RESPONSIBLE AI
A( PI) barrier to hallucinations In terms of the LLM integrations, Gorden says“ we try and facilitate quite deep integrations”. He explains that means native integrations rather than just relying on scraping and his team encourages the LLM providers to use Checkatrade’ s APIs directly“ so that they don’ t really have an opportunity to hallucinate”. Gorden says:“ They’ re not pulling from the stored context information. They’ re actually doing a real-time search and they’ re feeding that information back to the consumer. So, for example, if you think about Anthropic and ChatGPT, we’ ve got native integrations with both of those.”
“ We’ ve got user interface cards that represent trades. It’ s deeply integrated, which I think reduces that chance of hallucination quite significantly. On our platform side of things, we always make sure that any information we’ re showing always goes through some kind of a guardrail mechanism.”
Gorden acknowledges the issues of persistent AI hallucinations.“ I think the key for me is we can’ t prevent LLMs from making things up,” he says.“ At the end of the day, it’ s not a platform that we control. If someone’ s looking for that type of information, it can come natively to us and get the information in a very structured manner. If it’ s able to do that, it doesn’ t need to hallucinate. That’ s how we try to facilitate this process as much as possible.”
aimagazine. com 153