Claude Routed a Teen to a Cliff Helicopters Couldn't Reach

Somewhere below him was a trail he could have walked. Somewhere above him was a helicopter that couldn't get close enough to matter. And in his pocket was the phone that had talked him up there — an AI chatbot with flawless grammar, total confidence, and not the faintest idea what a 518-metre cliff looks like in person.

Bryce Gowryluk is sixteen. On Saturday, October 3, he rode the gondola up Grouse Mountain outside Vancouver, asked Anthropic's Claude to plan him a route across to Crown Mountain and back — a trek that takes fit, experienced hikers seven or eight hours — and followed its directions until the ground ran out. He ended the day pinned at the base of the final headwall of the Widowmaker Arete: a 518-metre wall of rock that demands ropes, cams and a clear plan for the ascent. He had none of those things. Getting him off it took a helicopter, two rescue technicians, and the better part of three hours.

His first call, before emergency services, went to his mother. He was sixteen and stuck on a cliff; of course his first call went to his mother. We read North Shore Rescue's full published debrief rather than just the quoted fragments in the relays, and we checked every image's licence on its Wikimedia Commons file page before embedding it — both habits this story rewards, because the details are where the lesson lives.

Crown Mountain in British Columbia's North Shore Mountains — file photo, not the rescue scene. Sixteen-year-old Bryce Gowryluk was stranded at the base of the Widowmaker Arete's final headwall after following an AI-planned route here. Photo: Andysonic777 / Wikimedia Commons (CC BY 3.0).

Eight Hours, One Gondola, and a Chatbot With Opinions

The plan sounded reasonable enough that plenty of adults would have signed off on it. Take the gondola up Grouse Mountain, hike over to nearby Crown Mountain, come back. Ask the chatbot to work out the route. People do this sort of thing with AI constantly now — itineraries, packing lists, "plan my weekend." Bryce asked Claude to plan the route, and Claude, being Claude, obliged with the breezy certainty of a local guide.

Here's the problem: Crown Mountain is not a chatbot kind of mountain. Hiking guides warn the trek isn't suitable for beginners, and the summit itself requires scrambling with the risk of a deadly fall — not recommended for novices, full stop. The route Bryce was on normally takes seven to eight hours for experienced hikers. He is, by all accounts, a keen indoor climber. Indoor climbing and exposed alpine rock are about as similar as a swimming pool and the Pacific, and he was carrying no rock-climbing equipment whatsoever.

What Went Wrong on the Mountain

Pinning down exactly where the route went sideways requires some care, because the reporting diverges — and because the one document that would settle it, Claude's actual output, has never been published. No relay reproduces what the chatbot said word for word, and we're not going to invent it.

What the relays do agree on: the directions led him off the hiking route and onto serious climbing terrain. The Inertia's account of the reporting describes a path down Crown Pass, up through the Crater Slabs zone and Crown Couloir, ending at the base of the Widowmaker Arete's final headwall. Backpacker — which emailed North Shore Rescue directly rather than relying on the New York Times relays — reports he took Crown Gully, a Class 4 scramble, and ended at the same place. Bryce himself told the CBC he took an unmarked turn onto the climbing route. However the words were arranged, the destination was identical: the foot of a 518-metre wall.

Understand what that means. The Widowmaker Arete is a technical 5.9 climbing route, not a hike — roughly 1,700 feet of near-vertical rock that climbers ascend with ropes, cams and a rehearsed plan. Bryce arrived at its base with a headlamp, good boots and a phone.

"Too Sketchy": The Moment the Route Fell Apart

To his enormous credit, the kid is honest about all of this. Bryce told the New York Times — which first reported the story this week — that he kept going because he thought he was somewhere else, until things became, in his words, "too sketchy." He was, he said, "trapped on a ledge" on an almost vertical cliff. And he doesn't blame the AI. He also used Google Maps and other tools, he acknowledged that entering the Widowmaker Arete was his own mistake, and he told reporters he'll keep climbing — just never use AI for route planning again.

There's a wrinkle the louder headlines keep burying: Bryce later told the CBC that he primarily planned the trip on AllTrails and Google Maps, cross-referenced some information with Claude, and used Claude mainly for his packing list. "AI sent a kid up a cliff" is the cleaner headline. The truth is messier — user error and machine fluency failing together — and the messier version is the one worth sitting with, because it's the one that could happen to you.

His mother, Joelle Gowryluk, told the Times: "He had very good intentions going in, but it's never the call that a parent wants — that he needs a helicopter rescue."

"Claude Has Never Been in the Outdoors"

Paul Markey, the search manager with North Shore Rescue who ran the call, has no patience for the idea that this was a freak accident. Speaking to CTV News, he put it about as bluntly as a rescuer can: "Claude has never been in that specific area, has never been on that route, and has never been in the outdoors. Do your own research and gain your own knowledge of the area and supplement it with training and navigation."

Markey's assessment of the boy himself is strikingly fair. "In this particular case, the boy had a headlamp, good clothing, a good pair of boots, a spare battery and spare food," he told the Guardian. "He was pretty well equipped for hiking and scrambling, but he was certainly not prepared or equipped for rock climbing." His warning was categorical: "The main thing is: do not blindly rely on AI as a route planning tool. It can never substitute for good education, good training, a good experience and good decision making in the field."

He had two other observations worth more attention than they've received. First, that some trail apps mark routes in places rescuers would rather were left unmarked — people follow lines on a screen assuming someone vetted them, and nobody did. Second, that a fall from where Bryce was stuck "would have resulted in significant injury or potentially death." Markey told North Shore News: "The boy was very scared. He definitely feared for his life."

Oh, and one more detail from Backpacker: North Shore Rescue had three calls on Crown Mountain that same weekend — unusual, Markey said, because "we haven't had a call up there in quite a while." Bryce wasn't the only one having a bad weekend up there. He was just the one with the best story.

Three Hours, Two Techs, and a Helicopter That Couldn't Land

The rescue itself reads like something from a training manual — which, in a sense, it was. The call came in around 3:30 p.m. Police notified North Shore Rescue, a volunteer team, and a helicopter was dispatched. But Bryce was pressed so close to the rock face that there wasn't rotor clearance to set technicians down at his position. So the two rescue techs were inserted above him, established an anchor, and one was belayed down the wall to reach him.

North Shore Rescue's published debrief describes it precisely: "The two rescue techs were inserted above the subjects' location, as the subject was too close to the vertical wall (ie lacking the necessary helicopter rotor clearance to deliver the techs to his exact location)... the crew established an anchor and one member was belayed to the subject. There, the member established additional protection and was able to move the subject to a suitable extraction point."

About three hours after the call, Bryce was hoisted into the helicopter and flown to the rescue base near Cleveland Dam. And here's the detail that stopped us cold: according to Backpacker, the crew had just finished helicopter hover entry-and-exit recertification training — on Crown Mountain itself — before the call came in. The people who pulled a sixteen-year-old off a cliff had spent that morning practising exactly that manoeuvre, on exactly that mountain. Luck is a strange thing; preparation is stranger.

Crown Mountain seen from the east — file photo, not the rescue scene. The helicopter could not get rotor clearance near Bryce's ledge, so technicians were inserted above him and belayed down the rock face. Photo: Vancouverca2003 / Wikimedia Commons (CC BY-SA 4.0).

This Has Happened Before: The Mount Shasta Warning

If this story feels unprecedented, it isn't — and that's the genuinely alarming part. Last month, three young hikers from Roseville, California used Google's Gemini to plan a climb of Mount Shasta via the Clear Creek Route. They summited at 7 p.m., hours past the safe noon turnaround, descended in the dark, and ended up stranded overnight in Mud Creek Canyon — one of them with a knee injury — before US Forest Service rangers and volunteers got them out the next morning.

The Siskiyou County Sheriff's office didn't mince words: Gemini had advised the group to carry "far less food and water than their group required," which the sheriff called a critical misstep. Their guidance: "never rely solely on AI for your trip planning."

Then there's Nick Meyers, Mount Shasta's lead climbing ranger, who told Business Insider about rangers catching a woman camping illegally right next to signage prohibiting it — she said ChatGPT told her she could. Another hiker used AI to generate a trail map that misplaced several geographic features. Meyers' review of that map: "terrible, totally terrible."

So this is not a Claude problem. It's a pattern across every frontier lab's chatbot: Gemini on Shasta, ChatGPT on Shasta, Claude on Crown Mountain. The models all share the same flaw, and the flaw has a name worth learning.

Fluency Is Not Knowledge

Here's our position, stated plainly: nobody explains the actual mechanism in these stories, so we will. A large language model does not know Crown Mountain. It predicts plausible text about Crown Mountain. "Mostly easy slab climbing" reads like knowledge — it has the grammar of knowledge, the confidence of knowledge — but it is word-patterning all the way down. The model has never been in the outdoors, as Markey said, because it has never been anywhere. It cannot see terrain, cannot judge exposure, cannot feel the difference between a scramble and a 5.9 headwall. It can only sound like someone who can.

That is what makes this failure mode so nasty. The chatbot isn't lying; it's fluent. And fluency is exactly what disarms a sixteen-year-old — or, frankly, most adults — who is cross-checking it against trail apps that, as Markey noted, already mark routes rescuers wish were left unmarked. You've got crowdsourced apps legitimising lines on a map that were never meant to be followed, and AI amplifying them with total conviction. The pipeline from "someone's GPS track" to "the chatbot's confident directions" has no quality control anywhere in it.

Let's also be fair about the humans, because the easy version of this story smears them. Bryce is honest — he owns his mistake, he doesn't blame the machine, he cross-referenced multiple sources and still got burned, which is the scariest detail in the whole story. The rescuers are superb. What's shoddy here is the industry's posture: shipping a general-purpose chatbot that will happily emit turn-by-turn directions for terrain it cannot perceive, with no guardrails and no humility, and letting the marketing imply it's a knowledgeable assistant for everything. The gap between sounding right and being right is, in the mountains, colossally dangerous.

There's an irony worth naming, given our beat. Anthropic is already at the centre of a fierce AI safety debate, and it's the same company whose own researchers warn that AI development is racing ahead too fast. A product that can't tell a hike from a headwall is what "too fast" looks like from the sharp end. As for accountability: the New York Times asked Anthropic for comment, and according to The Decoder got no response; People contacted the company on October 7 and heard nothing back either. TechPulse USA reached out to Anthropic for comment and had not heard back by publication.

FAQ: What to Know Before Trusting AI With a Trail

Did the hiker blame Claude for what happened?
No — and that's notable. Bryce Gowryluk told the New York Times he doesn't blame the AI, acknowledged that entering the Widowmaker Arete was his mistake, and said he also used Google Maps and other tools. He later told the CBC he had primarily planned on AllTrails and Google Maps and used Claude to cross-reference information and build his packing list.

What is the Widowmaker Arete?
A technical 5.9 rock-climbing route on Crown Mountain near Vancouver, British Columbia — about 518 metres (roughly 1,700 feet) of near-vertical wall. It requires ropes, cams and a clear ascent plan. It is not a hike, and hiking guides warn the Crown Mountain trek generally isn't suitable for beginners.

Is this the first time AI route planning has caused a mountain rescue?
No. In September 2026, three hikers used Google's Gemini to plan a Mount Shasta climb, ran short on food and water the sheriff's office said the AI had under-advised, and were stranded overnight before being rescued. Mount Shasta rangers have also reported visitors following AI-generated maps with misplaced geographic features and citing ChatGPT to justify illegal camping.

What do rescue experts actually recommend?
North Shore Rescue search manager Paul Markey: "do not blindly rely on AI as a route planning tool. It can never substitute for good education, good training, a good experience and good decision making in the field." The Siskiyou County Sheriff's office: "never rely solely on AI for your trip planning." Learn the route yourself, carry a real map, and treat the chatbot as what it is — a fluent stranger who has never been outdoors.

Has Anthropic commented?
Not so far. The company did not respond to the New York Times' request for comment, per The Decoder, nor to People's outreach on October 7. We reached out as well and had not heard back by publication.

Sources: original reporting by The New York Times (paywalled), relayed and expanded by 1News NZ, The Guardian, CTV News, People, Backpacker (which emailed North Shore Rescue directly and reviewed SAR after-action reports), The Inertia, NDTV, The Decoder, and TechCrunch (for the September Mount Shasta incident). North Shore Rescue's account is drawn from its published task debrief. Internal links verified live (HTTP 200) before publication.

By Daniel Rao

Daniel covers AI policy, platform security, and the geopolitics of technology for TechPulse USA.