What are chatbots?

You’ve heard the hype about chatbots, natural language processing (NLP), and generative AI. You’re ready to give it a try – but where to start? What is a chatbot, anyway, and how do you know which one to use?

You can think of a chatbot as a personal assistant in the virtual world. It can fetch information for you or help you create things like documents or images. It uses natural language processing to communicate in a human-like way. The first chatbot that captured the public imagination was ChatGPT, but today there are many others. Each has its own strengths and weaknesses, and knowing which tool to use for which purpose will help you get the most out of your interactions with them.

Before we jump in,  let’s spend a moment talking about the caveats. As with any tool, chatbots have limitations. They are trained on a specific dataset – in many cases consisting of trillions of inputs, such as all the available web pages in English or all the scientific and medical literature – and these training sets all have boundaries. For example, the first publicly available iteration of ChatGPT was trained on data from 2021 and earlier, making it useful for asking questions about historical or general information but not about contemporary questions. (Later iterations have filled that gap, so this is no longer an issue.) 

Like any computer interface, chatbots do exactly what you ask them to do. Getting the outcome you want requires knowing how to ask the chatbot for the information. This is called prompt engineering. There are many approaches to prompt engineering and, in fact, you can find courses and resources to help you learn this skill. It is often helpful to be as specific as possible in your request, and you can always ask follow-up questions or hone your request over several iterations.

Chatbots are also notorious for sometimes making things up, or hallucinating. This happens because these models are usually trained to take a partial pattern and predict how to complete it. A highly simplified example could be to take the words “the sky is…” and predict the next word in the sequence. The word that most commonly follows these three might be “blue”, and that might be the right answer 80% of the time. But maybe you asked a question about the composition of the atmosphere, or about clouds, or about the color of the sky at midnight – and “blue” would not be the right answer in any of those cases. It is very important to understand this limitation and to confirm any information you get from a chatbot before acting on it. This is similar to the responsibility we have for confirming that the route suggested by our GPS doesn’t lead us into a dead end.

Finally, remember that many of these tools – especially the free versions – are not confidential. Don’t input anything into a chatbot that you would not say out loud in a public place or write on your web site, and don’t use sensitive or personal information.

Now that we understand a bit about what chatbots are and what they can (and can’t) do, here is a list of some tools to get you started. This is in no way a comprehensive list, and new tools are being developed rapidly. 

• ChatGPT / GPT-4: The “OG” of chatbots, particularly useful for generating written content. GPT-4 is the most powerful and, therefore, useful chatbot created by OpenAI. It is available for a monthly subscription fee. If you don’t want to pay for access, you can use GPT-3.5 for free. This is essentially “last year’s model”, less powerful and trained on less data, but can also be quite useful. Try asking GPT to summarize a book or article in 250 words, to write a cover letter for your next job application or a performance review for a star employee, or to write code for something or help you troubleshoot your own code.
• Claude: Developed by Anthropic, Claude is like ChatGPT with a conscience. It was designed with a moral compass, so that it “knows” what is considered moral or ethical. This is intended to prevent harmful responses or interactions with users, which have been reported for other chatbots. Claude can also process more words at a time than many other models including ChatGPT, providing more nuance and context for its responses. Try asking Claude to proofread an essay based on a textbook, write a business plan, or translate song lyrics into another language.
• Gemini: Google’s chatbot, formerly called Bard, is now called Gemini. The free version is accessible as a standalone web page with features that are very similar to ChatGPT. A paid version, called Gemini Advanced, can also utilize your own documents in Gmail and Google Drive.  
• Perplexity: This chatbot is focused on providing factual information with accompanying references. It is accessible for free, with a paid version that offers additional features like image upload. Perplexity is particularly useful if you want to evaluate the sources of the information you are getting, as it provides direct links to the web sites from which it gathers the information that it presents. Try asking it about the history of paper money, or what is the difference between El Niño and La Niña weather patterns, or about chatbots. I used Perplexity to help me research this article.
• DALL-E: DALL-E is a chatbot that specializes in image creation. It was actually available publicly before ChatGPT. As with many chatbots, there are several versions of DALL-E available: DALL-E 2 is available as a standalone product from OpenAI, while DALL-E 3 is available to users of paid ChatGPT subscriptions like ChatGPT Plus or Enterprise. DALL-E 3 is also accessible through Microsoft’s Bing Image Creator.

These are just a few of the many chatbots that are available today. Each has its advantages and disadvantages, and all require some responsibility on the part of the user to make sure that information is accurate and ethical.

Which chatbots have you used and liked? What features do you like about them?