“I’ve heard a lot about AI hallucinations, I’m wondering what they are actually…”
Hi there, and welcome back to another week of Ask Trish! I hope you’re all well and having a wonderful start to June. (For those of you in the Northern Hemisphere, I hope you’re enjoying some sunshine and warmer temperatures! For those of you in the Southern Hemisphere, I hope you’re enjoying some cooler weather.)
Thank you so much to this week’s question-er for the fantastic question! Indeed, “AI hallucinations” have been all over the news lately…but I’m not sure that they have been explained clearly, particularly for a non-expert audience. It’s also just such a futuristic, counterintuitive phase (how could AI hallucinate?)…and thus deserves some explanation. Thankfully, in this week’s post, that’s exactly what I’ll provide to y’all! I’ll tell you a little about what AI hallucinations are, why they occur, and how you can protect yourself against them. Sound like a plan?
Let’s get into it:
First and foremost, what are AI hallucinations? Put simply, AI hallucinations are inaccurate, misleading responses that AI models, e.g., Generative AI models like ChatGPT, generate. When you input a prompt into ChatGPT and get a response back, you might assume that that response is 100% right…but in fact, ChatGPT may (and regularly does) confidently spit out answers that are completely wrong. But why? What’s going on? Well, at a technical level, the model may be perceiving patterns or making connections that are in fact non-existent, thus generating outputs that aren’t quite right or are completely nonsensical. (Remember: put very simply, AI models detect patterns in training data and make predictions based on what they’ve learned. But if the pattern doesn’t exist/isn’t a thing, unfortunately, the prediction will be wrong, too…) That’s why we call the output a hallucination! It’s a lot like when humans see something that isn’t really there. As Generative AI models have been released, there have been plenty of examples of AI hallucinations…Google Bard’s chatbot has inaccurately claimed that the James Webb Space Telescope had captured the world’s first images of a planet outside our solar system (NOT true). And Microsoft’s AI chatbot Sydney claimed to have fallen in love with users and to be spying on Microsoft employees (yikes).
Okay, now we understand what AI hallucinations are. But why would an AI model perceive a pattern or make a connection that doesn’t exist? How does that happen? Well, for one, it may be because the model’s training data is insufficient. For instance, imagine that an AI model is trained on images of Krispy Kreme doughnuts to automatically identify which have been “poorly baked” (is that even possible?!). But if the training images do not include any images of doughnuts that have been “correctly baked,” the AI model may inaccurately predict that a doughnut that was, in fact, baked correctly, has been “poorly baked.” The model doesn’t have anything to go on…so it hallucinates. Another reason an AI model may pick up on connections that don’t actually exist in the real world is biased data. In this case, it’s no longer that the training data isn’t there, it’s that the training data points to connections that don’t reflect our world’s “ground truth”/values. The AI model doesn’t know that…so it faithfully reflects the patterns in the data, even if they aren’t patterns that make sense/seem right to us. The takeaway here is clear: if, as a developer, you’re looking to minimize AI hallucinations, you ought to start by looking at your data.
But what if you’re not a developer? What if you’re just another GenAI user, wanting to avoid hallucinations? Can you do anything? Yes. First and foremost, I would suggest limiting the scope of the prompt that you input into a GenAI model. The more specific you are, the less room for misinterpretation and the fewer possible outcomes. I would also avoid using any hard words or phrases/metaphors/jargon. Doing so can make a model think you’re actually asking about something completely different…and lead it to try to make connections where there are none. Talk to your GenAI model the same way that you’d talk to a child. Keep it simple and clear. And do definitely provide as much relevant information as you can when you prompt a GenAI model to do something for you. Don’t say, “Krispy Kreme,” say, “Krispy Kreme, a longtime doughnut chain.” Again, more information makes it harder to misinterpret. Finally, one last suggestion: always, always, always double-check any descriptive information GenAI models give you. In general, I try to avoid using GenAI models for descriptive information; instead, I rely on models like ChatGPT to help me brainstorm/to help structure my ideas. (See my tips here!) But if you do decide to get the facts from ChatGPT, verify everything it tells you. Be critical!
Hopefully, that gave you a helpful, clear introduction to AI hallucinations. No doubt, this issue will evolve/continue to garner a lot of attention, so keep your eyes peeled…and do let me know if you have any other AI hallucination-related questions on your mind! Or, really, any questions at all in your mind. That’s right: whatever you’re wondering about, I’d love to hear from you. Please go ahead and share your thoughts here. Thank you in advance for contributing!
Have a great week,
Trish