The railway AI market is forecast to grow from $3.5bn to more than $6bn within four years, yet McKinsey finds that only 25% of companies surveyed have implemented multiple AI use cases at scale. Roughly 35% more run just one or two applications in production, with the rest still in pilot stage. The gap between market value and deployed capability frames an analysis by Nikita Chuyko, director of right-of-way diversion planning and integration at the Massachusetts Bay Transportation Authority, on where generative and discriminative AI actually fit rail preventive maintenance.

Global AI investment could reach $2.5 trillion in 2026, a 44% increase since 2025, according to the analysis. Passenger rail, freight rail and transit agencies have captured comparatively little of that implementation effort. Chuyko's argument is blunt: operators that cannot distinguish between the two families of AI tools risk buying capability they cannot use.

What is the distinction that matters?

Discriminative and predictive AI tools model conditional probability. They answer one class of question: given a defined set of parameters, what is most likely to happen? In rail, that means detecting squats, classifying corrugation severity, forecasting delays and estimating maintenance demand for track and rolling stock.

This is not new technology. A 2017 paper, "A Big Data Analysis Approach for Rail Failure Risk Assessment," combined ultrasonic measurement with camera images processed by Deep Convolutional Neural Networks to detect severe squats. A 2020 study, "Research on deep learning method for rail surface defect detection," layered additional object-detection algorithms onto DCNN-based video analysis to grade defects from corrugation, fatigue and stripping.

The limitation is structural. A well-trained discriminative model improves only marginally with more data, and its output remains locked to the specific task. It cannot, as Chuyko notes, be asked to compare its squat-detection approach against methods used by other organisations or generate a reading list. Operators buying such tools should expect precision, not flexibility.

Where does generative AI earn its place?

Generative models learn joint probability distributions over vast datasets and can weigh a wider range of variables. In maintenance, the analysis identifies three concrete applications:

• summarising standard operating procedures;
• generating training videos from SOPs and training materials;
• compiling inspection reports that merge inspectors' video, images and voice into a single document.

Two named operators claim results. SNCF, the French national operator, stated that predictive maintenance is a strong fit for generative AI, using Internet-of-Things sensors on rolling stock to significantly reduce downtime. Deutsche Bahn implemented BahnGPT in early 2025 for specialised applications in rail operations and maintenance. Neither figure — the downtime reduction, nor BahnGPT's measured effect — is quantified in the analysis. McKinsey's report cites maintenance co-pilots, service scheduling and early network infrastructure digital twins as further examples.

What could go wrong?

Chuyko identifies three failure modes. Sycophancy — the model's tendency to agree with a user who is wrong — carries real cost in a field where correctness is at a premium. Hallucinations persist even in current models, though less severely. Knowledge cutoff is the quietest risk: a model trained on stale data will confidently report false information if a track segment has been built or a fleet retired without retraining. That is a manageable issue for continuously retrained commercial models, but a critical oversight for a rail organisation building its own capability without a retraining procedure.

The analysis also makes a point often skipped in vendor material: reactive maintenance is sometimes the right answer. Where a failure carries low criticality, no significant safety risk and a straightforward fix, deferring work may be optimal. Preventive maintenance earns its priority only when failures would cause major disruption, safety incidents, long parts lead times or heavy labour burden.

What comes next?

Chuyko's prescription is selective deployment: discriminative AI for detection, categorisation and risk analysis; generative AI to turn those findings into recommendations and maintenance plans. He expects leading organisations to invest further in agentic AI that combines both. The claim is a projection, not a measured result, and the analysis offers no cost or return figures for it. What the evidence does support is narrower: railway AI spending is rising faster than railway AI adoption, and the operators that close that gap first will do so by matching each tool class to the specific failures that cost them most.