ChatGPT isn't Google

After posting that ChatGPT isn't Google, Izzy Piyale-Sheard asked a thoughtful followup question:

I appreciate Izzy's question and realized that it needs more than a quick response. So here's what I told Izzy, and others who may ask why and how ChatGPT isn't a search engine.

I'm happy to elaborate. I'll talk about ChatGPT and Google, but this applies to other LLM chatbots and search engines too.

The difference lies in what each tool is and how it works. ChatGPT accesses an LLM that was trained on a vast amount of data, including information from the internet. But when you interact with ChatGPT, you're not searching the internet itself. You're not searching its knowledge base. You're not searching at all! You're asking ChatGPT to create something new. The whole point of ChatGPT is that it's generative - it is designed to create something new (such as text, code, or an image).

ChatGPT works by finding patterns in the data it was trained on (the patterns in language, images, code, and so on). It uses prediction and probability modelling to generate a response that best matches what it believes you're looking for, based on what it knows about patterns and how humans speak. This is why it sometimes gets things wrong (aka a "hallucination"). ChatGPT strives for accuracy, but its goal is more in the accuracy of language, not necessarily the accuracy of facts. Hallucinations also happen because ChatGPT may not have enough deep information about a specific topic, or up-to-date data.

Google works by crawling the internet, indexing its contents, and ranking each webpage based on the quality and relevance of the information provided (and a few other factors, but let’s keep this high level). When you do a Google search, you are searching Google's index of the internet. Your search engine results page (SERP) lists the webpages that Google feels best match the intent of your search query.

Google’s SERP pages are a hot mess these days. They’re cluttered with Google pushing their own stuff, such as results from Google Maps, Google News, Google Images, YouTube, Google Shopping, and Google Business Pages. Then there are ads, and Google’s “knowledge panels”. It’s harder than ever to sort through that mess and find what you want in the organic search results.

When you do eventually find a result you want in Google, remember that it is something that already exists. We can debate whether or not Google is as effective now as it was in the past, but it is fundamentally still a search and retrieval system. ChatGPT is a generative system.

These differences between Google and ChatGPT are more prominent when searching for specific information. I recently saw someone post on LinkedIn that ChatGPT must be broken because it couldn’t accurately say how many employees a company has, even though that information is publicly available on the company’s website. I understand why this person thought ChatGPT should know that information. It has “read” the internet, right?

But the reason why ChatGPT couldn’t accurately answer the question is based on this core difference. ChatGPT isn’t looking to search and retrieve information for you. It’s looking to create new information based on patterns in language. ChatGPT is a mimic. It’s trying to behave like a human, and when it doesn’t know the answer, it will make one up. It can be helpful to think of ChatGPT as a child in this way. Children sometimes don’t understand what each word in a sentence means, but they string them together because they think that’s the order they should go in.

The example questions you provided (”What are some of the best remedies to cure a sore throat?” and “What are the quickest easiest ways to remove a blood stain from clothing?”) are fantastic questions for ChatGPT. They’re fairly broad and general queries, and ChatGPT can answer these with a high level of accuracy. These questions don’t require the specificity of Google’s search-and-retrieve system.

The more I play around with ChatGPT, the more I’m reminded that there is seldom a single right answer to most questions. Humans are good at asking questions and selecting an answer by taking in a large amount of information and weighing its importance, accuracy, and relevance to what we’re looking for. Artificial Intelligence systems are trying to do that too. They’re just still a ways behind us.