Help, my AI ʇɹǝıuıznןןɐɥ!
An AI Hallucination is when a large generative language model (LLM) generates false information or facts that do not correspond to reality. The hallucinations often appear plausible - at least at first glance - as fluent, coherent texts are generated.
However, it is important to emphasise that LLMs do not deliberately lie, but simply have no awareness of the texts they generate.
Large language models tend to be very confident in inventing new (false) information.
Thora Markert
Head of AI Research and Governance at TÜVIT
The technical reasons for AI Hallucinations can be manifold:
It is therefore possible for LLMs to generate hallucinations even though they are based on consistent and reliable data sets.
The containment of hallucinations is therefore one of the fundamental challenges for AI users and developers. This is because LLMs are usually a black box, which can make it difficult to determine why a particular hallucination was generated.
The term AI Hallucinations covers a broad spectrum: from minor inconsistencies to fictitious information. Types of AI hallucinations include
Sentence contradictions
Generated sentences contradict previous sentences or parts of the generated response.
Contradictions with the prompt
The generated response or parts of it do not match the user's prompt.
Factual contradictions
Information invented by the LLM is sold as fact.
Random hallucinations
The LLM generates random information that has nothing to do with the actual prompt.
If users rely too much on the results of an AI system because they look very convincing and reliable, they may not only believe the false information themselves, but also spread it further.
For companies that use LLM-supported services as part of customer communication, there is also a potential risk that customers will be provided with untrue information. This, in turn, can have a negative impact on the company's reputation.
LLMs are powerful tools, but they also come with challenges such as the phenomenon of AI hallucination.Through comprehensive audits, we support AI developers in identifying and minimising existing risks in the best possible way and further strengthening confidence in the technology.
Vasilios Danos
Head of AI Security and Trustworthiness at TÜVIT
The easiest way to recognise or unmask an AI hallucination is to carefully check the accuracy of the information provided. As a user:in a generative AI, you should therefore always bear in mind that it can also make mistakes and proceed according to the "four-eyes principle" of AI and human.
In order to counteract AI Hallucinations and other challenges posed by AI systems, it is advisable to have appropriate tests carried out by independent third parties. In the best case scenario, vulnerabilities can be identified and rectified before applications are officially deployed.