Hallucination Reduction

Programs like ChatGPT write text by appending, word by word, the most probable next word. This sometimes produces something that sounds fluent and convincing but is simply invented. Such invented content is called a hallucination in technical language. Typical examples are invented book titles, incorrect dates, or legal provisions that never existed. Hallucination reduction is the umbrella term for all techniques intended to make such errors less frequent. They cannot be eliminated entirely so far, but their frequency can be significantly reduced.

Why invented facts become costly

A spelling mistake is immediately noticeable. An invented source citation is not. That is exactly what makes hallucinations dangerous: they look like correct answers. The text is grammatically clean, the number sounds plausible, the tone appears confident. Anyone who does not verify it will adopt the error unchecked.

In practice, there have already been some spectacular cases of this. Lawyers in the United States submitted briefs containing court rulings that an AI had invented. The court imposed fines. Airlines and government agencies, too, have already had to answer for false information given by their chatbots.

For companies, this is the main reason why AI systems are often deployed only hesitantly. In medicine, law, or accounting, an incorrect statement is not a minor cosmetic flaw. Hallucination reduction therefore also determines in which areas such systems are permitted at all.

From source integration to self-verification

The most effective method is called Retrieval Augmented Generation, or RAG for short. In this approach, the system searches a genuine collection of documents, such as a company’s manual, before answering. The passages found are attached to the question. The model is then supposed to answer only from this material and not from the vague memory of its training. In a sense, answering from memory is replaced by an open-book exam with a reference work.

A second approach targets training. Humans rate answers and explicitly reward the model when it admits not knowing something. Without this training, a system learns that a confident answer is usually rated better than an admission of not knowing. It is precisely this tendency to guess that must be specifically trained out of it.

In addition, there are downstream verification steps. A second model reads the answer and compares each claim with the cited sources. If something does not match, the answer is discarded or flagged. A related idea is to ask the same question multiple times: if the answers diverge significantly, that is a warning sign of invented content.

How to recognize it in products

Source citations with clickable links are the most visible sign. Search systems like Perplexity or the AI answers in search engines rely on this. Every statement can be traced back to the original page. Sentences like “I don’t have information on that” are also a result of these efforts and not a weakness of the system.

In corporate products, hallucination reduction is usually invisible in the background. A customer service chatbot accesses the internal knowledge database instead of formulating freely. For sensitive questions, it hands off to a human. This is unspectacular, but that is exactly the point.

In news reports and vendor marketing materials, percentage figures often appear, such as “40 percent fewer hallucinations.” Such figures should be read with caution. They refer to specific test questions and cannot be generalized to every application. A common misconception is also that a larger model automatically hallucinates less. Larger models know more, but they also formulate their inventions more convincingly.