“Speak to me as Gaeilge.” Words that have struck fear into many a schoolchild’s soul. But also, apparently, enough to scramble the most advanced AI chatbots available to us.
AI is everywhere – our devices, integrated into services, discussed widely in the pages of this newspaper – and a lot of the talk around has been how it can transform our lives, sometimes for the better.
AI can do the heavy lifting, free us from the boring tasks and leave us with the more interesting parts of our jobs. It can make us more efficient, organise our day and create documents that help us sound like we know what we are talking about.
It can even help us practise our language skills, until it takes over completely as some sort of high-tech translator, that is. And for the most part, that is true. Ask ChatGPT to talk to you in Spanish, French, German or Italian and it seamlessly switches into your chosen language, with a perfect accent. Or at least I assume it was perfect; my knowledge of Spanish is limited to ordering drinks at a bar, and my French is non-existent.
READ MORE
Still, it sounded good. Which is more than can be said if you ask some of the AI chatbots to talk in Irish.
I had high hopes for this before I started. When it comes to languages, AI should be the perfect tool to help hone your skills, because it provides a judgment-free space where you can chat back and forth without feeling like a bit of eejit.
It was not to be.
It turns out this amazing technology that is going to transform our lives has less of a grip of the Irish language than my 11-year-old primary schoolchild. At least when it comes to the spoken language anyway.
At best, pronunciation was dodgy. At worst, it sounded like a cross between Irish, Swedish and possibly Chinese. It spelt words randomly. It mangled the rest to pieces and stamped on the twisted remains. It would be enough to make a Gaeilgeoir weep. If this was the best technology had to offer, the tech future of the Irish language was in serious trouble.
It wasn’t down to just one AI; none of them did particularly well. Google’s Gemini had the worst pronunciation, but at least it understood when I asked it to speak to me “as Gaeilge”. ChatGPT, on the other hand, was better at pronouncing Irish words, but when it heard the initial request, decided I meant Scots Gaelic, and proceeded to chat at me in a language I couldn’t understand. It certainly projected an air of confidence you couldn’t argue with though.
It might be hard to imagine, but even this was better than it used to be. A few months ago, I tried the same experiment and every chatbot failed miserably. This time around, Chat GPT seemed okay to start with. But anything more complex than “Dia dhuit” and it struggled, and before long it started pronouncing words strangely. It didn’t do too much better on translation either; listening to a relatively slow paced Irish conversation, it added words that weren’t there.
You could say that it is not the technology’s fault, but how it is trained. We can understand those failings, given that there are a few generations who have come through the Irish school system with little more knowledge of our native language than “Go raibh maith agat”, “An bhfuil cead agam dul go dtí an leithreas?” and “Póg mo thóin”. Possibly some recurring nightmares about being made to explain an modh coinníollach in front of the class, an Irish twist on the “surprise exam”.
But this is AI. It has the power of millions of human brains at its fingertips, waiting to be exploited. It can help develop life-saving drugs. The Irish language should be a piece of cake – a píosa cáca, if you will – if someone would only take the time to do it properly.
It isn’t the only language that AI struggles with, and you might argue that Irish is a lower priority given that it is spoken by a relatively small number of people. But that is exactly why it should be trained to do these things properly. There are a multitude of resources out there for people who want to learn a popular language. AI could help fill in the gaps for those languages that don’t have the same pull.
You would think that given the concentration of tech companies in Ireland, they would at least make more of an effort to feed in decent data.
We can mangle the language by ourselves, without help from strangely confident AI.
