Are AI Models Doomed To Always Hallucinate?

Large language models (LLMs) are AI models used to power chatbots such as Google Bard and ChatGPT. However, LLMs have been observed to produce false information, which is hardly ever questioned because these models are statistical systems designed to generate fluent and coherent texts from words, images, etc. LLMs hallucinate because they do not understand the reality the language is describing; they only generate grammatically correct words according to the prompt. In simple terms AI hallucinations occur when LLMs misinform and cook up facts, and these ‘lies’ can be plausible, clearly nonsensical, or a mix of real and fictional data. This is not a malicious act by the LLM but is a result of its inability to estimate its prediction uncertainty.

Popular examples of AI hallucinations include Google Bard’s untrue statement about the James Webb Space Telescope, and Meta Galactica’s citation of a fake paper on a topic. AI hallucinations can be little inconsistencies and fabrications, to huge falsifications or contradictions. These hallucinations can be sentence -contradictions, prompt contradictions, factual contradictions, and random, or irrelevant hallucinations. Harmless hallucinations such as claiming that in 2016, the Golden Gate Bridge was moved across Egypt are better than harmful, sometimes deadly, or costly claims by AI. A popular one occurred recently when an Australian mayor stated that OpenAI’s ChatGPT claimed that he pleaded guilty in a serious bribery scandal.

RELATED: WhatsApp Starts Testing AI-Generated Stickers

Do All AI Models Hallucinate? Why They Do, And Why It’s A Problem

Yes, all LLMs hallucinate, and will always do so. According to TechCrunch, researchers believe that model hallucinations can be used as a way to spread malicious code packages, especially to naive code developers. AI hallucination occurs due to the way LLMs today (and all other generative AI) are developed and trained. Some of the reasons researchers agree on include:

  • Quality of data: If the AI is fed bad source content information, hallucinations will occur. If the data contains errors, biases, noise, or falsehoods, the output would be a hallucination. A good example of this is the case of Reddit being used as a training data source for ChatGPT.
  • Methods of training and generation: If the data set is reliable, but there exists bias created by the model’s former generation, hallucination may occur.
  • Context of input: Unclear, contradictory, or inconsistent input can lead to AI hallucinations. This is why AI prompting is such an important skill to hone nowadays.

AI hallucinations are a big problem for both the AI companies and the users of these LLMs. This is because they erode user trust and foster feelings of disappointment, can spread harmful misinformation, be used in cyberattacks, and anthropomorphize inanimate models to some extent. This situation is further compounded by the fact that LLMs and the hallucination phenomenon are relatively new.

Even though model outputs are designed to sound plausible, users should read them with skepticism in order not to fall for hallucinations, which can be categorized as black box AI. Determining why an AI model generated a specific hallucination may sometimes be impossible, and trying to change the training data would consume a lot of energy and expense. Users have the onus to watch out for AI hallucinations themselves.

RELATED: Survey Discovers That ChatGPT Is Used (or Feared) By Only A Small Number Of Americans

Detecting And Preventing AI Model Hallucinations

Users can detect AI hallucinations by fact checking all output from LLMs. When working with unfamiliar or complicated material, users can instruct the model to self-evaluate, generate correctness probabilities, and highlight areas that might be untrue; as a basis for checking facts. Familiarizing themselves with information sources for the LLM, will enhance their fact checking efforts. For example, fact checking ChatGPT answers based on knowledge after 2021 would be smart, as its training data stops at 2021.

Users can reduce the occurrence of AI fibs/ hallucinations by doing the following:

  • Using specific and clear prompts: Users should ensure that their prompts are clear, unambiguous, and concise. Users should include context and step by step instructions in their prompts.
  • Strategy filtering and ranking
  • Using multi-shot prompting to provide examples of the desired output to help the model.

AI model hallucinations seem to be an unavoidable feature of LLMs for now, but ongoing research is trying to understand the why of occurrence, and the how of mitigation. OpenAI recently proposed an approach called process supervision, where AI models are rewarded for each correct step of reasoning instead of just rewarding the answer. This aims at training the models to follow a chain of thought approach to problem solving. Some researchers have also proposed instructing two models to collaborate until they come up with an answer.

NEXT: OpenAI- backed Language Learning App ‘Speak’ Raises $16M  To Expand To The U.S.

Leave a comment