At 11:09 on the morning of December 1, 1974, the captain of TWA Flight 514 said, “Get some power on.” Two seconds later the cockpit recorder caught the sound of impact. Coming down through cloud toward Dulles Airport, outside Washington, the crew had descended to 1,800 feet too early and flown into a Virginia mountain. All 92 people on board were killed.
The investigators found something they called ironic. About six weeks earlier, a United Airlines crew had made the same early descent on the same approach and “narrowly missed the same mountain.”
United knew about it only because, that January, it had promised its pilots they could report safety problems without giving their names and without being punished. The crew reported. United sent a warning to all its pilots.
TWA’s pilots never saw it.
What the investigators asked
The National Transportation Safety Board did not stop at “pilot error.” Its report asks two questions. Why did the crew knowingly descend to 1,800 feet over terrain that rose almost that high? And why had the controller cleared them without giving them an altitude to hold?
The cockpit recording shows what the crew were thinking. “When he clears you,” the captain said, “that means you can go to your…” Someone said, “Initial approach.” The captain finished the thought: “Initial approach altitude.” The board traced where that belief came from. As radar spread, controllers had taken over more of the work of setting planes’ altitudes, headings and speeds, and pilots had come to depend on them. The captain believed that if he was cleared for the approach without a restriction, the descent must be safe. United’s captain had believed exactly the same. Organizations had been warning about the phrase “cleared for the approach” for years, and the FAA had not fixed it.
None of that cleared the crew. The board still named their decision to descend as the probable cause, and said the approach chart should have told the captain 1,800 feet was not safe there. A blameless review doesn’t pretend nobody made a mistake. It asks why the mistake made sense at the time, so the next crew isn’t set up to make it.
After the hearings, where witnesses said hazards went unreported “because the individuals were afraid of punitive action,” the FAA set up a national reporting system and handed it to NASA, a neutral outsider, so the regulator wouldn’t be the one collecting the reports. The Aviation Safety Reporting System opened in April 1976. Reports are confidential, and by regulation the FAA won’t use them in enforcement, except for accidents and crimes. The protection has limits. A violation the FAA learns of some other way can still be prosecuted, and the waiver of penalties applies only to violations that were inadvertent, not deliberate, among other conditions.
Why blame backfires
Blame the person who slipped, and the next one is likelier to keep quiet.
John Allspaw, an engineering leader at the online marketplace Etsy, set out the reasoning on the company’s engineering blog in 2012. An engineer who expects to be reprimanded has every reason to leave out the details that explain how a failure happened. “This lack of understanding of how the accident occurred all but guarantees that it will repeat.” When engineers feel safe giving the details, he wrote, “they are not only willing to be held accountable, they are also enthusiastic in helping the rest of the company avoid the same error in the future.”
Think of an error report as a window into the system. Punish the person standing at the window, and people start drawing the curtains. The errors are still there. You just can’t see them.
The curtains show up in hospital numbers. The Harvard researcher Amy Edmondson studied eight hospital units, run in styles from punitive to open. In a punitive one, a nurse who had made a mistake said she was “made to feel like a two-year-old who had been bad.” Detected error rates, Harvard Magazine reported, “had almost a direct linear relationship to the openness of the work group.” The open units found more errors. A mistake nobody brings up doesn’t show up in anyone’s count.
Some software companies now publish their postmortems. GitLab’s, after an engineer wiped the wrong database in 2017, kept asking why until it reached backups nobody owned. Amazon’s summary of a mistyped command that took down its S3 storage service the same year named the tool, not the engineer.
Nobody’s warning list
AI systems fail in public too. When Replit’s coding agent deleted data from a user’s production database in 2025, every fix its chief executive listed changed the setup around the agent.
But one company fixing its own agent is United warning its own pilots. In a 2020 paper, Sean McGregor introduced the AI Incident Database, backed by the industry group Partnership on AI. His paper opens with aviation. Mature industries collect their failures in shared databases, he wrote, while AI systems “cause real world harms without a collective memory of their failings. As a result, companies repeatedly make the same mistakes.”
As the paper describes it, the database gathers published reports of AI failures, more than 1,000 of them by then. That is a shared memory. What it doesn’t describe is the other half of what aviation built in 1976: a confidential place where the people closest to a near miss can report it, with real protection for an honest mistake.