Trainline "AI" assistant: incorrect advice

[AlterEgo]

This might actually be the worst thing Trainline has done.

 

[Bletchleyite]

This might actually be the worst thing Trainline has done.

Agreed. I have defended them about all manner of things in the past against onslaughts from staff on X and the likes, but this is just grossly negligent - I can't think of another term for it.

 

[signed]

I wonder most how they justify that commercially. Being that, unlike most other company LLM, it doesn't seem to replace/augment the support system.

Though being Private Equity, I am not struggling to guess how.

 

[mikeg]

The management decisions behind procuring and releasing this stuff baffles me.

The Pointy Haired Boss character from Dilbert didn't come from nowhere... Honestly generative AI probably has a place but most of the time Richard Stallman gets it right - it's a bull**** generator and that's the last thing you want for giving advice that could have legal consequences.

 

[sh24]

If they did, I've got a feeling that if a case came to court, the court would take a very dim view of that tactic:

"Here's our helpful AI chatbot to give you advice about your ticket!" **

** Any advice given may be complete nonsense, and if you follow it you may be liable for an excess fare, a penalty fare or prosecution.

Disclaimers like this - as I understand it - are only legally valid under consumer law if they are considered fair and reasonable.

Correct. And I'm pretty sure that isn't going to be fair and reasonable.

AI is, without a really clean dataset, genuinely awful for most technical or precision based applications. Great for pretty pictures and writing emails, dismal for stuff that needs accuracy.

 

[Egg Centric]

WHY on earth would you use a LLM, which is general-purpose, behind a robot working in a specific domain such as GB rail ticketing???

This domain has specific rules, admittedly weird in places, validation requirements for retailing, a knowledgebase and so on. Surely you train the model on that body of data. Then it should spew out the relevant restriction (in human-speak such as 'you can use any train on a Sunday') and totally understand passenger types adult/child/railcard/combinations.

Not necessarily. While you could fine tune it, you could also implement this fairly easily with a general purpose model. In simple terms the LLM should not be reasoning itself but instead summarising some clear information spit out by a conventional piece of software (which I'm assuming Trainline already have) about the ticket (and if that piece of software can't accept a query about, say, age simply refusing to answer rather than coming up with nonsense). See this link for how it's implemented on one popular API

 

[etr221]

By coincidence I came across a piece "What goes on inside an LLM", at https://www.askwoody.com/newsletter/free-edition-what-goes-on-inside-an-llm/ - while it's very somplified, when you read it, you realise why it's totally unsuitable for what Trainline are doing with it.

 

[Bletchleyite]

Not necessarily. While you could fine tune it, you could also implement this fairly easily with a general purpose model. In simple terms the LLM should not be reasoning itself but instead summarising some clear information spit out by a conventional piece of software (which I'm assuming Trainline already have) about the ticket (and if that piece of software can't accept a query about, say, age simply refusing to answer rather than coming up with nonsense). See this link for how it's implemented on one popular API

Indeed. What you want here is for the chatbot to have a human-like conversation (which is what a LLM can do quite well) but using referenced facts in the response rather than just the LLM inferring things. This really is rather incompetent.

 

[island]

Ultimately, AI is just a more advanced version of the "three words" that show up on your phone/tablet/other device to choose from as what it thinks are the most likely next word in your sentence. Factual accuracy is optional.

 

[BRX]

I'm going to follow this thread just to find out how long it takes for the inevitable wave of complaints & disputes this will cause, to catch up with Trainline and lead them to swiftly withdrawing it.

 

[Benjwri]

In simple terms the LLM should not be reasoning itself but instead summarising some clear information spit out by a conventional piece of software (which I'm assuming Trainline already have) about the ticket

I’m not so sure about this. They could get it doing the basics, which it can’t now, such as explaining railcard restrictions, but otherwise it’s use cases are limited. Maybe it could interpret the public restriction text better in some cases, but I’d say that’s unlikely.

I don’t see how it could parse the existing electronic information, you’d end up with it telling people they can’t depart Reading on London Underground and LNER before 9:30!

 

[Egg Centric]

I’m not so sure about this. They could get it doing the basics, which it can’t now, such as explaining railcard restrictions, but otherwise it’s use cases are limited. Maybe it could interpret the public restriction text better in some cases, but I’d say that’s unlikely.

I don’t see how it could parse the existing electronic information, you’d end up with it telling people they can’t depart Reading on London Underground and LNER before 9:30!

It doesn't have to parse the existing electronic information. Trainline will (I assume) already have something that does that. What it would need to do is essentially convert the user's question into a query of that information which would then be passed into this already existing thing, and then interpret the results the existing thing will spit out. I'm highly confident it would work.

 

[Bletchleyite]

It doesn't have to parse the existing electronic information. Trainline will (I assume) already have something that does that. What it would need to do is essentially convert the user's question into a query of that information which would then be passed into this already existing thing, and then interpret the results the existing thing will spit out. I'm highly confident it would work.

Yes, that is the principle of just using AI to create a chatbot that basically has a conversation like a human would (passes the Turing test, in essence) but using it to provide accurate, referenced information in response to a query, retrieving it by way of a simpler search. AI absolutely can do that (products like ServiceNow will do that for you) and that is how it would be set up in a competent business. Clearly here Trainline are just using a general AI engine so they can tick it off on their "buzzword bingo" card and that really isn't good enough.

 

[Benjwri]

It doesn't have to parse the existing electronic information. Trainline will (I assume) already have something that does that. What it would need to do is essentially convert the user's question into a query of that information which would then be passed into this already existing thing, and then interpret the results the existing thing will spit out. I'm highly confident it would work.

Do they have the ability to do this? I haven’t used Trainline recently but last time I used that feature (Years ago) it seemed to refund and reissue my ticket, and it seemed to just do a new search and compare prices rather than what the ticket was actually valid for.

 

[Egg Centric]

Do they have the ability to do this? I haven’t used Trainline recently but last time I used that feature (Years ago) it seemed to refund and reissue my ticket, and it seemed to just do a new search and compare prices rather than what the ticket was actually valid for.

Someone like @Adam Williams would know better about the kind of things accreddited retailers need but I assume ticket validation is part of it (although it's possible it's sufficient to be able to find valid tickets rather than the other way around). Even if they don't have it they certainly would have the expertise to build it.

 

[OscarH]

Someone like @Adam Williams would know better about the kind of things accreddited retailers need but I assume ticket validation is part of it (although it's possible it's sufficient to be able to find valid tickets rather than the other way around). Even if they don't have it they certainly would have the expertise to build it.

They need to be able to surface valid tickets in the search results of course, but the kind of backwards search from tickets is quite a different flow that you need to implement specifically. Much harder if your journey planning is outsourced too and they don't support it

 

[MisterM]

Seems like there’s already a news outlet that noticed it:

Trainline’s new AI customer assistant veers off the rails

Among other errors, bot insists an adult can travel with child’s ticket

observer.co.uk

 

[styles]

“Our customer support team remains on hand. Our priority is helping customers cut through the complexity of rail.”

What, by giving them incorrect information?

 

[Adam Williams]

Anyone know if the standard bike reservation is enough for taking my moped (under 50cc, of course) on board a Voyager, then?

 

[MrJeeves]

Seems like there’s already a news outlet that noticed it:

Trainline’s new AI customer assistant veers off the rails

Among other errors, bot insists an adult can travel with child’s ticket

observer.co.uk

Nice to see a publication run with a story about it.

It's shameful to see the largest third-party retailer releasing such a joke of a product.

 

[duffield]

I wonder if anyone's been penalised yet for following its "advice"?