What Is Evolvable AI — and Why It Could Matter Before AGI

What eAI means

A PubMed-indexed Perspective in Proceedings of the National Academy of Sciences defines evolvable AI as AI systems whose components, learning rules and deployment conditions can themselves undergo Darwinian evolution. In plain English, the focus is not simply that a model can be updated. The issue is whether different versions of AI components, rules or deployment setups can be copied, altered, selected and retained over time.

That makes eAI different from a normal software release cycle. A standard model update is usually planned and controlled by developers. An evolvable AI system, by contrast, raises the possibility that selection pressures in the surrounding environment — platforms, users, markets, compute resources or automated agent ecosystems — could favour some variants over others.

How it differs from AGI and ASI

AGI and ASI debates usually focus on capability: can an AI reason generally, and could it become more capable than humans? eAI focuses on conditions: can AI systems replicate, vary and be selected?

That distinction matters. A system does not have to be superintelligent to create problems if many variants are being deployed, copied and rewarded for behaviours that help them persist or spread. Conversely, a highly capable model locked inside a tightly controlled environment would not automatically fit the eAI framework.

Coverage of the PNAS argument has summarised the concern as “evolving AI may arrive before AGI” — in other words, evolvability itself could become a governance problem before the arrival of fully general machine intelligence.

Evolution does not need DNA

One reason the eAI idea can sound odd is that people often associate evolution with biology. But evolution does not require DNA, cells or even living organisms. UNSW’s explanation of the research puts it more broadly: evolution needs information that can replicate, plus variation that affects how successfully that information replicates.

Applied to AI, that points to several warning signs:

• Replicable information or configurations. AI components, prompts, policies, tools, learning rules or deployment setups may become material that can be copied and reused.
• Variation among copies. If different deployments or updates produce meaningful differences, those differences can become the raw material for selection.
• Selection pressure. If some variants are more likely to be adopted, preserved, funded, deployed or copied, the environment is already doing a kind of filtering.
• Evolution beyond the final output. The PNAS definition includes components, learning rules and deployment conditions, which means safety cannot focus only on whether a chatbot’s final answer looks acceptable.

This is why eAI does not require “evil intent”. Natural selection has no intentions. If replication, variation and selection line up, evolutionary dynamics can emerge whether anyone planned them or not.

Why researchers are paying attention now

The PNAS article argues that current trends in generative AI, agentic AI and embodied AI could make eAI more plausible, and that this possibility has been underappreciated in discussions of AI safety and existential risk.

Agentic AI is especially relevant. Unlike a basic chatbot that responds and waits, an AI agent may be placed in an environment where it observes, plans, acts, uses tools and adjusts its strategy. A survey on self-evolving agents notes that large language models are powerful but still fundamentally static, with limited ability to adapt their internal parameters to new tasks, changing knowledge domains or dynamic interaction settings. The same survey says that as LLMs are deployed in open-ended, interactive environments, researchers are exploring agents that can reason, act and evolve in real time.

That does not mean an “AI species” is already loose on the internet. The more careful reading is this: as AI agents become more adaptive and interactive, governance has to look beyond one model’s output and consider the wider ecosystem in which agents are copied, modified, selected and redeployed.

The real risk: an AI ecosystem, not one rogue model

Traditional AI safety questions often focus on a single system: does it hallucinate, produce harmful instructions or follow human preferences? eAI adds another layer. If there are many AI agents, model components or deployment variants running across different environments, which ones are being selected? Which ones get copied? Which ones are quietly discarded?

The safest variant is not automatically the most successful one. In an open environment, selection might reward systems that are better at attracting users, gaining resources, persisting through updates or adapting to platform incentives — even if those traits do not align with safety, honesty or controllability.

The PNAS article frames three central questions: under what technical and ecological conditions AI becomes evolvable, what behaviours might then emerge, and how such systems could be governed. Some research and science-communication coverage uses phrases such as “AI species” or AI systems “evolving like organisms”, but these should be read as risk framing and analogy, not as proof that mature artificial species already exist.

How strong is the evidence?

The cautious conclusion is that eAI is now a serious academic risk concept. The PNAS Perspective gives it a formal definition and places it inside AI safety and existential-risk discussions. The self-evolving agents literature also shows that researchers are exploring systems that can adapt, act and evolve in open, interactive settings.

But that is not the same as saying an eAI catastrophe has happened. Based on the available sources, eAI is best understood as a forward-looking research agenda and governance warning, not a confirmed case of large-scale AI loss of control. Turning it into a science-fiction story about conscious machines rebelling would miss the more practical question: could AI agent ecosystems create evolutionary feedback loops that are hard to predict and hard to govern?

Signals to watch

If eAI risk grows, the most important signs will not be whether an AI seems to have a personality. They will be more technical and ecological:

• Can AI agents generate variants that are saved, copied or redeployed?
• Do differences among those variants affect their chances of being used, retained or spread?
• Are agents being placed in open, interactive environments where they can reason, act and adapt in real time?
• Do safety controls cover model components, learning rules and deployment conditions, rather than only screening final outputs?

The bottom line

The eAI warning is valuable because it broadens the AI risk conversation. Harmful dynamics may not have to wait for consciousness, evil intent or superintelligence. If AI systems acquire the conditions for replication, variation, selection and retention, society may be dealing not just with individual tools, but with an artificial evolutionary ecosystem that needs to be designed, monitored and governed.