The Global Rail AI Index

01 / SCOPE

What belongs in the index?

An entry should identify an organisation, a concrete railway-related application and a public source describing the use of AI or a specific AI technique. Coverage includes railway operations, infrastructure, rolling stock, passenger services, engineering, manufacturing and construction.

• A specific application: the description should explain what the AI does and the railway problem it addresses.
• Evidence of AI: a general claim about digitalisation, sensors or automation alone is insufficient to establish an AI application.
• A traceable source: the source should support the application described, including any stated deployment stage or result.

Each row describes an application associated with an organisation and country. One organisation or source may appear in several rows when different applications are described. Categories group applications by their main purpose; they are editorial labels, and some applications span more than one topic.

These criteria guide additions and review. Existing entries vary in the detail and quality of their evidence; inclusion does not certify that every claim has been independently verified.

02 / PROJECT MATURITY

An experiment and an operational system are different

Read the application description alongside its source. Project maturity is currently expressed in the text, rather than a separate, consistently populated status field.

An announcement or planned rollout is evidence of intent. Where it appears in an entry, it should be read as planned activity unless the source also documents actual use. If the stage is unclear, it should not be inferred from the organisation’s name or the presence of a source link.

03 / EVIDENCE

What a source check means

Sources include railway operators, infrastructure managers, suppliers, research organisations and industry publications. Direct project documentation is preferred when available. Supplier statements and press releases are attributed reports, not independent evaluations of performance.

1. Read for relevance. Check whether the source supports the named organisation, application, AI involvement and stated project stage.
2. Keep claims within the evidence. Preserve distinctions between reported results, expected benefits, trials and operational use. A result from one setting should not be generalised to all deployments.
3. Record supported dates. Keep publication, start of use and editorial access separate. Leave undocumented dates empty.

Checks are manual. Visiting the index does not re-check its sources or update their access dates. A successful retrieval records access to a page; it does not independently verify the project’s performance or current operating status.

04 / DATES

Three dates, three different meanings

Deployment date

The explicitly documented start of use of the described application, including pilot operation. It is not automatically the announcement or article date. “Not documented” means no supported start date has been recorded.

Source published

The publication date of the linked source. An article may describe an application that began much earlier. “Not documented” means no supported publication date has been recorded.

Source last accessed

The recorded date of a successful editorial retrieval, with the legacy year-only exception explained above. “Not yet recorded” means no successful-access date is recorded; it does not establish whether the link works today.

Dates preserve the precision available in the evidence: YYYY, YYYY-MM or YYYY-MM-DD. A year alone does not imply 1 January. “Date needs review” indicates an invalid recorded date, not a project status.

05 / LIMITATIONS

How to interpret the collection fairly

• Coverage is incomplete. Public availability, language, search visibility and editorial research priorities affect which applications are found. Absence from the index is not evidence that an organisation does not use AI.
• Counts measure entries. They do not measure investment, adoption rates, deployment scale or national AI leadership. Several entries may relate to the same wider programme.
• Country labels need context. A supplier’s home country, a research partner’s location and a deployment site may differ. Consult the description and source before drawing geographical comparisons.
• Evidence varies. A supplier claim, a research result and an operator’s deployment report provide different kinds of support. Listing an application does not endorse a supplier or establish its effectiveness or safety.
• Information can age. Projects may change, end or expand after publication. Manual maintenance does not provide continuous monitoring or a guaranteed update schedule.

Use the index to discover examples and starting points for research. Follow the linked evidence before relying on an entry’s technical claims, maturity or reported benefits.