The Right Data: Eliminating AI Bias and Hallucinations

By: Verbit Editorial
AI Image Generator on a screen
Filters

Artificial Intelligence and generative AI have dominated the headlines for the past year as businesses and institutions around the world have discovered the broad benefits that programs like OpenAI’s ChatGPT and DALL-E can provide. Their popularity shows no signs of slowing any time soon.

A recent report by Forrester suggests that global spending on AI software will reach $64 billion by 2025, up from $33 billion in 2021.

Verbit recently introduced its own generative AI product suite – Gen.V – which combines natural language processing, content generation abilities and AI to identify the most important elements of dialogue from transcripts. Gen.V is being used to analyze transcripts for customers and extract key information. It provides automatic summaries, keywords, SEO highlights and headline suggestions to name a few.

Though many are embracing AI technology and its growing potential, there is concern over the accuracy of the information that some generative AI tools return. AI models are advancing, but they still can make mistakes and produce incorrect answers, either via “hallucinations” or by demonstrating an “AI bias.”

people looking at images of others using technology

AI Bias and Hallucinations

Large language models (LLMs) are a type of AI designed to mimic human intelligence. They analyze incredible amounts of data – books, articles and web pages – and process the patterns and connections between words and phrases. They use this information to generate new content, such as blog posts, essays or simple social media posts.

Machine learning depends on the quality, objectivity and size of training data used to teach it. The quality of the output is influenced by the quality of the input – or, in simpler terms, think of the adage “garbage in, garbage out” – and incorrect or incomplete data fed into the system can result in inaccurate information being returned.

AI bias occurs when an AI model produces content based on datasets that contain human biases. Typically, AI bias stems from problems introduced by the individuals who train the machine learning systems. For instance, the information used to train the system could contain unintended cognitive biases, real-life prejudices, or, as noted above, incomplete datasets.

One commonly cited example of bias in generative AI is image generation tools that promote outdated stereotypes. For instance, producing an image of a middle-aged white man in a business suit when asked to generate a CEO or a young woman when asked to generate a flight attendant.

An AI hallucination is when an AI model generates incorrect information but presents it as if it were a fact. These hallucinations can occur for a variety of reasons, including:

  • Outdated/incorrect training data. An AI program is only as good as the data it’s trained on. If the AI doesn’t understand its prompt or doesn’t have sufficient information, it will rely on the limited dataset from its training to generate a response.
  • Limited training. If the AI is trained on a limited dataset, it may be unable to properly produce new data, resulting in non-factual outputs.
  • Poor prompts. Sometimes user prompts – either poorly worded or deliberately confusing – can result in AI hallucinations.
a computer with the ChatGPT homescreen

How Verbit is Eliminating AI Bias and Hallucinations

Verbit uses LLMs – like OpenAI – that have bias levers in place, and supplements them with additional levers of our own.

Our LLMs employ several techniques to identify and remove bias from training data, including, among other things:

  • Using human reviewers to flag biased text
  • Employing algorithms to identify and remove patterns of bias
  • Using debiasing techniques during data training, including adversarial training that teaches LLMs to be resistant to prompts designed to trick them into generating biased output, and counterfactual data augmentation that creates new training data by swapping bias attribute words such as he/she in a dataset.
  • Filtering output with post-processing techniques, including moderation API

As product accuracy and customer satisfaction are key concerns at Verbit, we also take additional steps to restrict incorrect outputs, including:

  • Limiting the scope and directing the LLM to a specific context or transcript, thus lessening the chance of bias. For instance, we train our customer-specific AI on customer-specific and customer-supplied materials. Our Gen.V solution for education clients ‘learns’ from existing classroom documents, course materials, previous lectures and any additional information that the school provides.
  • Conducting sample testing of the AI results to identify and prevent bias
  • Providing editing tools so customers can verify AI-produced results before sharing

OpenAI recently stepped up its efforts to remove hallucinations by developing new training methods for its AI models. The new process teaches the AI model to reward itself for each correct step of reasoning when arriving at an answer instead of just rewarding it for a correct conclusion. Programmers believe this process will result in better information since it encourages the model to follow a more human-like thought process and enables it to be more capable of solving challenging reasoning problems.

Rely on Us

The concerns over AI bias and hallucinations highlight the important role that humans play in this developing technology. As Verbit applys generative AI to our customers’ transcripts and captioning files on a regular basis, our team believes in maintaining the essential combination of technology and human professionals. Contact us to learn more about how we can put our gen AI product to work for you and find a helpful solution that best fits your needs.