Skip to main content

Hallucinations of AI and the quality of Inputs

 



In this article, we discuss why Generative AI hallucinates. Is it based on the quality of training data? How can we get output that has minimum hallucination? Is Artificial Intelligence biased? How to Mitigate Bias in Generative AI? We discuss these in this article.

Introduction

You should understand first how AI Models work. AI takes input from innumerable sources of information such as web pages on the World Wide Web, Books, Journals, Articles, and so on. These data are stored as nodes in Neural Networks. Generative AI just predicts the next most probable word. The next Node connection path it should take to churn out the answer to your question. That means no data scientist can ever examine or predict the output of Generative AI. It is just connections between words. How will you examine connections between words? 

Hallucinations of Generative AI

It is a common misbelief that AI Large Language Models hallucinate and give out non-existent information. It does not make up any story. It just reflects the information it received as input. That input is being weighted too. That means that information is re-iterated many times from many sources, based on the information prevalence in its sources. If you ask for rare data, it is less weighted, meaning it is not re-iterated from many sources. The authenticity of that data depends on those few sources. If that source is authentic, you get reliable information, if not, you get wrong information, which everyone calls hallucinations.

Bias in Generative AI

Biases in Generative AI work the same way. It is not targeting any ethnic group or community because it doesn't like them. It just reflects what is there in the World Wide Web and its other sources. You may not like the facts that are on the World Wide Web, like criminal cases in the courts by certain groups of people or the poor jobs they do. But facts are facts. It just reflects the truth, which is being re-iterated many times over. If you want to handle biases, these data have to be manipulated in ways people like to view. This method is a burden on LLMs and may really lead to hallucinating stories people want to tell. Most of all, we need fictional writers to write each of these stories. However, we can use filters and Retrieval Augmented Generation (RAG) to deal with the biases in the generated output.

Prompt Engineering

The solution to hallucinations is to write the correct prompt. You should explicitly tell Generative AI to give out information according to your needs. For example, there was an allegation from a lawyer, who tried a case in the US court, using Generative AI. He presented a case with arguments from previous cases like X versus Court, and Y versus Court. They found that all the cases Generative AI produced were hallucinations. They were not in fact hallucinations but poor prompting from the Lawyer. He should have explicitly asked the AI to give out real cases, with links to proofs. AI would have simply taken cases from Websites that churned out spurious information or simply it would have considered some cases from Fiction Stories and Novels. How can you blame AI for that? Always give an elaborate Prompt with complete details you want in the output. Proper Prompt Engineering is the only solution.

Quality of Inputs

This is an elaborate topic and area of research. The reason for the so-called hallucinations and probably some errors is because of the input that is fed into the LLMs. How to source only reliable information or at least rank the best sources of input first. Google has page ranking. But I guess LLMs have already done most of it. So, instead of going back to the start line, we can re-iterate the output of LLMs with first-quality data and present it to the end user. This is an area of research Data Scientists can presume for the development of Artificial Intelligence, which impacts every field and helps to increase employee productivity.

Photo Courtesy: freepik.com


Support Us: https://supportusbydeepak.blogspot.com

Labels:
Location: Chennai, Tamil Nadu, India

Comments

Post a Comment

Popular posts from this blog

Minimize Hallucinations by Updating LLMs with RAG

Large Language Models can be updated by using Retrieval Augmented Generation of High-Quality Data like Journals. This should be done on topics that the user asks using Gen AI. Unnecessary data that no one cares about need not be updated. This updated LLM fragment should be marked as high quality with more weightage given to it. If there is time-bound data, stale data should be marked, and RAG should be performed again every time or whenever necessary. Problem with Present Large Language Models The problem with present LLMs is in the frantic attempt to train LLMs, maximum internet was covered without considering the quality of the data in it. We need to rectify it. But it is impossible to start from the beginning and correct everything. Also, there is no need to do it. We can concentrate on the data that is most needed. Hallucinations Hallucination does not come from somewhere. It is the improper data fed into the LLMs during the training of the model. If we need to minimize hallucinati...
Post a Comment
Read more

AI learning methods other than Brute Force

  In this article, we discuss the various learning methods by which Artificial Intelligence can learn. We discuss methods like learning a subject completely by branching out to different topics and learning a subject starting from a high-quality source and then to a low-quality source of data available on the internet. Introduction We know that Generative AI tools like ChatGPT and Bard already have learned a good part of their knowledge by brute force. We are suggesting a few methods by which it can learn otherwise. These methods should help it to optimize itself, if not learn from scratch. Re-Inforced learning by adding weights ChatGPT and Bard should add weights to their important sources of information. These sources could be the most visited websites, the most referred websites, acclaimed books from renowned authors, research papers or articles from well-known magazines, Articles, and even the most watched YouTube Channels. These sources need to be given weights and they should...
Post a Comment
Read more