Large Language Models hallucinate: Here's what you need to know

LLMs are a specific type of AI model that is specifically developed to comprehend natural language. They have the capability to process and generate text, enabling their use in various tasks like language translation, summarisation, and question-answering.

Training these models involves using extensive datasets, the size of which depends on the intended purpose of the LLM. Smaller datasets generally consist of tens of millions of parameters, while larger ones contain hundreds of billions of data points. Despite all the training, these LLMs are not always 100 percent right.

Also read: What powers ChatGPT and Bard? A look at LLMs or large language models

In LLMs, parameters represent the ‘knowledge’ gained by the model during its training phase. More parameters generally result in more accurate predictions because the model has access to more contextual information. Some of the most prominent LLMs have hundreds of billions of parameters.

Hallucinations in LLMs

Hallucinations are a significant ethical concern that occurs when the model generates outputs that are grammatically and logically correct, but are disconnected from reality, often based on false assumptions. The issue of hallucination becomes particularly worrisome as LLMs become increasingly convincing, leading users without sufficient domain knowledge to excessively rely on them.

For example, a model may mistakenly attribute a quote to a celebrity, one that was never actually said. Additionally, these models might assert false information about a topic or cite non-existent sources in research papers, thereby spreading misinformation.

We’ve already seen the worrisome effects start to play out in the real world too.

Recently, a New York-based lawyer found himself in a challenging situation when the use of ChatGPT led to the inclusion of fabricated legal cases in his filing. The chatbot generated a series of fictional cases and arguments, leading to serious consequences for the lawyer.

To understand hallucination in AI models, we asked ChatGPT the question of who the founder of OpenAI, its maker, was.

The response was mostly correct, but it made one incorrect statement in the last sentence. The sentence said that Sam Altman, the CEO of OpenAI, stepped down on March 1, 2021, and co-founder Greg Brockman, who is President of OpenAI, took over the position.

Of course, this is incorrect. Sam Altman is still CEO of OpenAI.

Why do LLMs hallucinate?

Hallucinations in language models can occur due to limited or insufficient training data. To learn accurate language representations, models require diverse and comprehensive datasets. Inadequate or narrow training data can hinder the model's ability to grasp language nuances, leading to inaccurate responses and the generation of fabricated information.

Additionally, biases and prejudices can manifest in language models, contributing to hallucinations. These biases originate from biases present in the training data, which may reflect societal prejudices.

The form of generative AI that we are familiar with — chatbots such as ChatGPT and Bard — may seem to be generating coherent, relevant information, but their form of functioning is similar to that of an auto-complete mechanism.

As Vianai founder Vishal Sikka earlier told us, it is a regurgitating machine that generates the next words and next sentences from the dataset it has been trained on.

“When you ask it a question, it has really no idea about that question itself, it simply continues that sentence, adds another word, adds another sentence and so on. Because of the way we are, we think that it is this highly intelligent system that is communicating with us but it really has no idea. It is simply completing your sentences based on a synthesis of all that it has been trained on,” he said.

Effectively, LLMs utilise weights, statistics, and contextual cues to predict the most suitable next word or response based on the given prompt, previous tokens, and relevant data.

How to prevent hallucinations in LLMs

To address the issue of hallucinations in LLMs, researchers are employing different strategies. These approaches encompass enhancing the diversity of training data, mitigating inherent biases, and also implementing more effective regularisation techniques.

An instance of these efforts is evident in the collaboration between OpenAI and Google's DeepMind, where they introduced a technique called reinforcement learning with human feedback (RLHF) to combat the hallucination problem in ChatGPT.

This method involves the participation of a human evaluator who regularly assesses the model's responses and selects the most suitable ones for user prompts. The feedback provided by the evaluator is then utilised to fine-tune the behaviour of the model.