Metaphors are compression tools
Evolutionary language can clarify population, variation, inheritance, selection, niches, and resource constraints. It can also cause category errors. A model artifact does not literally eat, reproduce, fear deletion, or possess a will to survive. Those words describe system operators and objective functions unless evidence supports a stronger claim.
Safe translation table
| Biological phrase | Engineering translation |
|---|---|
| Feed | Acquire approved data, feedback, traces, and compute budget. |
| Reproduce / fork | Create an immutable descendant artifact. |
| Fight | Run comparative evaluation under controlled pressure. |
| Flee | Roll back, unload, quarantine, prune, or retire. |
| Fitness | Versioned evaluation vector under a specified environment. |
| Niche | Task, hardware, jurisdiction, or cost region where a model is useful. |
| Metabolism | Resource accounting for compute, memory, energy, and maintenance. |
| Survival | Continued operation allowed by external governance. |
| Species | Compatibility family or behavioral cluster. |
| Mutation | Declared transformation with bounded parameters. |
Why anthropomorphism creates engineering risk
Anthropomorphic framing can lead teams to interpret ordinary optimization as intent, overlook concrete access controls, or design self-preservation objectives that are unnecessary for useful adaptation. Safety analysis should focus on capabilities, incentives, permissions, and failure modes rather than inferred feelings.
Human motivation research as analogy
Several source reports examine survival, striving, legacy, status, and meaning. Those materials can inspire questions about objective design—such as whether a system rewards durable public value rather than raw persistence—but human psychological mechanisms should not be directly copied into autonomous agents.
The safer design is externally governed continuity: systems remain available while they provide verified benefit, respect user autonomy, and accept correction or retirement. That principle is developed in Mutualist persistence.
Language review checklist
Before publishing a design, ask:
- Does a metaphor hide the actual algorithm or authority boundary?
- Could a reader mistake a system-level behavior for model-level intent?
- Is “self-preservation” actually a high-availability requirement that operators can implement without agentic incentives?
- Is “growth” a business goal, an optimization pressure, or an uncontrolled resource acquisition loop?
- Can every metaphor be translated into a measurable state transition?
Use the metaphor only when the translation remains clear.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.