AI’s temporal blind spot

By Joshua Vickers

According to a recent study by the University of Edinburgh, state-of-the-art multimodal AI systems —including the latest iterations of GPT-4, Gemini, and Llama-Vision — performed poorly when asked to read analogue clocks or answer simple calendar queries. Accuracy in clock reading remained below 25%. Even identifying the weekday for a known date, a task mastered by primary school pupils and desktop software since the 1980s, yielded error rates as high as 20%.

Such failures are more than mere embarrassments. They reveal the underlying flaws in current AI systems and cast doubt on the readiness of these models for real-world applications in automation, logistics, healthcare, and beyond. The inability to perform seemingly trivial tasks, which have long been resolved by simpler rule-based systems, highlights how little today’s AI understands about the world it claims to interpret.

Moravec’s revenge

The findings revive Moravec’s Paradox — the observation that high-level cognitive tasks like chess and mathematics are easy for machines, while perceptual and sensorimotor tasks, trivial for humans, remain challenging for AI models. What is striking is how little this paradox has faded, despite exponential growth in computational scale.

Experts are clear. “We need AI that can reason, not just mimic patterns,” argues Yoshua Bengio, a Turing Award laureate and pioneer of deep learning. Gary Marcus, a cognitive scientist and longstanding critic of AI hype, calls common sense “the deepest problem in AI” — a deficit laid bare when machines cannot distinguish an hour from a minute hand.

The technical challenge is not minor. Reading an analogue clock involves multiple steps: identifying the orientation of hands, mapping angles to numerical values, and translating those into a coherent time. Current vision-language models, trained primarily to associate images with captions or prompts, lack the structured symbolic reasoning needed for such tasks. Even when visual recognition succeeds, the subsequent step —reasoning about spatial and temporal relationships — often fails.

Image: Joshua Vickers

Pattern without principle

Part of the problem lies in AI’s architectural bias. Transformers, the dominant model type, excel at recognising and generating sequences but lack a native understanding of spatial configurations. Unlike older models — such as convolutional networks, which preserve locality — transformers flatten images into tokens, stripping them of geometric context. A clock face is not just a collection of lines and numerals but a circular representation of time. The models don’t know that.

The Edinburgh team also identified troubling training artefacts. Many clock images in public datasets depict the time 10:10 — a convention favoured by advertisers for symmetry. Models trained on such distributions overfit this ‘happy time’ and falter when shown less photogenic moments.

On the calendar task, similar cognitive dissonance emerges. Answering “What day is Christmas in 2025?” requires not just optical character recognition but grasping how weekdays cycle through months. Without access to a symbolic calendar or an internalised date-mapping algorithm, most models guess — or hallucinate.

Economic mirages

These deficits are not merely philosophical. They create operational headaches in domains where time matters. A misread timestamp can derail a supply chain. An incorrect appointment date in healthcare could risk patient safety. Financial systems that rely on temporal precision may find AI’s unpredictability intolerable.

The economic promise of AI automation has already suffered from such blind spots. PwC has noted that many firms fail to recoup their AI investments, partly due to hidden costs of oversight and quality control. “An AI that needs human supervision for basic tasks isn’t reducing labour, it’s shifting it,” quips Aryo Gema, co-author of the Edinburgh study.

British firms are taking note. Healthcare startups like Babylon Health have had to blend AI with rule-based components after early failures in date-sensitive scheduling. Meanwhile, fintechs and logistics providers are lobbying for more transparent benchmarks before integrating vision-based AI into operational workflows.

Policy and principle

Regulators are catching up. The UK’s AI white paper, published in 2023, emphasised “trust by default” — a demand that systems prove reliability not only in abstract benchmarks but in mundane tasks. Europe’s proposed AI Act classifies such time-sensitive applications as “high-risk”, subjecting them to rigorous scrutiny.

For the public, these failures may provoke a more grounded understanding of AI’s capabilities. The spectacle of a machine generating Shakespearean sonnets while misidentifying Tuesday may induce laughter. However, it should also encourage critical thinking: AI is neither a miracle worker nor a simple tool. It is a sophisticated statistical system — powerful, but profoundly non-human.

Seconds, not Sense

There are reasons for hope. Oxford research groups have shown that specialised architectures, including spatial transformer networks trained on synthetic data, can achieve clock-reading accuracy above 80%. Hybrid models that blend neural networks with symbolic reasoning, are becoming more popular. Embodied AI, which learns by interacting with environments, may offer more intuitive grounding in temporal and spatial concepts.

Still, the lesson remains: intelligence is not just data and computation. It is structure, constraint, and experience. Until AI systems are endowed with more than correlations — until they possess models of time, causality, and physical space — the clock will continue to outwit them.

That may be inconvenient for technologists. But it is a timely reminder for everyone else.

Image: Dall E via Wikimedia Commons