It Felt a Bit Eerie": Exploring Humanlike Interactions During Collaborative Writing with an Artificial Agent

Synthesis: This comparative user study (n=48) examines how the temporal and visual dimensions of AI collaboration shape the experience of writing tasks, revealing that humanlike design features in AI agents create both positive social expectations and unexpected social costs.

The Humanlike-to-Machinelike Spectrum

Three AI-assisted text editor variants were tested along two dimensions:

• Temporal: synchronous (humanlike) vs. asynchronous (machinelike) suggestions
• Visual: presence (humanlike) vs. absence (machinelike) of a cursor

Synchronous suggestions increased efficiency but led to contextual misalignment — the AI's suggestions didn't fit the writer's intent. A visual cursor improved perceived intent understanding but evoked feelings of surveillance that participants described as "eerie." This maps onto emerging concerns in AICoFe: Implementation and Deployment of an AI-Based Collaborative Feedback System for Higher Education about the social dynamics of AI interaction in educational contexts.

What this means for practice

• Designers. Pair anthropomorphic features with explicit expectation management: the uncanny valley effect in AI collaboration means that without the mutual alignment natural in human-human collaboration, humanlike design features can backfire.
• Designers. Do not rely on learners to tell an AI collaborator from a human one, because users cannot detect AI teammates above chance; in LLM co-writing tools that ambiguity amplifies both the benefits and the costs of humanlike interfaces.
• Learners. Treat an agent's interface cues as design rather than understanding: the visual cursor improved perceived intent understanding but also produced feelings of being monitored, judged, and compared against, which participants described as "eerie."
• Learners. Watch for contextual misalignment when suggestions arrive while you are still writing, since synchronous suggestions raised felt efficiency but produced suggestions that did not fit the writer's intent.

Limitations

• Single task and single session: 48 participants (16 per variant) completed one academic-writing task in about an hour, so the findings rest on a short, low-stakes opinion-piece prompt and may not extend to other writing or to longer collaboration.
• Humanlikeness was simulated with heuristics: the agent's timing and cursor behavior were tuned to the mean and standard deviation of human keypresses rather than being genuinely human, and the authors note this may itself have shaped participant behavior.
• The authors state their conclusions rest on qualitative findings because of the sample size, so the study establishes exploratory patterns rather than measured effects.
• Convenience sample: participants were recruited through external Slack channels and the institute's paid-study listing, compensated $16 CAD, and averaged 28.1 years old (range 19 to 54).

Citation

Yin, M., Chiang, A., Cox, S. R., & Xiao, R. (2026). "It felt a bit eerie": Exploring humanlike interactions during collaborative writing with an artificial agent.