Artificial Intelligence (AI) “hallucinations” occur when an AI model provides information to the user that is not factually correct. Notable examples of AI hallucinations include Google’s Gemini telling users to eat rocks for their health and put glue on their pizza, as well as the tendency for image models to add extra fingers to human hands or push the boundaries of human proportion. 

Humorous as they may be, AI hallucinations have real consequences. While AI hallucinations are particularly concerning to users who may be unfamiliar with the subject matter they’re researching (such as Google’s AI bot “Bard” incorrectly presenting details about the James Webb telescope), they can even fool expert users. A BC lawyer was caught citing fake cases after the citations were provided in ChatGPT. 

Although human oversight could have caught these misattributed facts (and would certainly catch some of AI’s less intelligent renderings of hands), AI hallucinations can seem very real and convincing to the average user. Due to this, the user must be careful not to take the information at face value, as the model can present false data and insights that seems true. However, human oversight is not the only way to catch these AI hallucinations: many facets of a well built AI tool are designed to effectively prevent hallucinations from happening. 

Why do AI hallucinations occur? 

AI hallucinations happen most often because of the statistical nature of large language models. Unlike human intelligence, AI models do not truly “understand” information. Like the autocorrect on your phone stretched to a much larger scale, AI predicts patterns and generates a response based on these predictions. This predictive tendency is why asking an early LLM what to drink at a gas station might get you a response of “gasoline.” 

It is also why asking a model like ChatGPT the sum of 2 plus 2 isn’t the same as punching it into a calculator yourself. If 2+2=3 started appearing in responses across the training data, ChatGPT might begin seeing 3 as a plausible response. Gaps or data quality issues (such as bias) can also create AI hallucinations, as can a lack of constraints on the potential outcome of the query. 

How to mitigate AI hallucinations

1. Utilize Retrieval Automated Generation (RAG)

One of the most popular methods to mitigate AI hallucinations is retrieval automated generation (or RAG). This approach works to improve the reliability of AI models by retrieving information from a predefined set of data or documents – for example, a knowledge base of research papers from a particular medical journal. 

Consider this in the context of the example above of the AI model liable to suggest “gasoline” as a beverage choice at gas stations. With RAG, the model might first look at predefined menu options available: say, orange juice, soda, sports drinks, or water. Once these options have been retrieved, the model generates a response (“orange juice is the most popular drink in this region”). In the case of the lawyer who cited legal cases that did not exist, a RAG model would have only selected cases from a predefined database of options. 

Tools like RAG help to reduce the likelihood of AI hallucinations, since it gives the model a defined set of options – all human verified and periodically updated. Models designed to operate on RAG systems provide much more trustworthy, and sourced, responses. 

2. Rely on extractive AI for key inputs

Abstractive AI models take in key data insights and interpret, rephrase, and utilize these insights to generate text. Abstractive AI can create a medical records summary from a volume of documents so that a user gets a “big picture” look at the medical file in a fraction of the time. This is handy when a high level understanding is all the user needs, and especially handy when pressed for time. 

However, for users looking to make professional decisions based on AI generated data, it’s best to ensure that this data is true to the original source. Medical records can be indexed, organized, and highlighted to share key insights from the file without creating any new information. This respects the fidelity of the original source data and makes it easier for professionals to know and trust the data that they use. 

3. Utilize closed, high quality datasets 

Users in the medical, legal, or claims industries often have the resources to support high quality AI outputs already. Proprietary (and de-identified) patient datasets, peer reviewed studies, official health guidelines, policy manuals, and regulatory guidance makes up the “rulebook” AI models can play by. Using closed datasets built around trustworthy sources ensures that outputs will be trustworthy, too. Without the ability to introduce new information from unreliable or unvetted sources (which is responsible for the more humorous hallucinations mentioned earlier), these datasets can reliably give users the insights they need, every time. 

‍Human oversight and keeping humans “in the loop” of AI assisted workflows is crucial to avoiding hallucinations in your AI tools – but there are other ways to safeguard your data as well. While human oversight should always be a final fail-safe, it is possible to mitigate and reduce instances of AI hallucinations enough to consistently and confidently rely on AI in your processes.