AI-driven chatbots often exhibit a concerning tendency to generate false information while presenting it as accurate—an issue referred to as AI hallucinations. This phenomenon carries various negative consequences, from undermining the advantages of artificial intelligence to posing potential real-world harm to individuals.

As generative AI continues to gain prominence, concerns about AI hallucinations have intensified. In response to these concerns, a team of European researchers has been diligently working to develop remedies. Recently, this team introduced a promising solution, with the potential to significantly reduce AI hallucinations to single-digit percentages.

The innovative system at the heart of this solution originates from Iris.ai, a startup based in Oslo. Established in 2015, Iris.ai has engineered an AI engine specializing in the comprehension of scientific text. This software conducts in-depth analyses, categorization, and summarization of vast volumes of research data, offering valuable insights.

One notable user of Iris.ai’s platform is the Finnish Food Authority, which employed the system to expedite research related to a potential avian flu crisis, saving researchers up to 75% of their time.

However, AI hallucinations present a significant challenge. Current large language models (LLMs) are notorious for generating nonsensical and inaccurate information. These inaccuracies can have adverse consequences, including reputational damage and, in more severe cases, the propagation of harmful advice or malicious code.

Victor Botev, CTO of Iris.ai, notes, “Unfortunately, LLMs are so good at phrasing that it is hard to distinguish hallucinations from factually valid generated text. If this issue is not overcome, users of models will have to dedicate more resources to validating outputs rather than generating them.”

AI hallucinations also hinder the value of AI in research. A survey conducted by Iris.ai among 500 corporate R&D professionals found that only 22% trust systems like ChatGPT, yet a significant 84% still rely on ChatGPT as their primary AI tool for research.

Iris.ai addresses this issue through a multifaceted approach, focusing on fact-checking, semantic validation, and answer coherence. To verify factual correctness, they map out key knowledge concepts they expect to find in a correct response, comparing the AI-generated answer against these criteria and assessing their source reliability.

Semantic validation involves assessing the AI output’s similarity to a verified “ground truth” using a proprietary metric called WISDM, which evaluates topics, structure, and key information.

Another technique examines the overall coherence of the answer, ensuring it incorporates relevant subjects, data, and sources, enhancing answer quality and accuracy.

The integration of these techniques establishes a benchmark for factual accuracy and coherence, ensuring AI outputs closely resemble those of human experts.

Iris.ai has incorporated this technology into a new Chat feature within the company’s Researcher Workspace platform. Initial tests have demonstrated a significant reduction in AI hallucinations to single-digit percentages.

While the approach appears effective on the Iris.ai platform, scaling it for popular LLMs poses challenges related to user interpretation and understanding of results. To mitigate AI hallucinations, it is essential to enhance AI’s decision-making process, making it more transparent and explainable.

Notably, Microsoft has introduced the Phi-1.5 model, which uses high-quality, synthetically generated, and web-sourced data. This approach aims to reduce AI hallucinations by providing well-structured training data that promotes reasoning.

Additionally, addressing bias in training data is crucial. Victor Botev suggests training models on coding language, which emphasizes reasoning and leaves less room for interpretation, guiding LLMs toward factually accurate answers.

Despite its limitations, Iris.ai’s approach represents a step in the right direction, adding transparency and explainability to AI. By enhancing understanding of AI model processes and fostering external collaborations to build larger datasets and metrics, the field can continue to reduce AI hallucinations in the future.

For Botev, this work is crucial for building trust in AI: “It is to a large extent a matter of trust. How can users capitalize on the benefits of AI if they don’t trust the model they’re using to give accurate responses?”

By Impact Lab