Will ChatGPT Revolutionise Railway Signalling?

There has been a lot of publicity around ChatGPT and similar large language model (LLM) AI tools in recent months, so I thought I would take a look at how it will impact Signalling (certainly in the UK).??

HOW DOES ChatGPT WORK??

They work by analysing language-based content including descriptions, rules, regulations and specifications and applies a degree of logic to extrapolate the answer to the posed question. It does this through publicly available information and resources (although I would suggest that if you had a large enough data-pool this could be a private data-pool) looking for patterns and commonalities from which it can draw baseline attributes from which to derive a response.??

If you then pose it a question, it takes the key words and phrases and searches for all available information in that domain, ranks it by relevance and draws out answers based on common features and rules.?

RISKS AND CHALLENGES?

Firstly, the data has to exist – if the information is maintained in a closed databased (such as NRG) ChatGPT cannot see it?

Secondly, it has to exist in a digital form or one that is accessible automatically by a digital system – if records or regulations are on paper or acetates, the ChatGPT cannot see it?

Thirdly, it has no way to verify if the information is complete or trustable – the user needs to do this?

Finally, it is based on language not symbology so if you have data stored exclusively in a non-verbal schema (say a scheme plan) it cannot read it (although this is the logical next step in development).??

Any of these can disrupt the LLM doing its thing.??

COULD YOU APPLY IT TO OUR PROCESSES??

Short answer: No.??

Design is based on logical diagrams with a varied heritage and a mixture of COTS and bespoke systems which mean most signalling implementations in the UK are bespoke – the more bespoke, the less commonality. The principles of signalling are the same but the implementation is not, AI would be able to apply generic rules but not account for local specifics.??

It also cannot do stakeholder engagement where motivations and demands are not obvious or easily extractable.??

Even if we could use the AI to design signalling, the existing system assurance infrastructure in the UK does not have the technical facility to assure a product developed or checked by an AI device. Nor does it really have the time to upskilling effectively given the volume and pace of work required over the next 10-15 years. This doesn’t mean it cannot be done – just that it cannot be done right now.?

CONCLUSION?

This generation will not see the benefits or impact of ChatGPT – but the next one might. The current generation of CCS is too opaque to be read by any concentrated AI but if we are intelligent about how we build our integrated records database (common data environment, in the correct terminology) we might be able to use AI to support design development and decision-making in the future.?