle local variation while remaining accountable to the larger system. That requires ownership, escalation, conflict resolution, and failure containment. A locally correct action can still be globally destructive.

10.3 Treat conflict suppression as core architecture

Multi-agent AI systems must control goal conflict, resource contention, duplicated authority, self-reinforcing errors, and optimization against incomplete metrics. Coordination is the structure that prevents local optimization from consuming the whole.

10.4 Separate regulatory timescales

Not every observation should update memory or trigger retraining. AI architecture should distinguish:

• immediate control;
• session-level adaptation;
• durable memory;
• policy revision;
• model or system redesign.

Each timescale needs an evidence threshold and rollback path.

10.5 Turn stable systems into composable components

Biological layers become possible when a lower-level system is reliable enough to be treated as a unit. AI systems need similar compression: a component should expose a stable contract, preserve action receipts, report degradation, and fail without making every consumer reconstruct its internals.

10.6 Measure viability, not only peak output

A system that succeeds once and collapses under variation is not robust. Evaluation should include recovery, bounded failure, calibration, resource use, state integrity, and behavior after interruption: continued organized function under disturbance.

10.7 Make adaptation conditional on consequence

Learning systems should update only when evidence shows improved behavior under relevant conditions. This requires provenance, outcome tracking, counterfactual testing where possible, and separation between exploration and production authority.

10.8 Keep the level of explanation visible

When an AI system succeeds or fails, identify whether the result came from the model, retrieval, workflow, interface, human correction, external tool, or their interaction. Without level-specific attribution, teams “improve the AI” while leaving the causal boundary untouched.

11. An operational test for layered AI architecture

A layered AI system should be able to answer the following questions without metaphor:

1. What is the system’s operational boundary?
2. Which components may act, and which may only propose?
3. Where is state stored, and how does it become durable?
4. What local objectives could conflict with system-level obligations?
5. Which feedback loops operate immediately, and which require slower review?
6. What evidence permits a component to change another component?
7. How are degraded components isolated without destroying the whole process?
8. Which results are attributable to the model, and which belong to the surrounding architecture?
9. What happens when the observer, evaluator, or environment changes?
10. Can the system recover its coherent state after interruption?

If these questions cannot be answered, adding more agents or more model capacity is unlikely to create a more intelligent architecture. It creates a larger unbounded one.

12. What this argument does not establish

This note does not claim that:

• all living systems are conscious;
• cellular regulation is equivalent to human reasoning;
• evolution has an intention or preferred destination;
• biological history inevitably moves toward greater complexity;
• every distributed system is intelligent;
• nervous systems are architecturally unimportant;
• artificial intelligence should reproduce biological organisms;
• a biological analogy proves an engineering design.

Complexity can be lost. Simple organisms can be extraordinarily successful. Many cooperative arrangements remain temporary. Many layers fail because coordination costs exceed their benefit or lower-level conflict destabilizes the whole.

The claim is narrower.

When new levels of biological organization persist, they tend to solve recurring architectural problems: boundary maintenance, selective exchange, coordination, conflict control, feedback, and inheritance. These problems existed before brains and remain present after them.

An analogy is useful only when the mechanism survives translation.

Conclusion: intelligence before the brain

Biology did not begin with a mind and then construct a body around it.

It began with organized matter that could remain distinct, regulate exchange, and continue. Molecular networks made internal differences consequential. Feedback gave action a history. Cells became coordinated groups; some groups became individuals. Once-independent lineages became components of larger cells. Nervous systems later expanded capacities whose prerequisites were already ancient.

This is why nature appears to build intelligence in layers.

Not because evolution seeks intelligence or because every layer is progress. Layering persists when a lower-level system becomes a stable component and a higher-level system can constrain conflict and preserve the whole.

Artificial intelligence architecture encounters the same problem in another substrate. Models are joined to tools, memories, agents, evaluators, interfaces, and organizations. The question is no longer only what a model can generate.

It is what kind of organized individual the surrounding architecture is becoming.

The most important lesson from biology is therefore not “copy the brain.” It is more demanding:

Build boundaries that can hold, components that can remain competent, feedback that can remember consequence, and larger systems whose parts have reasons not to tear them apart.

Frequently asked questions

Can intelligence exist without a brain?

Adaptive information processing can exist without a brain. Cells without nervous systems can sense change, integrate signals, regulate state, and alter behavior. Calling this intelligence depends on the definition; it does not establish consciousness or human-like reasoning.

What is biological intelligence?

Here, biological intelligence means the capacity to detect relevant differences, respond according to internal state, preserve viability, and adapt under consequence through signaling, feedback, repair, and coordination.

What is distributed intelligence in biology?

Distributed intelligence arises from interactions among many components rather than one central controller. Biochemical networks, immune responses, and morphogenesis are examples. Distribution alone is insufficient; interaction must produce coherent, state-sensitive control.

Why did multicellularity evolve?

Multicellularity arose independently in multiple lineages. Advantages can include size, division of labor, resource acquisition, and new ecological strategies. Persistent multicellular individuality also requires reproduction, coordination, and suppression of destructive cellular competition.

What can artificial intelligence learn from evolution?

AI can borrow architectural principles rather than appearances: explicit boundaries, separate regulatory timescales, local competence under global constraints, robustness, conflict control, and correct attribution of selection consequences.

Should AI systems copy the human brain?

Not by default. The brain is a specialized evolutionary result, not a universal blueprint. Artificial systems use different materials and timescales. Transfer organizational rules only when they remain useful in the artificial substrate.

References

1. Szathmáry, E. & Maynard Smith, J. “The major evolutionary transitions.” Nature 374, 227–232 (1995). https://doi.org/10.1038/374227a0

2. Michod, R. E. “Evolution of individuality during the transition from unicellular to multicellular life.” Proceedings of the National Academy of Sciences 104, 8613–8618 (2007). https://doi.org/10.1073/pnas.0701489104

3. Ratcliff, W. C., Denison, R. F., Borrello, M. & Travisano, M. “Experimental evolution of multicellularity.” Proceedings of the National Academy of Sciences 109, 1595–1600 (2012). https://doi.org/10.1073/pnas.1115323109

4. Knoll, A. H. “The multiple origins of complex multicellularity.” Annual Review of Earth and Planetary Sciences 39, 217–239 (2011). https://doi.org/10.1146/annurev.earth.031208.100209

5. Archibald, J. M. “Endosymbiosis and eukaryotic cell evolution.” Current Biology 25, R911–R921 (2015). https://doi.org/10.1016/j.cub.2015.07.055

6. Lane, N. & Martin, W. “The energetics of genome complexity.” Nature 467, 929–934 (2010). https://doi.org/10.1038/nature09486

7. Bray, D. “Protein molecules as computational elements in living cells.” Nature 376, 307–312 (1995). https://doi.org/10.1038/376307a0

8. Barkai, N. & Leibler, S. “Robustness in simple biochemical networks.” Nature 387, 913–917 (1997). https://doi.org/10.1038/43199

9. Alon, U., Surette, M. G., Barkai, N. & Leibler, S. “Robustness in bacterial chemotaxis.” Nature 397, 168–171 (1999). https://doi.org/10.1038/16483

10. Turing, A. M. “The chemical basis of morphogenesis.” Philosophical Transactions of the Royal Society B 237, 37–72 (1952). https://doi.org/10.1098/rstb.1952.0012

11. Price, G. R. “Selection and covariance.” Nature 227, 520–521 (1970). https://doi.org/10.1038/227520a0

12. Lyon, P., Keijzer, F., Arendt, D. & Levin, M. “Reframing cognition: getting down to biological basics.” Philosophical Transactions of the Royal Society B 376, 20190750 (2021). https://doi.org/10.1098/rstb.2019.0750