AI hallucinations can create serious issues in areas like healthcare and law, but experts say they can’t be completely eradicated.
Sometimes artificial intelligence (AI) just makes things up and presents these hallucinations as facts.
This happens in both commercial products like OpenAI’s ChatGPT and specialized systems for doctors and lawyers, causing real-world problems in courtrooms, classrooms, hospitals, and more by spreading misinformation.
Despite these risks, many companies are eager to use AI, with 68% of large companies adopting at least one AI technology, according to British government research.
So why does AI hallucinate, and can we stop it?
Understanding AI Hallucinations
AI hallucinations are instances where artificial intelligence systems generate false or misleading information and present it as fact. These hallucinations can cause serious issues in areas like healthcare and law, leading to significant real-world problems by spreading misinformation.
Why AI Hallucinations Matter
AI is increasingly being integrated into various sectors. According to British government research, 68% of large companies have adopted at least one AI technology. This widespread adoption makes understanding and mitigating AI hallucinations crucial to prevent the spread of false information in critical areas such as courtrooms, classrooms, and hospitals.
A Little Personal Experience
I’m a writer, editor, and content manager, so I’ve seen a lot of extra “crazy” since the explosion of AI. Artificial intelligence is a powerful tool, there’s no denying it, but it can also be dangerous and chaotic in the wrong hands. You’ve got huge publications replacing their entire writing staff with AI models and cheap “ai wranglers” to cut costs, but that’s not the way AI is supposed to be used – it’s a productivity tool, not a replacement for human brains.
This misuse can lead to the dissemination of AI-generated hallucinations, which can be mistaken for facts. Now, yes, it’s pretty funny to read some of these hallucinations presented as facts by people who haven’t bothered to edit or check their work, when the content is obviously flawed and made up. But the danger lies in the hallucinations that seem plausible, making it easy for people to accept them as true without verification. And the hallucinations are real.
Why Do AI Systems Hallucinate?
Generative AI tools like ChatGPT are built on large-language models (LLMs) that work through ‘pattern matching.’ This means an algorithm looks for specific shapes, words, or sequences in the input data, like a question or task. However, the algorithm doesn’t understand the meaning of the words. It might seem intelligent, but it’s more like pulling Scrabble letters from a bag and learning what gets a positive user response.
These AI systems are trained on vast amounts of data, but incomplete data or biases can lead to hallucinations. All AI models experience hallucinations. Even the most accurate ones show factual inconsistencies 2.5% of the time, according to AI company Vectara.
The Real Dangers of AI Hallucinations
Depending on where AI is used, the effect of hallucinations can range from farcical to severe. For instance, after Google struck a deal with social media platform Reddit to use its content to train its AI models, its Gemini tool started pulling incorrect advice or jokes from the site – including a recommendation to add glue to cheese to make it stick to pizza.
In courts, lawyers have cited non-existent cases generated by AI chatbots numerous times, and the World Health Organisation has warned against using AI LLMs for public healthcare – saying data used to reach decisions could be biased or inaccurate.
How Can Hallucinations Be Reduced?
Reducing hallucinations involves improving the quality of training data, using humans to verify AI output, and ensuring transparency in how models work. However, implementing these processes can be challenging as companies are often reluctant to share their proprietary tools for inspection.
Many large AI companies depend on poorly paid workers in the Global South to label text, images, video, and audio for AI training. The work is long, exhausting, and often under poor labor regulations. LLMs can be fine-tuned to reduce hallucinations by using Retrieval-Augmented Generation, which enhances AI’s answers with external sources. However, this method can be expensive due to the necessary infrastructure, such as cloud computing space, data acquisition, and human managers.
Alternatively, using smaller language models could reduce hallucinations because they can be trained on more complete and specified data. This approach also has a smaller environmental footprint. Despite these efforts, experts from the National University of Singapore believe that hallucinations can’t be completely eliminated.
“It’s challenging to eliminate AI hallucinations entirely due to the nature of how models generate content,” the researchers wrote in a paper published in January.
