Lost in Time: Why AI Still Struggles With Clocks and Calendars

AI still struggles with something as basic as telling time

Our paper, Lost in Time: Clock and Calendar Understanding Challenges in Multimodal LLMs (co-authored by Miniml in early 2025), highlighted a simple but important gap: even advanced models can misread clocks, misunderstand dates, or fail at calendar reasoning.

Fast forward to today, and this is still showing up in real systems. Models like Sonnet 4.6 continue to get time-related questions wrong in practical scenarios.

Why this matters for business:

• These are not edge cases. Time, schedules, and deadlines are core to business workflows
• Failures are often silent. The answer looks plausible but is incorrect
• Multimodal AI is not yet reliable for operational decision-making without safeguards

The takeaway is straightforward: do not assume capability equals reliability.

At Miniml, we focus on how AI performs in real business operations, not just benchmarks. That means testing systems against real workflows and edge cases, and building bespoke guardrails as part of the solution so they can be used reliably in practice.

Because in production, small errors are not small. They compound.

AI adoption is accelerating, but trust should be earned through rigorous validation, not assumed from model performance claims.

Paper: https://arxiv.org/abs/2502.05092

Abstract

Understanding time from visual representations is a fundamental cognitive skill, yet it remains a challenge for multimodal large language models (MLLMs). In this work, we investigate the capabilities of MLLMs in interpreting time and date through analogue clocks and yearly calendars. To facilitate this, we curated a structured dataset comprising two subsets: (1) ClockQA, which comprises various types of clock styles-standard, black-dial, no-second-hand, Roman numeral, and arrow-hand clocks-paired with time-related questions; and (2) CalendarQA, which consists of yearly calendar images with questions ranging from commonly known dates (e.g., Christmas, New Year’s Day) to computationally derived ones (e.g., the 100th or 153rd day of the year). We aim to analyse how MLLMs can perform visual recognition, numerical reasoning, and temporal inference when presented with time-related visual data. Our evaluations show that despite recent advancements, reliably understanding time remains a significant challenge for MLLMs.

The Verge article: https://www.theverge.com/report/829137/openai-chatgpt-time-date