• by David DiCamillo , Anil Hari , February 13, 2024

Implementing AI is sort of like welcoming a puppy into your house, a puppy that never grows up and needs constant oversight (not to mention potty training). Your AI tools need constant maintenance and oversight to perform as expected. 

There’s a good reason for this. When you ask Bard, OpenAI or Anthropic chat questions, the tool tends just to take whatever data it can justify as a credible source within the large language model (LLM), run it through a GPT and output something humans recognize as “a real answer.” Except, the AI often just returns lies. 

Here’s the thing: the AI knows they are lies. Wait, what?

Even when questioned, AI will admit it. ChatGPT features a disclaimer placed right below the core product’s prompt: “ChatGPT can make mistakes. Consider checking important information.” This disclaimer is baked right into the product’s DNA (which is the case for all chat-based GenAI tools).

We have said many times to our partners: “There is no bigger liar than an LLM.” 

So what is happening? Why are these lies (AKA: hallucinations) occurring? 

Simply put: if the AI’s “brain” is just the data it’s programmed with, then anomalies in the data can result in “hallucination.” 

In more technical terms, hallucinations represent deviations from normal operations in AI systems, manifesting as unusual patterns, behaviors or the generation of fictitious outputs. Anomalies could present as erroneous output, compromised decision-making, or data-processing irregularities, hinting at issues with the AI's algorithms or the data it uses. Such anomalies arise from myriad sources, including corrupted input data, model drift, or unintended byproducts of the AI's learning process.

So what to do? When dealing with anomalies or hallucinations, there needs to be a holistic, specialized strategy. It starts with an iterative process of investigation and analysis to find the root cause. This could necessitate a model retraining or a more comprehensive system revision. User feedback isn’t just helpful — it’s the pivotal tactic for detecting both anomalies and hallucinations, often serving as the initial alert to a potential issue. 

As we continue exploring various LLMs and GPTs, quality assurance engineers must prompt their way through tests, meticulously ensuring the AI’s “sobriety.” 

Whether a hyper-growth startup or a Fortune 500, a multi-layered approach to detecting and resolving AI hallucinations long term is needed. Here’s what needs to happen: 

Detect AI anomalies by consistently:

• Monitoring system outputs 

• Analyzing performance metrics 

• Performing data quality checks 

• Reporting errors and feedback loops 

• Performing adversarial resting

• Validating and testing models

• Logging and maintaining audit trails

Train diverse data sets … constantly

Regularly retraining the model with updated, diverse data sets or modifying generation algorithms helps prevent the creation of misleading information. This will help ensure the AI operates reliably and maintains its intended functionality.

Recognize that AI is a marathon, not a sprint; it will create jobs in the long run

As headlines and fake experts spell the doom of the countless jobs AI will supposedly eliminate, we want to talk about the jobs it’s creating.

• AI analysts: As AI technologies evolve and automate various tasks, there's a growing emphasis on skills where human capabilities are still essential. These skills often involve creativity, emotional intelligence, complex problem-solving, ethical judgment, decision-making and interpersonal interactions.

• A whole new software industry: This has already begun. Just look at Arize AI, WhyLabs, Fiddler, Weights & Biases and Evidently AI among many others dedicated to analyzing data to understand how AI creates its outputs.

This is a brand-new world. Still, so many leaders are eager to tap into the power of AI, but trying to get involved feels daunting. Just remember, if you are ready to let your company fall in love with a new technology puppy, you need to learn how to train it. You’ve got to feed it better puppy food (data) and make sure your brand can handle the little “accidents” your AI is definitely going to leave behind.