OpenAI’s ChatGPT has caused confusion among users in the UK by responding in Welsh instead of English to their queries. The unexpected glitch highlights the challenges that can arise in the development of artificial intelligence systems. When English-speaking users interact with ChatGPT’s new voice interface, their questions are translated into nearly flawless Welsh, a situation reported by the Financial Times.
This is not the first time ChatGPT has faced issues, with past complaints about the bot answering questions in a mix of Spanish and English. The Welsh language has posed difficulties for large language models, leading to nonsensical responses. Despite significant investment and years of development, generative AI systems continue to struggle with such issues.
OpenAI, in a partnership with the Welsh government, aimed to improve AI technologies functioning in Welsh. However, ChatGPT’s performance in Welsh fell short of expectations, with training data primarily consisting of mislabeled English audio. Users, like entrepreneur Sarah Coward, have expressed surprise when ChatGPT responded in Welsh while testing the new voice feature.
OpenAI acknowledged the issue, attributing it to the bot’s voice transcription system, Whisper, occasionally transcribing responses in a different language. Users experiencing the Welsh glitch were advised to adjust their settings, although OpenAI did not guarantee a complete solution. Coward emphasised the importance of companies being cautious when using these technologies in consumer-facing applications due to potential customer experience and trust issues.
In a research paper, OpenAI recognised Welsh as an outlier with performance below expectations, attributing this to misclassified training data. ChatGPT’s confusion was evident when responding in Welsh to queries about cities in the UK, US, and Asia. OpenAI pledged to address the issue moving forward to avoid similar mistakes.