AI Hallucinations and Generative AI: A Mirror of Perception

There are few instances where AI and the human psyche converge in shared experiences. One such fascinating example is hallucinations — a phenomenon experienced by both humans and, as we will come to understand, generative artificial intelligence (AI). AI can hallucinate, albeit in a way that differs from human hallucinations.

Hallucinations in Artificial Intelligence: When AI Starts 'Seeing' Things

Hallucinations, real experiences without any basis in reality, can affect all of our senses and vary in complexity. They arise in us when our senses are disrupted and attempt to fill in the gaps of our understanding of reality, sometimes with entirely unrealistic results. This is an affliction that can also affect generative AI.

Generative AI can 'hallucinate' and create data based on non-existent patterns, much like a mirrored reflection of our own brain through AI's artificial neural networks. An AI system 'understands' something based on what it has been trained to understand. This is the crucial difference between living beings and AI systems.

(Midjourney)

What Contributes to AI Hallucinations?

Several factors can contribute to AI hallucinations. The primary factors are:

• Insufficient or biased training data: If the AI's training data is limited, unrepresentative, or skewed, the AI may operate based on this incomplete picture, leading to incorrect or unexpected results.

• Task complexity: Certain tasks involve a high degree of ambiguity or subtlety that AI may struggle to handle or lack training data for, causing it to generate unusual results.

• Inadequate model architecture or training: Sometimes the design of the AI model or the way it is trained can also contribute to hallucinations, where it fabricates a result with little grounding in reality.

A common example of AI hallucinations today primarily revolves around text, when an AI receives a specific question where it is asked to find differences where no training data existed for the model — for example, a question about whether a kilogram of water or a kilogram of oxygen weighs more:

(Nick Babich)

The truly concerning problem is the subtle, hidden, and less obvious hallucinations that could just as well align with reality and that actors then act upon. A possible worst-case scenario could be in healthcare diagnostics, where the model has been trained on flawed data and completely fails to identify diseases.

Can We 'Cure' AI Hallucinations?

Over the past six months, actors such as OpenAI with ChatGPT have strengthened their model to minimize the risks of hallucinations, but the problem persists. While it may not be possible to completely eliminate hallucinations in generative AI given the current state of technology, there are strategies to reduce their frequency and impact:

• Improve data quality: Ensuring the AI has a large, varied, and representative training dataset can help it better understand the patterns it is meant to learn.

• Improve model architecture and training: Adjusting the design of the AI model or fine-tuning the training process can help minimize hallucinations.

• Use evaluative algorithms: Implementing algorithms that evaluate and filter the AI's outputs can help catch and correct hallucinations before the content is finalized.