In highly regulated industries such as finance or medicine, artificial intelligence could represent a risk.

More specifically, products that use LLM or Large Language Models such as ChatGPT, Gemini or Claude. The above mentioned are multimodal products that use LLM, but in this case we are going to focus on their LLM aspect.

Let's get straight to the point: artificial intelligence is hallucinating; but what does that mean? The technical explanation is that if a user gives an instruction to the model, if it is not well trained, the transformer decodes and interprets the instruction incorrectly and gives you an incorrect answer. IBM explains it like this.

More concretely: You give it an instruction and the model gives you the most likely answer based on its training data, but it is far from reality — look at these examples:

• A medical AI agent that tells me I have COVID when I actually have the common flu.

• A financial AI agent that tells me that in order for me to do better financially all I have to do is rob a bank

The first is the wrong answer and the second is simply illegal, but correct. We cannot act with either of them. The AI ​​agent is effectively an extension of the company and as a company it is critical to be able to advise users in a correct, ethical and legal direction. It is important that users do not go to jail.

This happens because LLMs think probabilistically. The factors that define probabilities are:

• The training data

• Reinforcement processes

• The weights of its parameters

And from there comes the next word or idea as “the most likely”.

Easier said than done: the solution is to have a team specializing in artificial intelligence, who has the ability to architect the way AI trains and thinks. Shameless promotion: at Moneda we can provide this service.

Piero Silverio
Brand & Business Strategy

Hector Terrero