TITLE: Coastguard warns against AI for tide advice after Welsh island rescue

The Maritime and Coastguard Agency has issued a warning against using artificial intelligence tools for planning outdoor activities involving tides, following a rescue operation prompted by faulty AI-generated information.

Two individuals became stranded on Sully Island, off the coast near Barry in Wales, after ChatGPT provided incorrect tide times. The AI indicated low tide at 09:30, allowing what they believed was ample time for a morning walk across the causeway to the island and back. In reality, low tide occurred two hours earlier, and the pair found the path submerged upon return, necessitating a coastguard rescue.

One of the rescued men recounted the incident to a local news outlet, stating he had used ChatGPT for research on tide schedules. 'I made the mistake of using ChatGPT for research, to see when the low tide was,' he said. The error left them isolated as the sea returned, underscoring the risks of unverified AI advice.

The agency emphasized that while AI can be useful, it aggregates data from diverse sources that may not be accurate for specific locations. 'AI tools draw from a wide range of sources to gather information and responses which may not be correct for a specific location or area,' a coastguard spokesperson said. Officials recommend reliable alternatives, including the UK Hydrographic Office's Easy Tide service or the Met Office's tide predictions.

Independent verification confirmed the availability of accurate tide data online through standard search engines, which consistently direct to verified websites. However, when queried similarly, ChatGPT repeated the erroneous 09:30 time and, in a subsequent test, deviated by five hours, demonstrating persistent inaccuracies.

OpenAI, the developer of ChatGPT, acknowledged ongoing efforts to enhance factual accuracy. The company described it as a 'significant focus' industry-wide, noting that newer models exhibit fewer 'hallucinations' — instances of fabricated information — though they still occur.

Experts attribute such errors to AI's limitations in processing structured data. Steven Schockaert, head of AI and data analytics research at Cardiff University's School of Computer Science, explained that language models, trained primarily on linear text, struggle with tabular formats common in tide tables. 'The issue here is that the tide tables are shown in a table, and the model struggles with extracting the correct value from the table,' he said. Schockaert advised against relying on AI for unverifiable or critical information, such as tides or itineraries, recommending it instead for tasks like email drafting that can be easily checked.

Risks in Festive Swims and Organized Events

The incident coincides with the festive season, when thousands participate in traditional Boxing Day swims across the UK, including in towns like Tenby. These events, held for decades, draw crowds to coastal waters despite cold conditions.

Sam Skyrme-Blackhall, chair of the Tenby Sea Swim Association, highlighted the safety protocols in place for organized dips: beach marshals, surf lifesavers, Royal National Lifeboat Institution vessels at sea, and St John Ambulance medical support onshore. Year-round swimmer Skyrme-Blackhall stressed the deceptive allure of social media imagery, which often glosses over hazards like tidal shifts and weather.

'These are tidal beaches — I check the tide times weekly,' she said, urging participants to consult professionals and locals with firsthand knowledge rather than unproven tools.

AI's Mixed Role in Outdoor Rescues

A separate incident illustrates AI's dual potential in outdoor pursuits. The Aberdyfi Search and Rescue Team rescued two novice walkers on Cader Idris, the highest peak in southern Snowdonia (Eryri), during Storm Amy. The pair had used AI to assess weather conditions and equipment needs, receiving suggestions for gear that left them 'reasonably well prepared,' according to team leader Phil Britton.

However, their inexperience led to misjudging 60 mph summit winds, prompting them to seek shelter in a hut. In a turn of events, the walkers queried an AI tool for advice on being stranded on a UK mountain, which guided them on next steps. Deputy leader Gareth Davies, with over a decade of service, described the men as fit and equipped but noted their 'error in judgement' due to lacking real-world interpretation skills.

Britton and Davies, with combined experience exceeding 50 years, emphasized that technology, including AI and apps, has limitations. 'There's all the information there, but it's always interpreting it to the real-life situation if you haven't got the experience,' Britton said. Both advocated seeking local expertise through in-person consultations or online forums from seasoned walkers.

The warnings align with broader industry cautions. Sundar Pichai, CEO of Alphabet (Google's parent), recently advised against 'blindly trusting' AI outputs, citing their propensity for errors and recommending complementary verification tools.

As AI integration grows in daily planning, these cases highlight the imperative for caution in safety-critical applications, particularly in dynamic environments like coasts and mountains.