AI chatbots have become a part of the daily lives of millions of people, but few even understand how they work. Portal theconversation.com shared Five facts about how AI really works and how to best use them.
The bot learns from people's responses
AI chatbots are trained in multiple stages, and the first stage is pre-training, where the models learn to predict the next words in a large amount of text. This process allows them to achieve a common understanding of language, facts, and logic.
If you ask a pre-trained model any questions, it can easily provide detailed instructions – even if the user asks something suspicious. To make AI useful and safe for humans, annotators help guide the model's responses in a more effective direction—a step called “matching.”
After this period, the AI can refuse to answer any malicious questions, citing the law or other reliable sources of information. Without additional customization, chatbots will be unpredictable: there will be nothing to stop them from spreading misinformation or harmful content.
Bots learn not through words but through tokens
Humans naturally learn language through words, but AI uses small units of data called tokens for this purpose. They can be words, subwords, or even just strings of letters. And while the encoding process often follows logical patterns, it sometimes produces unexpected results that reveal both the strengths and weaknesses of the AI's ability to interpret language. For context, modern chatbots typically have a vocabulary of 50,000 to 100,000 tokens.
Knowledge about bots becomes outdated every day
AI chatbots do not continuously update themselves, which makes it difficult for them to answer any queries related to recent events, new terms, and information other than expiration dates—or the last date the training database was updated.
For example, the current version of ChatGPT has a deadline of June 2024. If you ask who is the current president of a particular country, the AI itself will have to go to the search engine, “read” the results and give an answer based on them. Many other AIs also use search engines to ensure their answers are as up-to-date as possible. Updating AI chatbots is a complex and expensive process. Improving the effectiveness of updates remains an open scientific problem.
Bots are susceptible to hallucinations
AI chatbots sometimes suffer from “hallucinations” – meaning they produce false results or simply make them up with complete confidence in their accuracy, as the model predicts text based on patterns without checking the facts. Such errors are a result of the internal structure of the AI; they prioritize coherence over accuracy, rely on imperfect training data, and completely lack understanding of the real world. While ChatGPT has received some improvements, such as a fact-checking tool and the ability to directly ask the bot to cite reliable sources, there are no tricks that can prevent the illusion in the first place.
Bots do math using computers
Recently, some chatbots have switched to the so-called. a logical reasoning model that allows them to break down a problem into individual components to solve a complex problem. For example, if you ask ChatGPT “What is 56,345 minus 7,865 multiplied by 350,468”, the AI will give the correct answer because it “understands” that multiplication comes before subtraction. And to perform calculations, the bot uses a built-in calculator, which allows you to achieve accurate results.
