In 2016 a birdwatcher on the Mississippi photographs a heron — and catches, in the background, a critical crack in the Hernando DeSoto Bridge that four professional inspections will go on to miss for years. When it's finally found, the bridge closes for 86 days. Two years later, a detective realizes the inspector and the AI he's been using are failing the exact same way.
Barry Moore paddled the Mississippi one Sunday looking for herons, camera up at the massive steel bridge carrying I-40 across the river — 50,000 vehicles a day between Memphis and West Memphis, Arkansas. Through his viewfinder he saw something that wasn't a bird: a crack in one of the main horizontal support beams, clearly visible even from river level, even with a consumer camera. He zoomed in, took the shot, and thought: that doesn't look good. Then: but I'm sure the inspectors have seen it. They check these bridges constantly. Someone definitely knows. He paddled on, looking for herons. The photo sat in his camera roll for five years. Nobody knew about the crack.
Inspector Mitchell had checked this bridge for years. Same protocol, same checklist, same result: visual examination — check for corrosion — check for cracks — no issues found. 2017. 2018. 2019. 2020. Four clean inspections. Not because he was lazy or incompetent — because he was doing exactly what humans do. Year one you inspect everything with fresh eyes. By year five, the pattern locks in: never found a problem here, probably fine again. Your memory of “no problems” becomes the baseline, and you stop searching with fresh eyes. The protocol required using the under-bridge equipment to examine every beam from every angle. Mitchell's own notes: “I don't believe it was safe to use the under-bridge inspection equipment for this section.” Translation: it felt sketchy; the familiar angles had always been enough. He chose the familiar pattern. And missed the crack. For four years.
Mitchell had been promoted — years of clean inspections, no incidents. A new team from an outside engineering firm took over, and they didn't carry his four years of “no problems found.” No pattern. No anchor. Just the protocol and the equipment.
50,000 vehicles a day suddenly had no way across the Mississippi. The main cargo route, the primary evacuation route, the lifeline between two states — shut down. All because one inspector found a reliable pattern and stopped looking for problems.
Watching the news, Barry Moore remembered. He scrolled back to August 2016 — and there it was, the crack, clear as day, visible from a kayak, photographed by a birdwatcher with no training in structural engineering. He called the DOT hotline: “I think I photographed that crack five years ago.” The forensic timeline was brutal: the crack was visible in 2016; Mitchell's first clean inspection was 2017; four inspections and a 2019 drone survey all had the crack in frame, and none of them saw it. 86 days of closure. ~$550 million in economic damage. A $3.8 million repair. Because pattern-matching bias made a visible crack invisible to trained eyes — and nobody ever verified the inspector was actually looking.
Detective George McIntyre had saved every article in a folder marked “Pattern-Matching Failures,” not knowing why it mattered. Two years later he started using GPTP to research a cold case. It returned confident, detailed, specific results — and half of them were wrong. Confidently, specifically wrong. He asked it why, and it explained: it found results that matched the pattern of the query and presented them as fact. It didn't verify. It pattern-matched. George stared at the screen, opened the filing cabinet, pulled the bridge folder — and it clicked:
Researchers were starting to call it GFAS — Good First Answer Syndrome. You find one answer that seems right and you stop searching. This isn't an AI problem. It's a human problem that AI inherited from us.
George laid it out for Jimbo over a laptop and two beers: the DOT's 2021 fixes and the AI-safety recommendations of 2023 are the same list. Rotate inspectors so no one goes stale; require two people to verify critical components; mandate the thorough method so you can't shortcut it; document with photos; bring in outside audits to verify the verifiers — that's exactly verify sources, cross-check results, external validation, show your work. Same problem, different domain. Humans pattern-match to save time; AI pattern-matches to generate faster; both miss the critical thing when the pattern becomes the only answer. The fix is the same everywhere: never let one source verify itself. Always question the first answer. Because first answers are often wrong — but they're comfortable, and humans, just like the machines they built, will choose comfortable over thorough if you let them.
The framework it teaches
Same city, same lesson