The rise of AI has created “AI hallucinations” (Naveen Rao, VP AI, Databricks) AI has seen huge leaps and rapid expansion over the past few years – and if the experts are to be believed, this trend will continue throughout the new decade.
According to Rao, even as large language models (LLMs) have transformed natural language reason and become powerful productivity tools, their potential to produce misinformation and disinformation is a challenge that companies will have to grapple with.
AI hallucinations are events when AI model outputs are plausible sounding but are false, nonsense, or constructed. While seemingly innocuous in lower-stakes consumer applications, this is a much more serious problem in the enterprise, where it can potentially result in ill-informed decisions, reputational harm, and possibly legal liabilities.
The point is that while reasoned (but mistaken or reckless) human errors can be explainable by way of intent, the AI hallucinations under discussion are most fundamentally “process glitches” about the workings of the model.
Databricks, the [construction, hyperparameter tuning and deployment of their models boosts the popularity of Databricks’ collaborative machine learning platform, not to mention that the robust model lifecycle management can enhance Databricks’ model reliability services. Rao highlights the strides the company has made towards adding strong governance layers to generative AI that control how AI agents use certain access rights and entitlements. He also calls for a modular approach to creating AI applications, leaving behind monolithic LLMs and toward systems that have interconnected, independently certifiable parts. This enables better details and control of outputs.
Industry analysts and companies including Databricks are probing potential remedies for hallucination. These techniques include obtaining higher-quality, more diverse training data sources, using advanced techniques for prompt engineering to provide better guidance for AI models, and using human-in-the-loop (HITL) methods for critical applications.
Relational retrieval knowledge bases are a common method for grounding large language model (LLM) answers to knowledge spatial databases, and are considered by some as a powerful way to ensure that LLM generations are anchored in real enterprise-truth and not make-up information.
Whilst Rao admits there’s been great strides made in AI, he says we should not be surprised if AGI is not as imminent as some might believe: “It’s a much more difficult problem than people realize.” For businesses, the emphasis is squarely on creating trustworthy, accurate AI systems that provide a clear ROI, and controlling AI hallucinations is a key factor in achieving this. The challenge remains, but so does the commitment to the idea of building AI that is more reliable and trustworthy in the real world.
