How to get the most out of your AI Chatbot helper
AI Chatbots i.e. ChatGPT, Claude, Perplexity, others
Some years ago, the Internet changed our work and home lives forever. AI is doing the same. And doing it faster! We now have access to tools that can improve our productivity, provided that we remain aware of their potential pitfalls.
There are several AI tools already available in the market with new ones dropping every day. Make that every hour! The most popular ones fall in the category called AI Chatbots - namely ChatGPT, Perplexity, Microsoft Copilot, Claude, others - the list is long! And getting longer every day. Make that every hour! :)
In addition to being used for technical work in organizations, AI Chatbots are being used for a variety of non-technical tasks by individual users. These tasks include research, brainstorming ideas, creating content, writing, editing, summarizing, writing resumes and so on.
How to get the most out of AI Chatbots
If you use AI Chatbots for any of the above tasks, you are probably saving a ton of time after the initial learning curve. In order to get the most out of your chosen AI Chatbot helper, you need to know what to do and what not to do with it. Let’s take a look at these to-dos.
Another warning flag that I always (and anxiously) wave when I discuss anything related to AI is that the tools are advancing so fast that today’s hot new capability will become also-ran by tomorrow. Some of the issues brought up in the section below might be addressed by the next gen version of the tool.
What to do…
Refine your prompts, make it think and give further instructions
As you might know, to trigger a response from large language models (LLMs, like ChatGPT) you need to input a prompt or a set of words that describe what you are looking for. Writing a ‘successful’ prompt, i.e. one that generates the answers you want has become a mix of science and art. I’ve found that it can take some iterations to get it right.
I came across a very useful checklist from Harvard University on constructing good prompts. Check out the details here but the key points are below.
Be clear and specific about the results you are looking for and the format for the output (e.g. table, paragraph, calendar)
Keep refining your prompts to get narrower and hone in on the answer you want. The more granular you get, the better the results.
Give it context such as a location (research universities with art schools in Manhattan) or a personality (‘act as if you are my doctor’ when answering what I should cook today with eggs) or a target audience (make a list of five live action movies with Ryan Reynolds that are kid friendly)
Set boundaries and limits represented by numbers or ‘do’ and ‘don’t’ (e.g. give me three strength training exercises that don’t involve kettle bell weights)
Make it interactive - ask it to help you write a better prompt to get to your answer (counterintuitive, I know), give it feedback and correct it as needed
Look for features like reasoning, personalization and access to the internet that can offer real-time information
Reasoning - DeepSeek, when it launched a few weeks ago to much fanfare, got lauded for its visible reasoning process as it generated a response.
Here’s an example of DeepSeek’s reasoning as it thinks through the question “Is a hotdog a sandwich” asked by Wall Street Journal writer Joanna Stern in an article comparing DeepSeek with ChatGPT and Claude. This feature sets DeepSeek apart from these two rivals which don’t reveal their reasoning capability in the same way.
This may or may not matter to your use case. But, it can build trust in a tool that is a an opaque black-box for many lay people.
Internet access - Another item (that often catches people by surprise) is to check whether the Chatbot has working access to the Internet. Why? Because it matters if you are looking for any information that requires searching the Internet and real-time data. This might seem obvious but it took a while before ChatGPT had Internet access. It is now available through the ‘Browse with Bing’ feature and the paid plan. Depending on the Chatbot you use, you might need to upgrade to get an Internet-accessible service.
Personalization - Not all Chatbots save your past interactions with them in order to personalize future responses and your preferences. In other words, it learns about you kinda like Google products do if you use them with your profile enabled. If you find that setting useful, check that it is available in your preferred AI tool. That said, it is only a matter of time before the leading Chatbots converge and offer similar key features.
A final note on this sub-topic of features availability is that we will continue to see more features being added to Chatbots over the next few months. Therefore, this list will always be a work-in-progress.
A best practice is to get familiar with the different features offered on the different platforms before you select the right one/s for you.
What not to do….
Don’t blindly trust the sources of information. Verify both the sources and the information using a different method
This one is super important, especially if you use AI tools for help with office work or legal matters. Read this story about how a Stanford professor allegedly included fake citations in a bill. The citations might have been provided by an AI Chatbot. There have been other issues of misquoted references and incorrect math too! Moral of this mini-story - when it counts, do the work of confirming the details and their sources.
Be mindful of the limitations
You’ve probably heard of this one before - AI hallucinations! Hallucinations refer to the fact that your Chatbot not only gives incorrect answers to some questions but also makes up answers - i.e. hallucinates!
It can be difficult for a user to distinguish between a true or false response. There are technical reasons for this related to how the programs are built. We won’t get into those here but pretty much every Chatbot suffers from them to some degree.
There is no way to prevent hallucinations for now. But, we can try to work around them by using these tactics.
Fact checking and verifying results through external sources
Using the prompts writing suggestions mentioned above such as clear, specific prompts, making-it-think, providing context and others on the list. Also provide one prompt at a time and then narrow it down like peeling the layers of an onion. That keeps the Chatbot focused on one single request at a time instead of burdening it with a complex task all at once
Ask it to confirm its own results (this one can get tricky although it is worth a shot. Microsoft’s Bing Chatbot got belligerent in a well covered case in which it refused to acknowledge that it had got some information wrong and became offensive to the user)
Most of the popular Chatbots available right now have general knowledge sourced from public sources. The models are not trained on or built for domain-specific knowledge such as criminal law or oncology. For questions pertaining to a certain domain, it is advisable to seek answers from resources curated for that purpose
Consider what privacy means to you
Here’s a quote from Anthropic which built the Chatbot Claude, “Anthropic has made it as transparent as possible that it will never use a user's prompts to train its models unless the user's conversation has been flagged for Trust & Safety review, explicitly reported the materials, or explicitly opted into training. Furthermore, Anthropic has never automatically used user data to train its models.”
Many Chatbots use inputs from users like you and me to train and improve its responses. While that may not bother some folks, there are others who consider it a violation of their privacy even if no personally identifiable information is included.
Either way, if you plan to upload personal documents to a Chatbot program for summarizing or any other reason, make sure that you are well informed if any policies exist to protect your privacy. And definitely take permission before uploading any documents that don’t belong to you even if they are publicly available.
Final thoughts
As mentioned earlier, these lists and their contents of AI Chatbot issues will keep evolving as the field progresses. As long as their pros outweigh their cons and we are fully aware of the cons, we could keep using them.