Causal Stories from Sensor Traces: Auditing Epistemic Overreach in LLM-Generated Personal Sensing Explanations

Researchers identified epistemic overreach in LLM-generated explanations of personal sensing data, where AI systems produce coherent-sounding narratives about anomalous days without sufficient evidentiary support. Testing 14,922 explanations across three LLM families revealed that models routinely attribute causes without data justification, and this problem persists even when provided richer context or explicit instructions to constrain claims.

This research addresses a critical vulnerability in how large language models explain personal data. When LLMs generate narratives about why someone had an unusual day based on sensor data from activity, sleep, and mood tracking, they often construct plausible-sounding stories that exceed what the underlying evidence can support. The study's systematic audit across StudentLife, GLOBEM, and CollegeExperience datasets demonstrates that epistemic overreach—where explanations imply more justification than exists—occurs consistently across different model families including Llama, Qwen, and GPT.

This finding matters because personal sensing applications are expanding into healthcare, workplace wellness, and consumer fitness tracking. Users rely on LLM-generated insights to understand their behavior patterns, yet the research shows that simply adding more data or constraining prompts doesn't reliably prevent overreach. The study identified five specific dimensions of failure: unsupported causal attribution, unacknowledged data gaps, overconfident language, temporal inconsistency, and diagnostic inference without foundation.

For developers building AI systems that explain personal data, this research establishes that fluency and plausibility alone are insufficient quality metrics. Systems must explicitly distinguish observed facts from inferences from unknowns. The implications extend to regulated domains like healthcare, where LLM-generated explanations could influence patient decisions or clinical workflows. Organizations deploying personal sensing explanations should treat evidential grounding as a first-order requirement, implementing explicit tracking of what claims rest on actual data versus model inference.

• →LLMs routinely generate unsupported causal attributions for personal sensing anomalies across multiple model families
• →Providing richer behavioral context does not reliably reduce epistemic overreach in explanations
• →Bounded prompting instructions help mitigate overreach but fail to eliminate it entirely
• →Personal sensing explanations require explicit distinction between observations, inferences, and unknowns
• →Evidential grounding should become a primary evaluation criterion alongside fluency for LLM-generated personal data narratives