Generative AI hype soared in 2023. Today, that hype has transformed into action and most enterprises are using the technology in some capacity.
While some organisations may not have determined the best use cases yet, business leaders remain optimistic about its potential. The vast majority (86%) of CEOs surveyed by Gartner believe AI can help maintain or grow company revenue.
However, several quirks are holding adoption back. One of these is the tendency for GenAI to completely make things up.
What are AI hallucinations?
AI “hallucinations” stem from the way large language models (LLMs) are trained. The models are fuelled by enormous datasets that contain both reality and fiction. GenAI uses probabilistic algorithms to guess what the user desires based on prompts fed into it. If the data is bad or the platform can’t recognise the context, it can return incorrect and sometimes egregious results.
These hallucinations could be obvious lies, such as when Meta’s platform erased the Trump assassination attempt from history. Occasionally they are outrageous falsifications, such as when Microsoft’s Copilot professed to be a demigod and demanded worship from us pitiful humans. Or they can be smaller, subtler untruths, such as inventing academic citations or faking names, dates and other events.
What AI hallucinations mean for business
These examples could be seen as amusing at best and mildly irritating at worst. However, when critical customer-facing applications depend on generative AI, hallucinations can create undesirable business outcomes. For example, in February, an Air Canada chatbot offered a discount to a customer based on a false interpretation of the company’s policies, leading to an unwarranted refund which the airline was forced to pay.
Academics and GenAI providers are working to eradicate the prevalence of these hallucinations. Despite making some progress, AI is still hallucinating. This is preventing business adoption with 60% of decision-makers saying these hallucinations are their biggest concern with the technology, according to a KPMG survey.
Hallucinations may be an unfortunate reality of generative AI but organisations can mitigate their impact. Businesses should ensure employees are using the right models for the job and refrain from using consumer-facing AI tools such as ChatGPT for more sensitive matters.
Leaders should also insist on regular data audits, continuous training and user feedback to establish whether the use of GenAI within the company is effective.
By using internal platforms designed for a specific use case – such as combing through client record databases or surfacing examples of case law for legal departments – businesses might suffer fewer AI hallucinations and save time by automating laborious work.
However, working with generative AI requires sharp critical-thinking skills. Only subject matter experts will be able to detect a blatant error in an otherwise convincing AI-generated screed. Despite the hype, the big tech providers often mention the need for human oversight when using AI and, with the current state of play, it’s hard to conclude otherwise.
