Avoid Getting Spooked by AI Hallucinations

In this season of ghost stories and eerie decorations, a new digital specter has emerged: AI hallucinations. As the use of AI continues to rise, we are confronted with a chilling reality that has been haunting the digital realm.

AI hallucination describes the situation when AI technology produces information or responses that lack grounding in actual facts or established knowledge. What is particularly striking is that AI tends to be very confident in the accuracy of these made-up answers.

With ChatGPT, Bing’s Chatbot and Google’s BARD all have been the creator of some public examples which have surfaced:

– In February 2023, during its initial public demonstration, Google’s chatbot, Bard, incorrectly claimed that the James Webb Space Telescope had captured the first image of a planet beyond our solar system [The Verge].

– In June 2023, lawyers in New York State were fined $5,000 for submitting ChatGPT-generated hallucinations in court. The two lawyers submitted fake legal research created by ChatGPT, citing non-existent cases [Yahoo Finance].

– A rather disconcerting conversation with Bing’s Chatbot ended with the Chatbot declaring its love for a New York Times reporter [The New York Times].

There are many reasons for why AI results may contain hallucinations, some identified causes include:

• Insufficient or Biased Training Data: AI models rely on extensive and unbiased training data to understand language and provide coherent responses. When the data is limited or contains biases, these models may struggle to grasp language complexity, leading to hallucinatory or biased outputs.
• Unfamiliar Idioms and Slang: The use of idioms or slang expressions in prompts can confuse AI models, especially if they have not been exposed to these language nuances during training.
• Input Data Noise: In cases where language models receive noisy or incomplete input data, such as missing information, contradictory statements, or ambiguous contexts, they may falter in providing accurate predictions. This uncertainty in input can lead to hallucinatory outputs.
• Adversarial Attacks: Deliberately crafted prompts designed to confuse AI models can generate unintended and often absurd responses. These adversarial attacks exploit the limitations and vulnerabilities of AI systems.

Despite its current shortcomings, steps can be taken to minimize inaccurate responses:

1. Clear Prompts: Use straightforward and easy-to-understand prompts to reduce the chances of AI misinterpretation.
2. Provide Context: Including context in your prompts helps the AI generate more accurate responses by narrowing down possibilities.
3. Specify Response Type: Limit potential outcomes by specifying the type of response you want from the AI.
4. Express Preferences: Communicate your desires to the AI and inform it of any content you wish to avoid, guiding its responses effectively.
5. Verify Results: While refining your prompts is essential, it is crucial to verify and double-check every output the AI provides, whether you’re using it for coding, problem-solving, or research.

AI doesn’t have to be scary, as we all learn to become experts in prompt engineering, let’s continue sharing what we learn and how we can best approach these new tools to enhance our everyday lives.

When implementing AI strategies within your organization, make sure to refer to this FORUM article on where to begin and this FORUM article regarding the questions your association should be asking before creating policies regarding its use.

*Note: The feature image on this article was created using Canva’s AI text to image tool.