AI Will Fail. That’s A Good Thing.

We’re in the early days of generative AI. And as many of us experiment with the technology, we’re finding useful applications.

But, AI is evolving towards a future state that promises to be more useful than it is right now. And in order to get there, one thing needs to happen, over and over again.

AI needs to fail.

In fact, embracing and learning from failure is essential to unlock the full potential of generative AI. Because I believe that AI will become an extraordinarily capable creative collaborator.

Let’s break down what I mean by that. What does a great creative collaborator do? To me, a collaborator supports our weaknesses. When we fail in our creative process, a collaborator knows how to push us forward. And, similarly, when our collaborator fails, we know how to push them as well.

That’s what true collaboration is. We complement each other’s skills and fortify each other’s weaknesses for mutual growth. We work together towards a better outcome than either could create alone.

In order to reach this future state for generative AI, we need to understand exactly what the failure states are on both sides of the collaboration. Where does AI fail in the creative process? Where do we fail? And how can we mitigate, move through, and learn from those failure states to continue to improve?

When it comes to AI, there are three failure states that stand out to me. And each of them has a potential mitigation.

The first failure state was termed the “10,000 Bowls of Oatmeal Problem” by Dr. Kate Compton. It illustrates the tendency of generative systems to output lots of similar-seeming artifacts. If we imagine AI generating 10,000 bowls of oatmeal, we know that each bowl is, technically, slightly different than the other—but not to the end user or audience. To them, it’s just a lot of unremarkable oatmeal. My long-time collaborator Dr. Piotr Mirowski calls this failure state the automation of mediocrity. And Will Storr, in his book “The Science of Storytelling” reminds us that our brains are wired to understand the world through compelling stories. That means that AI fails both as content and as effective communications when it generates a thin gruel of middle-of-the-road content.

We can mitigate this failure state by prompting AI and pushing it in the right direction. As we get better at prompting, and AI gets better at understanding our intentions, we can push it to create more and more differentiated content. Our creativity is key to unlocking diverse regions of interesting outputs that populate the landscape hidden within the model.

The next failure state for AI is something I call the Tinder Effect. AI can generate many outputs very quickly, with very little effort—a simple click of a button. This creates an expectation that the next click will yield better content, leading to a failure state characterized by lost creative momentum and continuous, unproductive searching for the next option. The Tinder Effect is related conceptually to the infinite scrolling design pattern which also problematically lacks stopping cues.

We can mitigate the Tinder Effect by disrupting the rapid content creation loop. Rather than acting as an audience that passively receives content from AI, we need to position ourselves as curators. What I mean is that we need to give our input, reviews, critiques, guidance, or commentary on the model’s outputs. Decide which is best, or push the model in a new direction. By taking an active role in content creation, we can make choices that push the model into better directions and maintain a creative connection in the collaboration.

The third failure state for AI is something I call the Me Me Me Effect. Essentially, we sometimes want AI to create content the way that we’d create it. Here’s the thing: generative models don’t necessarily create content the same way that we do. Try as we might with prompting or fine-tuning, we might not be able to truly match our own voice. That might be because they’re trained in a way that doesn’t match our creative instincts and tastes. But it also might be because the model is pulling us outside of our comfort zone in a way we didn’t expect.

This one is harder to mitigate. Lots of research is ongoing on how to best prompt, tune, guide, and encourage models to align with our preferences. One thing we can do in the meantime is to approach creative AI with a collaborative mindset—learn to expect that AI will push you in new and complementary directions and get results that you value. These results may not be the outputs you will eventually use, but they may provide the creative spark or association that prompts you to explore new and interesting directions.

Failure states exist on both sides of this creative relationship. We humans fail to complete the creative process all the time. We get stuck looking at the blank page and end up never starting. Sometimes, we feel like an empty bucket–we’ve run out of inspiration and can’t think up any new ideas. And sometimes, the timer runs out when we need to come up with a great idea in a short time frame and can’t get there soon enough.

Like a great creative partner, AI can help us to overcome these failure states. With AI, we never have to look at a blank page. We can reignite our ideation by generating one hundred new ideas with the click of a button. And we can save time by arriving at creative solutions faster.

True collaboration means each partner brings something different to the table. You accentuate each others’ strengths. And you mitigate each other’s failure states.

We have to learn to embrace that relationship with AI. Because if you can approach AI with a collaborative mindset, you can push yourself to be more creative than ever.

So the next time you sit down to create, invite AI to collaborate. Just remember—like any collaborator—AI will let you down, once in a while.

That’s part of the process.