AI still makes mistakes when answering very simple topics

Earlier this year, Scott Gutterman, chief digital officer of the PGA Tour, witnessed a ChatGPT error when answering the number of times Tiger Woods had won the tour, demonstrating the limits of artificial intelligence, as described in Wall Street Journal.

Although AI models are trained on large amounts of data, they often lack specific knowledge, leading to inaccurate responses.

As companies embrace AI, many are realizing that generic models like those offered by OpenAI need to adapt to industry-specific data to be truly useful. However, this customization process can increase costs and complexity, requiring tight data control.

AI still needs human review when it responds to something

• The PGA Tour now uses an approach called Recall Augmented Generation (RAG) to reduce errors by incorporating specific information, such as round rules, directly into its AI queries.
• All generated responses are reviewed by humans before being released.
• However, RAG has limitations and is best suited for low-risk missions.
• For more important cases, such as agricultural advice, “fine-tuning” of models using proprietary data is recommended, although this is expensive and requires expertise.

Read more

Some companies are exploring the possibility of creating custom models to understand specific nuances, such as home loans. As industry-specific AI models emerge, the responsibility for personalization remains in the hands of companies.

In the case of the PGA Tour, Guterman highlighted the importance of teaching AI to differentiate between event wins and major wins, as ChatGPT mixed up this information by claiming that Tiger Woods had won 15 times on tour, when in fact he had won 82.