Despite the increasing possibilities of artificial intelligence, there are still often noticeable errors in the responses of AI chatbots. OpenAI is now researching to improve this.

The popularity of artificial intelligence chatbots has skyrocketed since OpenAI’s ChatGPT. The learning models that AI trains with are extensive, but not error-free and omniscient. It often happens that systems like ChatGPT or Google Bard simply write something themselves if they don’t know the answer and “sell” these inventions as correct information.

Some examples

At Google Bard, this became immediately painfully clear during the performance. The system then claimed that the James Webb Space Telescope captured the first image of a planet outside our solar system. That is not true; That honor goes to the European Southern Observatory’s Very Large Telescope in Chile, which captured an image of an exoplanet in 2004. That’s almost seventeen years before James Webb came into being.

Even the mere sucking of the digital thumb can have legal consequences. A New York attorney recently considered bolstering the evidence in favor of his team in his client’s lawsuit against an airline by bringing before the court a number of similar cases in a voluminous filing. However, the man used ChatGPT to put these things together, and it turns out the system had put them together one at a time. The judge immediately ordered a hearing on sanctions against the attorney. The outcome of the legal dispute is not yet known, it is still ongoing. But the plaintiff will no doubt be less reassured after this debacle.

These examples undoubtedly also play a role when the CEO of OpenAI warns the world about the negative sides of artificial intelligence. However, that didn’t stop the company from launching an official ChatGPT app for iPhone in the meantime.

Solution

To avoid these situations, OpenAI will develop improved and new training methods. The idea is to reward the AI ​​for reliable results and prevent unwanted results. The intent is not to limit this to the end result, but to apply it to every step; “Process Monitoring” versus “Outcome Monitoring”. This would make the thinking of the system more transparent.

According to the OpenAI researchers, this way of working has several advantages. By monitoring the entire process and not just the outcome, the AI ​​argues much less “the end justifies the means”. According to the researchers, human surveillance at each step leads to a process that can be interpreted more clearly. The disadvantage of this method is that it slows down the training of the AI. How to reward AI for something hasn’t been revealed, but it won’t be with a reward.