Hallucinations are the Feature, Not the Bug

We’re coming up on the third year of LLM hype after ChatGPT 3.5 launched in November 2022. At the time, everyone was struggling to understand the technology and its implications for society in the present and the future.

The primary thing that I have learned since joining the team to integrate LLMs into Shaping Tomorrow: hallucinations are not bugs. They are, in fact, the main feature for LLMs. Clearly, this depends on what you want the LLM to do, but there are few things they do so far that a human being would not do much better, albeit with much more time. I’m not discounting the efficiency that LLMs offer, but where we should be steering the technology is in supporting humanity to think more creatively.

That’s right, I’m suggesting that the slop is actually the real value when used by reasonably intelligent humans who can think for themselves. It’s just a matter of critically recognizing the slop, thinking about its utility, and then creating something new.

I had been using a variety of automation tools in my futures work for almost 10 years by the time of 3.5’s launch. Then, I started experimenting with incorporating summarizers into our existing automation a few months before the launch, probably August 2022. I thought it was useful, but I wasn’t certain what our clients would think about it.

When 3.5 launched, I started scrambling to understand what value we could derive from the technology that would benefit our clients. Shaping Tomorrow has grown exponentially over the last 3 years as a result. Prospects have been curious about how we use the new technology, both enthusiastically and skeptically. Some are happy to see anyone who could help them use this newfangled AI, and others are concerned about the future implications to human dignity, employment, and the environment. It’s been great to have philosophical and future focused conversations in a professional setting where such topics are usually focused on pragmatic outcomes and/or the near-term.

In the process of testing different prompts to add value to our existing processes, we noticed that the LLM was conflating two ideas expressed in two different sentences into one wacky idea. It told us about a new technology in forestry that could immediately create credit cards from twigs. The machine titled this insight “Growing credit on trees”. I dug into the sources used for training the output, and I found two sentences, one about how greeting cards are made and another on using fallen twigs to make paper.

This hallucination was completely useless to me, but the idea of making paper from fallen twigs was something I hadn’t seen previously. This was my first clue that hallucinations could be of value even if the hallucination itself was mostly useless.

The more I used the system to create articles using our sources, the more I began to recognize the hallucinations for what they were, confabulations of unrelated ideas. And, I started running into more confabulations that got me thinking about future scenarios. These confabulations were, sometimes, suggesting where trends might converge in the future, and convergence is a big part of futures and foresight.

A prime example is the convergence of mobile computing and gaming when the iPhone was released. Nokia had some ideas about how mobile computing and gaming could mix, but they focused on making gaming mobile. Apple, on the other hand, considered how the strengths and weaknesses of both gaming and mobile computing might create a new way of gaming using mobile phones. Mobile gaming was one of the killer apps for the iPhone.

So, I came across several hallucinations that seemed to indicate what a convergence of trends might look like. For instance, I ran a workshop for a human resources (HR) client. I ran a list of trends relevant for HR, and the very first one was Biometric Menstrual Tracking.

The participants went a little quiet. One of the ladies wondered what business that would be for HR. I said I didn’t know, and I also commented that my mother wouldn’t want me even speaking about the topic. That got a slight laugh, and the two men loosened up enough to ask the ladies about whether they were doing enough to support their female employees. Then, the discussion started to take shape as they discussed the importance of removing the common stigmas around women’s health.

As they did that, I looked at the sources for the list of trends. I found nothing relevant. So, I searched our wider system and found one story about Spain’s then recent legislation allowing women to take a small number of days off work for extreme cases related to their health.

I brought this to the group, and one of the ladies told us about her experience in Korea where a similar policy is observed. However, the group was still confused about how this specific type of biometric tracking might relate to HR. The general consensus was that it was creepy.

Now, was this a hallucination? Did the LLM confabulate this source with a more specific HR source about biometric tracking? I couldn’t figure it out even after speaking to the IT team.

Regardless, the machine brought a trend that I would never have brought to the group because of both my gender and the stigma my upbringing associated with the topic. Plus, it ignited a discussion that brought value to the client. We criticized the output. We used liminal thinking to consider its value anyway. Then, the team began to get creative about how they wanted to better address women’s health in their (already progressive) HR policies.

Clearly, my examples are related specifically to futures and foresight. However, LLMs are still hallucinating with no sign of ceasing even as they improve. So, we need to figure out what to do with them. My argument is that these hallucinations which can seem inconvenient for tasks that require accuracy such as customer service are actually the primary value drivers for tasks that require any forms of creativity such as strategy, innovation, workshop facilitation, and marketing. The only way to make hallucinations valuable is for humans to engage their critical, liminal, and creative thinking.