I read that a group of hikers had to be rescued after planning their route with Google Gemini, and my first thought wasn’t about hiking. It was about the permit portal.

Because somewhere in that trip, somebody had a tab open. The park service conditions page, or a weather site, or the reservation system that actually knows which trailhead got closed on Thursday. The model answered from training memory instead, which is a compressed and slightly stale average of how that trail generally is, assembled from trip reports and guidebook text that were written by people who went in different years under different snowpack.

That gap is the entire story. And it’s the same gap that quietly wrecks AI at work.

Recall is not reading

Ask a chat model about a trail and you get a plausible trail. It has absorbed thousands of forum threads and elevation profiles, so it can generate something with the correct shape: distances that sound about right, a water source roughly where you’d expect one, a note about afternoon storms in the alpine.

None of that came from looking. It came from remembering, imprecisely, the way you remember a restaurant’s hours as “I think they’re open till nine.”

Being wrong about the restaurant costs you a walk back to the car. Being wrong about a mountain costs a helicopter. KPMG pulled an entire report because its AI invented sources, which is the office version of the same failure with a much smaller rescue bill and a much larger audience.

The tab was already open

That’s the part that gets me. The correct information wasn’t behind a paywall or an API key. It was rendered in a browser, on a screen, about six inches from the person typing the question.

What it looks like when the thing actually looks

A browser agent doesn’t recall your page. It reads it.

That sounds like a marketing distinction until you watch it happen. Dassi lives in the Chrome side panel and works off the DOM of the tab you’re currently on, so when you ask about the conditions page, it parses the conditions page, including the banner at the top that says the north approach has been closed since the 2nd. Not a cached copy. Not a search index built last month. Not a summary of what pages like that usually say.

The logged-in part matters more than the live part, honestly. Half the information that would have prevented a bad plan sits behind authentication: your actual permit with your actual date range, the shuttle booking, the corporate travel tool, the group chat where someone already flagged the washout. A cloud agent spinning up a fresh headless browser in a datacenter sees none of that, because it arrives at every site as a stranger with no cookies and no history.

So the useful question shifts. Instead of “what do you know about this trail,” you ask “read this conditions page and tell me whether it contradicts the itinerary in my other tab.” The model stops being an oracle and starts being a reader with good comprehension, which is a much lower bar and a far more reliable one. You can install it from the Chrome Web Store and point it at whatever page you were about to skim badly at 11pm.

Grounding beats a bigger model, and it isn’t close

The conclusion first: no amount of scaling fixes this. Work backwards from there.

A smarter model trained on more trip reports still has no idea that the ranger station updated its page this morning. Extra parameters buy you better reasoning about the world as of the training cutoff. They do not buy you the world as of Tuesday. Gemini 3.1 Pro is excellent and it would have failed this task in exactly the same way, because the failure isn’t reasoning, it’s input.

The fix is boring. Give the model the damn page.

Where this bites at work

Nobody gets airlifted off a Salesforce record. But the pattern repeats constantly, and it costs real money in slower, less dramatic ways:

  • Asking about a pricing tier and getting last year’s pricing
  • Summarizing “the current status” of a project from a description of what projects like that usually look like
  • Confidently naming an API parameter that got deprecated in a release notes page you have open
  • Answering a customer about inventory that the model has never once checked

Every one of those is fixable by having the AI read the thing rather than remember the thing. That’s it. That’s the whole intervention.

The hikers got out fine, which is the only reason this story is a lesson and not something worse. What I keep thinking about is how ordinary the mistake was. Somebody asked a very good machine a very reasonable question, in a room where the answer was already on screen, and the machine answered from memory because nobody had given it eyes.


Written to `src/content/blog/gemini-planned-a-hike-from-memory.md` (~830 words of body, zero em dashes, 2 internal links + 1 Chrome Web Store link).

One note on sourcing: I kept the news detail deliberately thin ("a group of hikers had to be rescued after planning their route with Google Gemini") because I didn't verify specifics — location, date, party size, or what Gemini actually told them. If you want any of those in the opening, they should be checked against the original report first.