Artificial Intelligence Is Amazing, But Please Don’t Use It For These Common Tasks

This article written by Reken co-founder and CEO Shuman Ghosemajumder was originally published on Inc.com.

I once asked an early version of a modern AI chatbot to generate a profile of me. Asking about a subject you know every detail of is a great way to test AI models, since you can spot hallucinations more easily.

The profile it generated was comprehensive, and it all seemed accurate until I spotted an interesting line: it said I was the author of a book called Click Here to Kill Me. From the description, it was about “how to protect yourself from online scams and fraud,” which was consistent with my areas of professional expertise.

But knowing that I had never authored such a book, and being amused by the title, I was curious. I Googled it and confirmed that in fact no such book existed. (I later realized that the AI may have been thinking of the real book Click Here to Kill Everybody by Bruce Schneier.)

I asked the chatbot for more details. It happily told me all about this fictitious book: its length, date of publication, and other details. Only after additional interrogation did its ruse collapse. The chatbot admitted that there was no such book and said it had “made it up.”

This story illustrates a truth about modern AI: It’s not what it seems to be. That doesn’t mean it’s not a huge breakthrough, but it does mean that the key to maximizing its benefit and minimizing its harm is to understand what it's good for and what it’s not.

We, as a society, have talked about artificial intelligence since the 1950s. The holy grail of AI has always been systems that could match or exceed the capability of the human brain. Until just a few years ago, people could clearly see we were far from that point, as software could not understand the nuances of human language or communicate in ways that demonstrated anything resembling real intelligence.

All of this changed with the advent of generative AI, and specifically chatbots based on large language models (LLMs) trained on vast amounts of written text. Suddenly, computers seemed like they could comprehend what we were saying, and write (or speak) to us articulately and confidently, leveraging the vast knowledge of the world’s information that the models had been trained on. Users would regularly find mistakes in their answers, but were told that those were “hallucinations,” a term that sounded like a special and rare case of error that might soon be overcome.

Actually, LLM-based chatbots don’t understand at all. Instead, they statistically analyze the world’s writing to predict and construct responses to questions. It turns out that these answers are often accurate enough to be useful. However, just as often, AI answers are riddled with errors or nonsense and can even be completely fictitious.

The difficulty is telling the good answers from the bad ones. If you’re an expert on a particular subject—such as your own personal background—then you can easily spot mistakes. But most people are not AI researchers testing chatbots, so the vast majority of questions posed to them are in areas where the user is not an expert, and is therefore unlikely to easily identify errors. And so, the AI seems not just intelligent, but all-knowing, with an imposing command of every subject outside of the user’s area of expertise. In other words, the AI we have today is perfectly designed to fool people into thinking it’s the superhuman intelligence we have been promised by science fiction.

It doesn’t actually deliver this, of course. But that doesn’t mean it’s not incredibly useful. Perhaps the most important rule for using chatbots is this: Do not rely on them to be accurate. This is why OpenAI’s ChatGPT has the disclaimer, “ChatGPT is AI and can make mistakes. Check important info.” Google’s Gemini has a similar warning. Both of them state this in small print that most people don’t even notice, and if they do, they may not heed the advice.

This isn’t a problem for all tasks one might use AI for. In creative work, there is often no right answer, and so the output from AI may be eminently usable. However, in terms of getting factual answers to questions, you should take everything an AI system gives you with a grain of salt—and then verify it carefully before using it anywhere.

This means that there are many things that generative AI today should simply not be used for. Here’s a brief overview of topics to be extra wary of.

Don’t Use AI As A Calculator

Many people regularly use chatbots as calculators or to help them with financial analysis. We have relied on dedicated calculators and spreadsheets for decades, so it’s only natural to think that AI would be an improvement on those older technologies—and when people try it, it often seems to work impressively well. ChatGPT and other chatbots have been benchmarked as performing similarly to humans on standardized tests and solving competitive math problems.

However, chatbots regularly fail on simple arithmetic or basic counting questions. (They’re also unreliable at accounting and sports statistics.) The problem of chatbots misstating the number of r’s in the word strawberry, for instance, became a famous meme in the AI community.

When you ask a chatbot to perform a math operation or solve a problem, it will give you a detailed response. It will do so even if the response is nonsense, so it’s up to you to verify whether there is any faulty logic in its steps. The reason this happens is that it’s not actually solving the problem. Instead, it has been trained on the text of the solutions of similar problems, and it statistically constructs what an answer to this particular problem might look like. Given lots of data, this can appear convincing. And it can often be helpful. But it’s completely unreliable. This is why one notable mathematician recently declared that the value of LLMs for solving advanced math problems was “basically zero.”

To verify whether a chatbot’s answer is correct, you often need to do exactly the math that you would do if you didn’t have AI in the first place. So if you need a reliable answer to a math problem, AI is not an effective shortcut to doing the calculations with non-AI tools.

Don’t Rely On AI To Summarize Documents Accurately

Many people use AI to summarize stuff—documents, videos, emails, articles, and more. But when AI hallucinates in those contexts, it sometimes introduces content that wasn’t in the document at hand, or it may twist the meaning or omit important information.

The level of potential harm varies depending on how the information is used. If AI summarizing an email makes a mistake that it presents just to you, that is likely nowhere near as harmful as AI summarizing a news story and presenting wrong information to millions of people online. Last year, Apple was forced to pull its AI-generated news notification summaries from its beta version of Apple Intelligence because it generated demonstrably false information, although Apple later brought the feature back.

Don’t Use AI To Create Maps And Diagrams

Image generation is a popular AI application, but avoid relying on a map or diagram that’s generated by an AI model. A detailed map is a great example of content that looks more accurate the less familiar you are with it. If I looked at a map of downtown Palo Alto, I would be able to identify even small mistakes, but if you showed me a map of Anchorage that was 99 percent made up, I could be easily fooled.

The same applies to technical diagrams. This is why the folks on the plane in that episode of Friends believed Phoebe when she said there was something wrong with the “phalange”—they didn’t know enough about airplanes to know there was no such part. AI regularly makes up stuff, as Phoebe did, but the bigger problem is that it doesn’t know it’s giving you a falsehood. In that sense, it’s more like the famous quote from George Costanza on Seinfeld: “It’s not a lie if you believe it.” AI regularly believes its own mistakes.

So, How Should You Use AI?

Just because you can’t rely on AI systems to produce accurate results in these areas doesn’t mean that you can’t use them to help you. In the case of calculations, there may be a problem that you don’t know how to approach, so you can ask a chatbot to give you an idea of how to solve it. When summarizing a document, you can take what it gives you as a starting point for structuring your own reading and research. For maps and diagrams, generating images with AI can often give you a sense of what you’re visually looking for; it can pick various aesthetic choices that help you decide what you want or could create.

Despite the very real deficits, AI is still a great brainstorming partner. It’s excellent at filling in gaps in your thinking, giving you ideas you might not have considered, and helping you see things in new ways. All of that is incredibly useful in terms of helping you solve your own problems faster, smarter, and more accurately—even if the AI can’t do it all on its own.