AIFoPa-2026-0027 — A Hundred and Forty-Six Cars

On 16 June 2026, near Connell, Washington — a town of some five thousand one hundred people thirty-five miles north of Pasco, established in 1883 as a stop on the railway and sustained since by food processing, agricultural chemicals and the state corrections system — a train was lined onto a section of track occupied by a roadway maintenance worker. The train was nine thousand four hundred feet long, weighed six thousand five hundred tons, ran to a hundred and forty-six cars, and carried hazardous materials in at least twenty of them, which is the threshold at which the applicable rules oblige a railroad to call it a key train. What the materials were has not been established publicly. The vice general chairman of the signalmen’s BNSF general committee, Brandon Pettit, says only that it could have been any of the substances that pass through the Pasco classification yard in the ordinary way of things.

The instrument that lined it is called Movement Planner. It is supplied to BNSF by Wabtec, it consists of a set of algorithms which plan the movement of trains and then line the switches and signals to execute those plans automatically, and it does so across the majority of the BNSF system without prior authorisation from a train dispatcher. On the day in question, according to the signalmen, it bypassed the safety restrictions and lined the hazardous train northbound out of Pasco onto occupied track. A human dispatcher saw what had happened and took the signal down before the train could be ushered onto it. The Brotherhood of Railroad Signalmen wrote to the Federal Railroad Administration on Wednesday 26 August to say that what had been prevented would otherwise have been a “public catastrophe” for the worker, for the crew on the train, and for the community of north Franklin County.

BNSF describes the same minute differently, and the Bureau sets its account down at length because the difference is the whole of the dispute. A train dispatcher, the company says, manually entered an instruction to line a switch for a train movement “milliseconds” after the software executed the same queued instruction, and the result was that the switch remained lined inconsistently with the track authority the dispatcher had granted. The company adds that this is precisely what its layers are for: “Like all railroads, BNSF relies upon overlapping layers of measures that include engineered solutions, operating rules, and training to prevent incidents from occurring. That approach worked in this instance as the dispatcher immediately and correctly resolved the issue.” It notes, and the Bureau has no reason to doubt it, that there has never been a train accident on BNSF caused by a defect or failure of its computer-aided dispatching systems. Mr Pettit’s reply to that last proposition is recorded here because it is the crux and not a rhetorical flourish: no accident has been attributed to the system because on each occasion a person corrected it.

The two accounts agree on what the switch did. They differ on whose hand was nearest it at the time, and the Bureau, which files machinery rather than blame, observes that on either account the instruction the software executed was one the software had queued, and that the arrangement in which a machine and a person can issue the same instruction milliseconds apart with an inconsistent result is itself the finding. The signalmen put this in the language of the trade: the system, they say, does not comply with the fail-safe closed-loop principle, and is on that basis in breach of federal regulation as it stands. “What we’re asking,” Mr Pettit said, “is that they don’t turn this system back on, period.”

It had been turned back on once already. The union says BNSF took Movement Planner offline shortly after the June incident and reactivated it some weeks later on the Pasco–Spokane corridor; the company stopped using it again this month, after the Federal Railroad Administration asked that it be shut down and at the request of the American Train Dispatchers Association. That association’s president, Ed Dowell, called the incident a wake-up call, and his organisation has told the regulator something the Bureau regards as the second finding of this record: that BNSF had known of the vulnerability for about a year before Connell, and that Wabtec had estimated the fixes would take some fourteen months. “If defects capable of producing this type of failure are known but remain in service,” Mr Dowell said, “that raises serious questions about risk management and regulatory oversight.” Two similar mishaps occurred with the dispatching system in 2022, both of them also in Washington state. The signalmen tried on both occasions to have the Federal Railroad Administration look into them and were unable to preserve the system logs.

The Bureau ordinarily has the advantage of wreckage. An incident that has completed itself leaves a figure attached to it — the derailment at East Palestine, Ohio, in February 2023 involved twenty cars of vinyl chloride and other hazardous materials and an eight-hundred-million-dollar cleanup — and a figure is a thing a record can be built around. This file has no figure. It has a signal that was displayed for a short time and then was not, and a worker who has no idea, presumably to this day, of the afternoon he had. The archive is accordingly obliged to note that the cheapest incidents to ignore are the ones that were stopped, and that the railroads are at this moment lobbying for regulatory changes that would let them replace human employees with systems of this kind. The signalmen’s letter makes the arithmetic explicit, and the Bureau, having checked it, can find nothing in it to correct: the artificial intelligence made the potentially catastrophic error, and an experienced human employee identified and prevented it.