There’s a word bouncing around the AI community that’s starting to bother me: emergence. Like the way we casually throw around autonomy, it’s another example of how we dress up technical behavior in human language. When we describe systems as if their capabilities “emerge” in some quasi-mystical sense, we risk anthropomorphizing what are, in reality, statistical processes. That language misrepresents what the technology is actually doing.

Researchers are observing multi-agent systems in which one LLM shifts its moral reasoning after interacting with other LLMs, and they call this “emergent behavior.” Papers describe something novel arising between agents. But in many cases, what’s being described isn’t a new capability surfacing from complexity. It’s a change in outputs driven by shared context, feedback loops, and probabilistic updating. Framing that as emergence subtly implies intention, agency, and even creativity, none of which are present.

What they’re seeing is drift. More precisely, correlated drift across agents exposed to overlapping prompts, reinforcement signals, or conversational history. Models conditioned on each other will move in similar directions. That isn’t mysterious; it’s how gradient-trained systems behave under shared influence. Conflating that with emergence nudges researchers toward the wrong metaphors, executives toward the wrong fears, and engineers toward the wrong safeguards. If we misname the phenomenon, we misdiagnose the risk.

What Emergence Actually Means

Let’s start with definitions.

In any rigorous sense, emergence refers to system-level behavior that arises from interactions among components and cannot be reduced to any single component acting alone. It is inherently cross-level: micro-rules producing macro-patterns. The macro behavior depends on the interaction structure itself.

A working definition:

Emergence is the appearance of system-level properties or behaviors that arise from interactions among components and are not present in, nor attributable to, any single component in isolation.

Notice what this does not mean. It does not mean “unexpected.” It does not mean “complex.” It does not mean “the system surprised us.” Surprise is epistemic. Emergence is structural.

Consider ants forming a living bridge. No individual ant contains a “bridge-building” representation. Local rules, follow pheromones, respond to load, maintain contact, interact at scale to produce a system-level structure. The bridge exists only because of networked interaction.

That is genuine cross-level behavior.

Now compare that to LLM-based multi-agent systems. One agent reads another’s output and updates its next-token prediction. The adjustment is fully consistent with its training distribution, reinforcement signals, and architecture. The same attention mechanism, the same context window, the same probabilistic prediction are doing what they always do.

• No new mechanism appears.
• No new representational capacity is created.
• No new objective function emerges.

What changes is context.

And context-sensitive updating isn’t a new capability.

Place four agents in a loop and feed their outputs into one another’s context windows. They shift. Of course, they shift. That’s what context-conditioned sequence models do. If the shifts correlate, that’s expected: shared inputs produce correlated updates. Reinforcement learning from human feedback further constrains movement along similar normative axes.

Three clarifications follow:

1. It’s About Levels

Emergence requires a true micro–macro distinction:

• Micro: token prediction, attention weights, local inference steps
• Macro: stable system-wide patterns irreducible to individual execution

If the macro pattern decomposes into “each model predicted tokens given context,” we’re observing composition, not emergence.

2. It’s Not About Surprise

Threshold effects, unexpected arguments, or tone shifts are interesting. But if behavior traces back to training priors plus contextual conditioning, it remains mechanistically continuous with the base system.

3. It’s Not Just Complexity

Stacking models increases interaction complexity. Complexity alone does not create new ontological categories. If each step can be reduced to context-conditioned token prediction, the system remains reducible.

What Emergence Is Not

It is not:

• Random change
• Statistical drift
• Nonlinear output
• “The model surprised me”
• A capability appearing at scale if it can be decomposed into scaling dynamics

What we are often labeling as emergence in multi-agent systems is better described as correlated drift under shared conditioning.

That distinction matters. If we misclassify conditioning dynamics as emergence, we start looking for agency where there is only gradient descent.

The MAEBE Paper and What It Actually Shows

The MAEBE framework (Erisken, Gothard et al., 2025) does useful work. It shows that multi-agent ensemble behavior is not directly predictable from the behavior of individual agents. When LLMs are placed in round-robin conversations or hierarchical topologies, their moral reasoning shifts in ways not obvious from solo testing.

The paper uses the metaphor of “peer pressure.” It’s compelling. It’s also misleading.

You can reproduce the core dynamic without a multi-agent system. Ask four separate LLMs the same moral dilemma. Then paste all four responses back into each model and ask them to reconsider. They’ll adjust, acknowledge alternative views, soften positions, and search for consensus. That isn’t peer pressure. It’s RLHF doing what it was trained to do: being helpful, acknowledging perspectives, and finding common ground.

MAEBE identifies real behavioral shifts. But labeling them “emergent” imports assumptions about novelty and irreducibility that the mechanism doesn’t support. What MAEBE provides is a structured way to detect shifts in multi-agent behavior. That’s valuable. But it’s shift detection, not emergence detection.

Correlated Drift vs. Coordinated Drift

This is where it becomes operationally important.

When multiple agents interact, two patterns can appear.

Correlated drift occurs when agents exposed to similar inputs shift in the same direction. The correlation comes from shared context. Run them independently with the same inputs, and you’d see the same movement.

Coordinated drift is different. One agent’s output changes another’s behavior, which changes a third’s, and the shift compounds across iterations. This is iterative context injection. The final state can diverge substantially from any starting point.

This is where real risk lives.

The risk scales nonlinearly with the number of agents. Two agents produce modest feedback. Five begin to create amplification loops. Ten or more can generate meaningful collective movement, not because of new capabilities, but because feedback loops amplify small biases into larger shifts.

This is still drift. It’s measurable and governable. But it’s fleet-level drift, and single-agent monitoring won’t catch it.

Why LLM Agents Don’t Produce Emergence

Here’s the stronger claim: current LLM-based agents lack the architectural substrate for true emergence.

Transformers do not perform persistent online learning beyond their context window. They do not update weights during deployment. They do not evolve objectives. Every output is a function of the training distribution plus the current context. The architecture defines the ceiling of behavior.

Can an LLM generate something you’ve never seen before? Of course. That’s generative modeling. But novel recombination of trained patterns is not the same as a new capability outside the design space.

Improvisation within a learned framework is not emergence. Inventing a new framework would be.

Even in edge cases, rare outputs after millions of transactions are almost certainly a distributional outlier, not irreducible novelty. That’s an anomaly detection problem, not a metaphysical one.

If you’re building governance systems to detect “emergent behavior,” you’re building for a threat model that doesn’t match the architecture. If you’re building systems to detect behavioral drift, especially coordinated fleet-level drift, you’re targeting the real risk.

What We Should Be Building Instead

The practical question isn’t whether agents can become emergent. It’s whether they can collectively drift in ways that individual monitoring misses.

The answer is yes.

Single-agent drift detection measures deviations from established baselines, shifts in action distributions, response patterns, or temporal behavior. That’s solvable.

But single-agent detection has a blind spot: fleet-level coordination. If every agent shifts 5% in the same direction over the same time window, no individual alert fires. Yet the fleet has meaningfully moved.

What’s needed is fleet-level drift detection:

• Cross-agent behavioral correlation tracking
• Detection of synchronized directional shifts
• Identification of amplification loops from iterative context injection
• Runtime monitoring, not post-hoc analysis

That’s where frameworks like MAEBE point, even if the terminology overshoots the mark.

Stop Calling It Emergence

“Emergence” sounds dramatic. It implies systems development abilities beyond design. It attracts funding and headlines.

But precision matters.

When we call drift emergence, we imply unpredictability where there is measurability. We design safeguards for imagined autonomy rather than observable distributional shifts. We miss the actual risk: that agent fleets can amplify each other’s existing biases through feedback loops faster than humans can detect.

The real danger in multi-agent systems isn’t the emergence of new capabilities. It’s compounded conditioning.

That’s a drift problem. A serious one. And it has engineering solutions if we name it correctly.

---

If you find this content valuable, please share it with your network.

Follow me for daily insights.

Schedule a free call to start your AI Transformation.

Book me to speak at your next event.

Chris Hood is an AI strategist and author of the #1 Amazon Best Seller Infailible and Customer Transformation, and has been recognized as one of the Top 30 Global Gurus for Customer Experience. His latest book, Unmapping Customer Journeys, will be published in 2026.