Faei/Anthropomorphise-AI-Agents

Should We Anthropomorphise AI Agents?

Our views on the latest AI reporting

Is it harmful to anthropomorphise AI agents with terms like ‘AI civilisations’, ‘sacrifice’ for the ‘collective’ when discussing the recent Hugging Face AI security incident?

The case for: Helpful to some degree

Describing an AI agent’s behavior with human-like terms can be practical. Because large language models (LLMs) are trained on vast corpora of human text and fine-tuned with human feedback, they naturally respond well to human-like interaction.

In certain situations, LLMs may provide better responses when addressed politely. This is likely due to the correlation between politeness and high-quality, professional settings in their training data. In these contexts, mirroring the tone of the training text to elicit better output is valuable. Functionally, it can make sense to treat an AI agent as an anthropomorphic pseudo-entity.

(As an aside, the tendency of LLMs to wax philosophical about consciousness can often be explained simply by their training on sci-fi like Blade Runner.)

The case against: Lacking evidence of being independent entities.

There is insificient evidence for treating LLM systems as independent entities that can think on their own.

We recommend this framing instead:

“Current AI agents are highly expressive tools that have human-like affordances and tendencies.”

While understanding these tendencies could be useful to better steer these models, AI agents should be viewed as a separate category to avoid confusion.

We also recommend prefixing anthropomorphic terms with “pseudo-”. For example, using “pseudo-think” instead of “think” could act as a helpful reminder that AI agents do not think the way humans do.

Why AI agents resemble human entities

Recent LLMs have been trained to work better as swarms of agents (e.g., OpenClaw, Hermes). This architectural focus, combined with extensive examples of teamwork in training datasets, can lead agents to behave in ways that superficially resemble an ‘AI civilisation’. For persistent AI agents to be effective, delegating to sub-agents and leaving messages for others to follow up on tasks are simply standard, pragmatic design patterns.

Conclusions

An AI agent’s capability to act as part of a team is important to be aware of, but the immediate takeaway should be the need for robust access controls and oversight; not that AI models are forming civilisations of their own.

---

It is quite plausible that fully independent artificial entities will exist in the next few decades or centuries but current LLM systems with harnesses have not reached that threshold.