Why does AI get things wrong?
Artificial intelligence can feel like it gets things wrong – a lot.
As a coach, consultant, or small business owner, you may have had false or factually incorrect outputs from AI tools like ChatGPT. And inaccurate responses can take time to fix – which is frustrating if you’re using these tools to save time.
However, this is not a reason not to use AI – often, it’s a case of rubbish in, rubbish out; and besides, I believe we’re setting an unfairly high bar for AI to meet.
Let’s reframe the problem: can we help AI to get things right more often?
In this blog post, I’ll explain the three common reasons AI can produce errors, and why these kinds of output can seem more frequent with AI tools. I’ll also offer practical tips on how to help AI make less mistakes.
How does AI get things wrong
1. Hallucinations: When AI makes stuff up
One of the most intriguing (and frustrating) ways AI can go wrong is what’s known as a “hallucination”, or more accurately, a “confabulation”.
This is when AI generates information that sounds plausible but is completely fabricated. For example, you might ask ChatGPT for insights on a specialist topic, and it confidently provides an answer that, when you check it, is factually incorrect or cites names and places that don’t exist.
Why does this happen? AI models like ChatGPT are trained on vast amounts of data, but they don’t ‘understand’ this information in the same way humans do. Instead, they predict likely word patterns based on probabilities. So, if it can’t find a concrete answer, it may piece together something that looks right, even if it’s not. This is problematic when you need reliable data for business decisions.
2. Fake citations: When AI links to nowhere
If you’ve ever asked ChatGPT or a similar model for references or sources, you may have been given a link or citation that doesn’t actually exist. This is because AI (mainly) doesn’t have real-time access to the web or a specific set of sources – instead it generates responses based on patterns in the data it was trained on.
Which is not to say AI is ‘lying’ to you; it’s merely trying to satisfy your request in a way that fits the pattern of a legitimate citation. The problem arises when you rely on these citations to back up an important point or decision.
3. Reward hacking: When AI optimises for the wrong thing
AI models are designed and trained to optimise their outputs based on specific goals, such as providing an answer or maximising user engagement. However, unless these goals are aligned with yours, you can end up with AI material that satisfies the AI’s built-in parameters rather your business objectives.
For example, you want meaningful information that connects with your audience, but an AI system might generate clickbait content for you because clickbait gets the highest engagement. Another example would be a social media AI algorithm that curates your feed based on its own drivers (engagement!) and sometimes shows irrelevant or unhelpful content because it know you will scroll past it, thus spending longer in the app.
As a recent article in The Verge says,
“Hallucinations occur when an AI unintentionally generates incorrect information, often due to knowledge gaps or flawed reasoning. In contrast, reward hacking happens when the … model strategically provides incorrect information to maximize the outcomes it was trained to prioritize.”
Why do AI mistakes feel more significant than they are?
When we interact with AI tools like ChatGPT, we tend to expect them to get it right first time. If and when they fail, it’s frustrating, and we may question the tool’s reliability. This contrasts sharply with how we engage with other instances of AI. On social media, for instance, if an AI algorithm serves up a post we don’t like, we scroll past it. We don’t dwell on the mistake – we move on. Similarly, with Google search results, if one link doesn’t answer our question, we click on the next one.
The difference is that when we ask AI a specific question, and it provides a single answer, we treat it as a final output. We’re applying a higher standard to this kind of AI use, and that leads us to amplify any errors. It’s important to remember that AI is not infallible, and like any other tool, getting successful results depend on how we use it – and that sometimes requires a bit of trial and error.
How to avoid common AI mistakes
Here are a few practical steps you can take to get the most out of AI tools:
- Use AI as a support tool: Treat AI as a helpful assistant, not a final authority.
- Be crystalclear with your prompts: To avoid ‘reward hacking’, clearly define what you want from your AI tools. Be specific with your prompts or instructions. The clearer your input, the more likely you’ll get meaningful, relevant and accurate output. (I have a whole section in my AI video course – see below – about how to prompt more effectively.)
- Look beyond one answer: If you don’t get the correct answer the first time, rephrase your question or try a different approach.
- Double-check the facts: Always verify any critical information included in an AI response, especially if it affects your business.
By understanding the limitations of AI and taking a few simple precautions, you can avoid errors and inaccuracies and get the most out of these powerful tools.
If you want to learn more about how to get more from AI, I’ve created a video course to help you get started and build your AI skills. The course is created specifically for coaches, consultants, and small business owners and is focused on how AI can make marketing easier and more effective. There are 15 short step-by-step guides on how to use the many different aspects of AI effectively for your business.
