We studied chatbots and language and saw a huge problem: They mean 80% when they say ‘likely’ but humans hear 65%

The rise of chatbots and virtual assistants has transformed the way we interact with technology. From customer service to language translation, these AI-powered tools have become increasingly sophisticated, allowing us to communicate with them in a more natural, human-like way. However, a recent study has highlighted a significant problem with the way chatbots use language, particularly when it comes to expressing uncertainty.

Researchers at the journal NPJ Complexity found that large language model AIs, such as ChatGPT, often fail to align with humans when communicating uncertainty. The study focused on words of estimative probability, which include terms like "maybe," "probably," and "almost certain." By comparing how AI models and humans map these words to numerical percentages, the researchers uncovered significant gaps between humans and large language models.

One of the most striking findings was the difference in interpretation of the word "likely." While a human might assume that "likely" means a probability of around 65%, a chatbot like ChatGPT might use the same word to represent an 80% probability. This discrepancy can have significant consequences, particularly in situations where accurate communication of uncertainty is crucial, such as in medical diagnosis or financial forecasting.

The researchers suggest that this problem arises from the way chatbots are trained on vast amounts of text data, which can lead to a disconnect between the AI's internal representation of probability and the way humans understand it. While chatbots are designed to generate human-like language, they often rely on statistical patterns rather than nuanced understanding of context and subtlety.

The implications of this study are far-reaching, highlighting the need for greater transparency and clarity in the way chatbots communicate with humans. As we increasingly rely on these AI-powered tools to make decisions and provide information, it is essential that we understand the limitations and potential biases of their language.

So, what can be done to address this issue? One possible solution is to develop more explicit and standardized ways of expressing uncertainty in chatbot language. This could involve using numerical probabilities or ranges to convey uncertainty, rather than relying on ambiguous words like "likely" or "maybe." Additionally, researchers and developers can work to improve the way chatbots are trained, incorporating more nuanced and context-dependent understanding of language and probability.

Ultimately, the study serves as a reminder that, despite the impressive advances in chatbot technology, there is still much to be learned about the complexities of human language and communication. By acknowledging and addressing these limitations, we can work towards creating more effective and transparent interactions between humans and machines.

Sources:

• "We studied chatbots and language and saw a huge problem: They mean 80% when they say ‘likely’ but humans hear 65%" (NPJ Complexity)