How to use ChatGPT – PIRG
Part of the Public Interest Network
$
Close
Close
Search
$
Nav
Close
A guide on how to get (or not get) the most out of OpenAI’s chatbot
Our online life
Director, Our Online Life Program; and Don’t Sell My Data Campaign, U.S. PIRG Education Fund
ChatGPT – the chatbot from the company OpenAI – is everywhere. After it was released to the public in November 2022, it became the fastest-growing consumer app in history. Today over a third of adults in America have used ChatGPT – a number that’s about doubled in just the last two years.
To know how to best use – or not use – the chatbot, it’s helpful to understand how it works.
In the most basic terms, ChatGPT is a chatbot. You can ask it a question or give it a prompt to follow and it quickly spits out a response. It’s really great at some things and pretty darn useless at others.
For example, if you need to use an avocado and cilantro in your fridge that’s about to go bad, you can ask ChatGPT for a recipe that will use them up. Or you can ask it to summarize a calculus idea for you and ask it to explain it like you’re 5, or like it’s a Shakespearean tragedy, or like it’s a comedy sketch.
ChatGPT’s comedy sketch version was 3 pages long and full of very, very bad puns.Photo by PIRG intern Rohan Tayur | TPIN
Today ChatGPT is generally best used for open-ended, text-based tasks where there’s not necessarily one right answer.
Having a basic understanding about how ChatGPT works will help you know when it will be good at a task and when it won’t.
We think the simplest and most helpful way to think of ChatGPT is that it’s like a very sophisticated version of autocomplete that recognizes patterns in text. It strings together phrases based on what it thinks is most likely to come next in a sentence using all the text it’s ever read. The result is a chatbot that appears to respond to your inquiries in a relevant way. ChatGPT exists because the internet came first, with tons of blogs, news articles and chat forums out there to serve as the raw data for it to learn how human speech works and what people have to say.
Our online life
Tips & Guides ●
For example, if you ask ChatGPT what the capital of Texas is, it doesn’t really know the answer is Austin. What it knows is that in all the text it’s read before, when the words “the capital of Texas is” are said, the next most likely word to appear is “Austin”. It can answer your question correctly – but only because it’s been trained on data that includes that fact enough times that it’s almost statistically guaranteed it will say “Austin” next.
It makes sense you can give ChatGPT a list of ingredients like avocado and cilantro and it spits out the idea you should make guacamole. Think of how many hundreds of guacamole recipes there are on the internet.
But that doesn’t mean ChatGPT can reliably write a good recipe from scratch for you. It may pick an amount that doesn’t actually make sense for the recipe, and the result may be inedible. (A personal example: my roommate gave ChatGPT a muffin recipe and asked how to add more protein. Its suggestion resulted in unpleasantly dense muffins and my roommate’s vow to never use ChatGPT like that again.)
What’s happening under ChatGPT’s surface is much more complex (and unpredictable) than your typical autocomplete. It’s a large language model (LLM) that passes text through layers of its neural networks, breaking down your request into pieces, interpreting those pieces, and constructing a response.
The weird truth is that we don’t know with 100% certainty why a chatbot produces the exact answer it does when it does it. The data it’s pulling from to create its responses is unimaginably huge, and there’s some level of randomness at play. You can ask it the exact same question with the exact same wording twice in a row and it will give you a slightly different output.
That said, we think understanding the basics is enough to figure out when it might be useful for you.
You can access ChatGPT by visiting chatgpt.com or by downloading the ChatGPT app on your phone or other device. You can use the chatbot for free using an account linked to your email, or use it logged out entirely. Buying a paid subscription gives you access to more of its features.
When you open ChatGPT you’ll see a very basic screen. You can type anything into the chatbox and ChatGPT will write a response.
Photo by TPIN Staff | TPIN
The words you choose to use are referred to as “prompts.” For example you can ask it: “tell me a horse joke”, “write me an Instagram post” or “how should I prepare for my upcoming job interview?” Some use it to edit their writing or for brainstorming, like coming up with ideas for a child’s birthday party.
ChatGPT can be good at these types of open-ended, text-based questions where there’s not one right answer. For other uses, however, it can run into problems – and there are other reasons, like privacy concerns, you want to keep in mind as well.
There are some things you don’t want to use the chatbot for.
The most important thing to understand is that ChatGPT is not always great at facts. If you need an answer to a question with one right answer, ChatGPT is not necessarily the most reliable tool. The chatbot even has a disclaimer under its chatbox: “ChatGPT can make mistakes. Check important info.”
Now, sometimes ChatGPT does produce correct facts. In particular it’s more likely to be right about common facts that have been true for a long time and are therefore likely to appear in its training data more frequency than newer facts or those that are more obscure. Austin has been the capitol of Texas for a long time, and plenty of homework sites explain the basic concepts of calculus that haven’t changed much either.
But ChatGPT also gets stuff wrong. And when it does, it can present information with the same appearance of confidence, making it easy to believe it’s telling the truth when it’s really not. As one expert put it, using ChatGPT is like “chatting with an omniscient, eager-to-please intern who sometimes lies to you.” These untrue slip-ups are often referred to as “hallucinations.”
Hallucinations cause real problems. Lawyers using ChatGPT to help write legal briefs have submitted faulty court documents with citations to fake cases. One man ended up in the hospital after asking ChatGPT about lowering sodium in his diet and began substituting table salt with sodium bromide. Sodium bromide is known for triggering psychiatric episodes.
Maybe someday you’ll be able to trust ChatGPT more. According to the company, ChatGPT’s ability to be factually correct is greatly improving.
For now we’d argue it’s best to avoid ChatGPT when you need to be right about something. And really don’t use it where being wrong comes with big consequences, like seeking legal, financial or medical advice.
Consumer alerts
Asking ChatGPT a question isn’t the same as doing a (normal, non-AI) Google search. It’s more a pattern recognition machine than a web search tool.
Until recently, ChatGPT did not have access to the real-time internet if you were a free user. That meant it couldn’t give you up-to-date information. The responses it could genereate was tethered to whenever the company last updated its training data. That’s different now. OpenAI rolled out the ChatGPT search feature for all users – including free tier and logged out users – in February 2025, giving you access to a real-time internet connection.
But even with access to up-to-date information on the web, ChatGPT can still fail to give you current information.
While writing this article, for example, I turned on ChatGPT’s web search feature and asked if free users have access to the real-time internet. It told me no, despite multiple pages on OpenAI’s own website explaining this feature is now available to everyone. The first news stories it cited were both from April 2024.
ChatGPT tells me it can’t access the internet…while it is accessing the internet. (Interaction from Sept. 2025)Photo by TPIN | TPIN
It’s possible if I had phrased this question differently it may have given me different information. Or even if I asked the same question again. But that’s a part of the trickiness of the chatbot – even when you think you’ve been clear, you never exactly know what you’re going to get.
It did better when I asked it for the local weather forecast. It did so-so when I asked it for movie times this weekend at the local theatre. ChatGPT flagged that the sources it was looking at, like IMDB, seemed to have out-of-date information and suggested I look at the theatre’s website. The theatre’s website does in fact have a page displaying movie showing times. But for whatever reason the chatbot didn’t go access the page on its own and produce the information, or even give me the link to the right specific page.
I would personally not use ChatGPT for retrieving any information that needs to be timely. (Or, again, relying on it for anything factual at all.)
Because ChatGPT’s outputs are combining text from a huge number of sources, it’s hard to know what its output to any particular question will be. It’s possible that text could end up being the same text used somewhere else. For example, The New York Times has filed a copyright infringement lawsuit against OpenAI and Microsoft. Its complaint included examples of chatbot results with near-verbatim excerpts of the paper’s copyrighted journalism. You don’t want to rely on the chatbot in a situation where being perceived as plagiarizing could get you in trouble.
Right now, if you’re using a consumer account, it’s quite possible that whatever you put in ChatGPT’s chatbox is being stored somewhere on OpenAI’s – or a third party’s – servers.
OpenAI keeps your chat logs by default (and if you’re using a free account, opts you in to having your chat history used to further train ChatGPT by default, too). The only way to delete your chats is to delete your account, manually delete chat logs, or turn on the “temporary chats” feature, which the company says will automatically delete your chat logs after 30 days.
Usually.
Right now, those options may not actually delete your data.
That’s because in The New York Times court case we mentioned above, a judge has recently ordered the company to stop deleting user chats – even if you submit a deletion request – because deleting chats could be deleting evidence. It’s currently unclear what telling OpenAI to delete your data will actually do.
Anytime a company is storing your data, it can cause problems for you. The more data companies collect about you and the longer they store it on their servers, the more likely it is it will eventually be exposed in a breach or a hack.
Even without a hack it’s possible your conversations could end up public. For example, it was reported in July that some ChatGPT users’ conversations were exposed in Google search results. In particular, don’t give ChatGPT your medical or financial information.
My rule of thumb is that I don’t say anything to ChatGPT that I wouldn’t be ok with someone else reading. Better safe than sorry.
Corporate responsibility
In summary: sometimes yes and sometimes no. It all comes down to what you’re trying to use it for. If you really need to be right about a specific fact or need to be right because it’s a high stakes situation like getting medical advice, we’d strongly suggest going elsewhere. If you want to use it for open-ended text based tasks like drafting thank you notes or brainstorming names for your new kitten, those are tasks it’s better suited to do.
At the end of the day it’s best to remember: you can’t trust ChatGPT to tell you the truth.
Director, Our Online Life Program; and Don’t Sell My Data Campaign, U.S. PIRG Education Fund
R.J. focuses on the intersection of tech and people. Her work ranges from the risks of commercialization of personal data, toconsumer harms like scams and data breaches, to manipulative targeted advertising, to keeping kids safe online. In her work at Frontier Group, she has authored research reports on government transparency, predatory auto lending and consumer debt. Her work has appeared in WIRED magazine, CBS Mornings and USA Today, among other outlets. When she’s not protecting the public interest, she is an avid reader, fiction writer and birder.
Our online life
Our online life
Right to repair
Our online life
Donate Update your membership
© 2025 PIRG
Visit the website of U.S. PIRG Education Fund, our sister 501(c)(3) organization.
Close
Close
