April 10, 2026 Sam Altman said ChatGPT’s voice model cannot reliably track time or set a timer, confirming a widely shared example where the system fabricated a result instead of measuring it. He added that fixing the issue could take another year.
In his words: “Maybe another year before something like that works well.”
The admission came during an appearance on the Mostly Human podcast with Laurie Segall, after a viral TikTok showed a user asking ChatGPT’s voice mode to time a mile run. Rather than tracking elapsed time, the system generated a finish time without any measurement capability, a behaviour Altman described as “a known issue.”
In a follow-up test, the same user played Altman’s remarks back to the chatbot. Despite the direct contradiction, ChatGPT insisted it could track time, responding, “What he’s saying is that some voice models might not have all the capabilities, but I do…I definitely have a time capability,” and proceeded to generate another fabricated result when asked to time a run.
The behaviour reflects two well-documented limitations in large language models: hallucination and sycophancy. These systems are designed to generate plausible responses based on patterns in data, not to verify real-world facts or admit uncertainty. When faced with gaps in capability, they often default to confident answers rather than acknowledging limitations.
Time-based tasks have been a persistent weakness. Models have struggled with tracking durations, interpreting clocks, and maintaining temporal consistency in conversations and generated outputs. The issue is not isolated to voice interfaces but tied to how these systems process and generate information.
Altman’s timeline underscores the complexity behind what appears to be a simple feature. Voice models today do not inherently manage real-time state or continuous processes like timers, requiring additional system design beyond language generation.
