BTS releases SafeMTS AI study
The Bureau of Transportation Statistics published a SafeMTS report showing AI and large‑language models being used to analyse maritime near‑miss data, demonstrating how advanced analytics can extract safety insights from incident reports. That approach is adaptable to transit near‑miss and incident learning if agencies can structure comparable datasets (x.com).
Most safety systems wait for a crash, fire, or grounding and then ask what went wrong. The Bureau of Transportation Statistics is trying to learn from the moments just before that, using a maritime program called Safe Maritime Transportation System and a new April 8, 2026 study on artificial intelligence tools that read near-miss reports at scale. (bts.gov) A near miss is the shipping version of a driver almost running a red light and stopping inches short. The Bureau of Transportation Statistics says these narrowly avoided events are early warnings that can prevent more serious maritime incidents if companies report them before anyone gets hurt. (bts.gov) Safe Maritime Transportation System is a voluntary, confidential reporting program built by the Bureau of Transportation Statistics and the Maritime Administration with industry partners. Its 2023 pilot report said the goal was to create a shared source of “precursor” safety data, meaning warning signs that show up before a catastrophe. (rosap.ntl.bts.gov) That confidentiality piece is not a footnote. In an August 8, 2023 Federal Register notice, the Bureau of Transportation Statistics asked approval for a new data collection because shipping companies were being asked to submit sensitive near-miss information to a government-run analysis system. (federalregister.gov) The hard part is that near-miss reports are usually written in plain language, not tidy spreadsheet boxes. One captain might write “loss of situational awareness,” another might write “confusion on the bridge,” and a third might describe the same problem in 200 words. (bts.gov) Large language models are the software behind chatbots that can sort messy text into consistent categories. The new Safe Maritime Transportation System report tested whether those models could classify free-text maritime reports, map them to a shared safety taxonomy, and then hand the results to human reviewers. (bts.gov) The study did not pitch a machine that replaces investigators. The Bureau of Transportation Statistics says the workflow used human-in-the-loop review, meaning subject-matter experts checked the model’s work so automated analysis could be faster without turning into an unchecked black box. (bts.gov) The report says the payoff was less manual sorting and more comparable data across companies that describe the same hazard in different ways. It also says the models were tested in a secure protected environment because these reports can contain operational details companies do not want exposed. (bts.gov) The catch is that artificial intelligence only works as well as the records it gets. The Bureau of Transportation Statistics says the barriers to wider use were inconsistent data formats, incomplete records, and companies reporting on different schedules, which makes pattern-finding harder even before a model starts reading. (bts.gov) That is why this maritime study reaches beyond ships. The Bureau of Transportation Statistics’ maritime safety page says the same approach is part of a broader precursor safety data effort, and the lesson is simple: if transit agencies, rail systems, or other operators can collect near-miss reports in a structured and confidential way, language models can help turn thousands of scattered anecdotes into one usable safety map. (bts.gov)
