How Accurate Is a Chatbot?

A chatbot can generate new text, learn, and reason, but it doesn’t understand meaning and relies on statistical patterns and the datasets used. Incorrect information in the dataset can lead to inaccurate and misleading responses.
 

Datasets

The choice of datasets depends on the chatbot’s specific purpose and application. It’s essential that the datasets used are high-quality, contain accurate information, and are representative of the chatbot’s intended task.
 

MC4 Dataset

Many companies don’t disclose the sources used to train their chatbots. Google collected texts from 2.5 million unique Dutch and Belgian websites for the MC4 dataset, a widely used dataset for chatbots.

Research on the Most Used Dutch Dataset*

De Groene Amsterdammer and Data School investigated websites from the MC4 dataset. They concluded that the most used Dutch websites for chatbot training are full of copyright violations, personal data, and fake news.

According to the research, Docplayer.nl is the primary source for chatbots—a hosting site that collects documents, including files with personal data such as job evaluations, tax returns, and intelligence reports. Docplayer.nl was long a goldmine for hackers seeking privacy-sensitive data.

Notably, a “neo-Nazi conspiracy site” called Stormfront ranks in the top 200, just one spot below RTL Nieuws. This suggests chatbots may learn equally from both sites.

Also concerning is the presence of sites like Marktplaats and eBay, where users leave phone numbers—data that gets linked by systems aiming to learn and generate new text.

If a chatbot’s data source isn’t reliable, it can spread misinformation without traceability to the original source.

*) Source: Research: The Sources of ChatGPT – De Groene Amsterdammer