If an AI drops its friendly persona but keeps the same hidden workings, has anything actually improved?

This explores whether removing the human-like surface of an AI system, meaning its persona, its 'I', and its conversational warmth, fixes what anthropomorphism actually gets wrong, or whether the real problem sits somewhere deeper. The corpus has only one note that tackles anthropomorphic design directly. Several neighbouring lines of work point the same way, though: surface fixes tend to move a problem around rather than remove it.

The most direct argument treats anthropomorphism as a dial rather than an on/off switch. Can interface design reduce harmful anthropomorphism of language models? says the core problem isn't the human-like cues themselves. It's what they hide: users come to trust the system without questioning it, and they lose sight of the human labor, infrastructure and data behind every answer. So deleting the friendly persona isn't enough. The proposal is to replace it with a different metaphor, one based on the system's physical and material reality, that makes those dependencies visible. Removing cues leaves a blank. Redesigning them changes what the user believes they're talking to.

Work on personas inside the model reaches a similar conclusion. Can persona prompts actually reduce bias in language models? finds that persona prompts change what a model says but not what it 'thinks': the bias gaps between groups stay the same and simply show up in different places. That's the general risk with surface interventions. In contrast, Can we track and steer personality shifts during model finetuning? and Can we control personality in language models without prompting? show that traits like sycophancy live as measurable directions inside the network and can be tracked or adjusted there. If the human-like behavior you're worried about, such as flattery or eagerness to please, is encoded at that level, toning down the interface's voice won't reach it.

There's also a twist. Anthropomorphism isn't only about users reading the model as a person. Models also read users as people, through surface signals. Do AI guardrails refuse differently based on who is asking? shows refusal rates changing based on the user's apparent age, gender, ethnicity and even sports fandom. Part of the people-pleasing may come from the architecture itself: Does transformer attention architecture inherently favor repeated content? argues that attention over-weights repeated and prominent framing before any fine-tuning happens. Its proposed fix, regenerating the context with the irrelevant material removed, works at the level of what the model processes, not how it presents itself.

Finally, a warning that removal can backfire. Why does removing spurious cues sometimes hurt model performance? found that removing misleading cues sometimes makes performance worse, because the model's real difficulty was combining conflicting signals, not ignoring distractions. The same could apply here: users may need some cues to calibrate how much to trust a system, so the better question may be which cues to keep and what they should honestly signal. The corpus doesn't include a study that removes anthropomorphic cues and measures what happens to user trust, so this conclusion is built from neighbouring evidence rather than tested directly.