ChatGPT kicked off a new era for the Internet with its explosive November 2022 debut, and it remains an intriguing starting point for those exploring the benefits of generative artificial intelligence (AI). This AI chatbot can significantly reduce the time it takes to complete certain tasks. But it won’t replace any of the tools you already use.
ChatGPT has a few notable limitations that its main competitors, Google Bard and Microsoft Bing Chat, have overcome. First, ChatGPT does not cite its sources, making it impossible to truly trust its information. Second, ChatGPT's data only goes up to 2021, whereas Bard and Bing Chat can both surf the web for news and up-to-date information. At this time, no generative AI chatbot is an Editors' Choice winner because they all have serious drawbacks. ChatGPT is worth exploring casually, however, especially because of its strong, human-like writing skills and helpful coding advice for programmers.
What Are ChatGPT, Generative AI, and LLMs?
ChatGPT falls under the umbrella of generative AI. Unlike other AI systems, such as the one that might recommend what to watch next on Netflix, ChatGPT generates its own content.
It produces written responses to prompts from a user. Other generative AI systems, such as DALL-E and Midjourney create images, but ChatGPT specializes in human-like writing. It can also generate tables full of data and even snippets of computer code. Someday, it may even write its own Netflix show, as the latest season of Black Mirror suggests (and writers in Hollywood are striking against).
ChatGPT and its competitors, namely Google Bard and Microsoft Bing Chat, are often referred to as large language models (LLMs). The technology is built to continuously “learn” from conversations with users. Behind the scenes, AI trainers program the LLM to recognize human language patterns and reproduce them. This includes learning new languages, dialects, and slang. You can ask ChatGPT questions like, “Write a recipe and instructions for beef bourguignon in the voice of Julia Child,” or as one Twitter user did, “Make your answer more bro-ey.”
Even after months of exploring ChatGPT, I am still discovering the size and scope of its capabilities. It changes every day as the model learns more about human life. For that reason, it would be impossible to know everything it can do at one point in time—for example, how does ChatGPT edit legal documents? Does ChatGPT give good medical advice? Can it code an app? How good is it at Mandarin? On any given day, you have to explore it for yourself.
How Much Does ChatGPT Cost and How Do You Access It?
You access ChatGPT on its website, chat.openai.com, as well as on mobile for both Android and iOS. There is a free version, which OpenAI calls the Research Preview, and a paid plan for $20 per month called ChatGPT Plus. Both use an AI model called ChatGPT-3.5 by default, but the Plus version also offers the option to use the more powerful ChatGPT-4 as well as a few other features (more on that later).
To use ChatGPT, you create an account with an email address or by authenticating via an existing Gmail, Microsoft, or Apple account. The company asks for your date of birth, and you need to verify a phone number (which can't be a VoIP number) to prove you’re human and that you live in a country where ChatGPT operates.
Once you’re in, the chat function dominates the interface. It’s so simple and straightforward that I’ve been surprised by how many people I’ve heard say, “Even my [insert non-tech savvy relative] tried it!”
The mobile and desktop versions sync to display the same chat history. They both mostly offer the core chat function, plus the ability to see your conversation history, rename past conversions, and share links to them.
Is It Worth Upgrading to ChatGPT Plus?
For most people and use cases, a ChatGPT Plus account is not necessary. You may want it if you're interested in following along with ChatGPT's latest initiatives, which typically roll out to Plus users first. For example, I had fun using the custom instructions feature to program it to refer to me as Arnold Schwarzenegger, and to replace "yes" and "yeah" with the German "ja!" while this feature was not available on the free version.
ChatGPT Plus Has Both ChatGPT-3.5 and -4
Outside of niche features, the biggest fundamental difference between the free and paid version are the AI models offered. Both default to GPT-3.5, but with a Plus account you can experiment with the more advanced GPT-4. That being said, OpenAI admits GPT-3.5 should suffice. It calls it the “fastest model, great for most everyday tasks” while GPT-4 is its “most capable model" for answering questions that require "reasoning and advanced creativity.” From what I gather, that means GPT-4 helps with more complex calculations, typically for STEM fields. However, Bing Chat also runs on GPT-4 and it is free.
50 Exchanges Per 3 Hours
ChatGPT limits the number of exchanges you can have with GPT-4. It's capped at 50 messages every three hours. There is no limit on the number of exchanges with GPT-3.5. Bing Chat also caps the number of exchanges at 30 per conversation topic. But you don't need to wait three hours, as you can instantly start a new chat.
ChatGPT Plug-Ins
Plus users also get access to plug-ins, or third-party enhancements in the Plugin Store, which is like an app store for ChatGPT. There are plug-ins that search scholarly articles instead of scraping the whole web, create and edit visual diagrams in the chat app, plan a trip using Kayak or Expedia, and parse PDFs. After upgrading to a Plus account, you enable plug-ins through a dropdown menu under GPT-4.
There are more than 70 plug-ins as of this writing, although you can only see the full list of them once you have put down the $20 a month, so it's a gamble whether you'll find a plug-in that's worth it to you.
OpenAI also offers its own plug-in, called Code Interpreter. It helps users perform complex calculations in the Python coding language. "Having access to a very eager junior programmer working at the speed of your fingertips can make completely new workflows effortless and efficient, as well as open the benefits of programming to new audiences," according to OpenAI. Code Interpreter may be worth a Plus account, although I don't have the programming skills to vet it myself.
An Assistant Writer for Bite-Sized, Routine Tasks
One of the best uses for ChatGPT—even the free version— is writing, which tracks for a large language model that specializes in mimicking human language. That said, its uses are limited. For example, I find ChatGPT's writing to read like a dry book report. Plus, it repackages information already on the web rather than generating new content.
Its writing chops are good enough, however, to take tasks off your plate when newness isn’t a priority. For example, it can craft a routine out-of-office email, though I wish it would add "generated by ChatGPT" at the end. It can also provide inspiration for small writing tasks that you plan to polish up and complete on your own.
ChatGPT's Writing Skills
To quantify GPT-3.5’s writing skills compared with its competitors, I put 20 of its answers into a calculator for the Flesch-Kincaid system to determine the grade level of its writing. Impressively, it scored an average of collegiate-level writing (13th grade, or first year of university). Competitors Bard and Bing Chat both scored lower, at 9th and 11th grade, respectively.
That said, ChatGPT's Flesch-Kincaid score for readability, a separate measure, was the lowest at 40. Bing Chat scored 49 and Bard 59. Readability and grade level are inversely proportional; a higher grade level means more complex sentences.
While the grade level results are impressive, the score likely comes from having clear sentence structure rather than any flare or metaphors to keep the reader engaged. ChatGPT has a clinical approach to writing. It always uses an introductory sentence to orient the reader, transitional sentences to switch to a new idea between paragraphs, and a conclusion to summarize what it just said.
ChatGPT also tends to give longer, more verbose responses than Bard and Bing Chat. When I asked each of them to “write an autobiography on Martin Luther King, Jr.,” ChatGPT gave a 500-word answer compared with just 80 for Bing Chat. I then asked ChatGPT to “write a 2,000-word essay on MLK, Jr.,” and it provided one that may work for a middle school or early high school level. I could not have submitted this essay to a college instructor and expected it to pass muster, no matter what Flesch-Kincaid says.
Once I got over the initial amazement of interacting with a chatbot that can write a full essay in a snap, I realized the real-world applications are limited. Tools to detect AI-generated content are on the rise, particularly in schools. Professionally, AI-generated content is not only unprotected by copyright law, it’s also unoriginal and stale. But it’s still fun to ask it to "tell a joke," even if it usually tells the same ones over and over again.
The best way to use ChatGPT for writing is to take inspiration from it, or even parts of its text, and then mix it with your own original work and editing.
Programmers Love ChatGPT, and It Can Help You Become One
ChatGPT’s language skills extend to coding languages. Based on information from online forums and my personal network, the people who seem most satisfied with ChatGPT are programmers. Google made a comment to this effect at its 2023 I/O conference, when a presenter noted that “coding is one of the most popular uses of Google Bard.”
ChatGPT can be a great junior programmer companion (it passed a Google interview to become one) to help with debugging or reducing time spent searching for coding answers on sites like StackOverflow. If you aren’t a skilled programmer yet, ChatGPT might make it easier to learn by generating snippets of basic functions for you in a way no coding course can.
You can easily test the code by pasting it into your system to see if it works, a luxury you don't get when asking ChatGPT for general information that's hard to fact-check. The same goes for Excel functions, which ChatGPT can generate, too. If they don’t work, you can go back to ChatGPT and simply say, “That didn’t work. Is there another approach you can try?” It may even ask questions to better understand your task and refine its output.
A ChatGPT Plus account may be worth it for serious coding endeavors, as long as they aren’t on confidential work projects, but the few programmers I know who use it are on GPT-3.5. I asked Hao Zhang, an assistant professor at University of California, San Diego, who is testing and building AI models, why he doesn’t use ChatGPT Plus or Bing Chat for coding, since Bing Chat is free and it also runs on GPT-4. He said ChatGPT-3.5 is enough, and he prefers ChatGPT’s simple interface without Microsoft ads and all the extra links and functionality.
A Helpful, Semi-Accurate Research Starter
In my own work, I’ve only found ChatGPT helpful for research, though it has limited uses and I still default to Google Search. Most of the time, I don’t feel like taking the time to craft a ChatGPT prompt or carry on a conversation. Scrolling through links works just fine.
That said, ChatGPT can do things Google Search can’t. I asked GPT-3.5 to create a table of electric vehicles (EVs) on the market today. It listed only seven models and their starting prices, which I could copy with one click. The information was incomplete and the prices were inaccurate, but it got me going on the task rather than beginning from scratch.
Since I regularly report on EVs, I know that there are at least three times as many on the market. The incorrect prices are from 2021 because that's when ChatGPT's training data stops. For this reason, the free version of ChatGPT is not a good companion for shopping, booking travel, or doing any online research that you hope will result in a purchase.
ChatGPT does not cite its sources, whereas Bard and Bing Chat do. With no sources, it's hard to build off the research because you don't know where it came from. The tool’s confident, human-like prose can be tempting to take at face value, but it’s important to check any information ChatGPT generates or risk consequences. Fact-checking every single word that ChatGPT generates is exhausting and should deter most people from using it. Still, for personal rather than professional use, I find at least a portion of ChatGPT's output to give me a helpful starting point.
If you’re going to use any generative AI model, ChatGPT and Bing Chat are likely more accurate. An experiment by a team at UC Berkeley found that votes from more than 40,000 people determined GPT-4 gives the best answers of any generative AI model on the market today, followed by GPT-3.5.
GPT-4 With Web Browsing Is Slow, Useless, and Now Disabled
Although OpenAI has temporarily disabled its Browse with Bing feature, it plans to bring it back, so it's worth noting that the feature needs some work. Prior to it being removed, Browse with Bing routinely took much longer than GPT-3.5 to generate answers. It often errored out with a, "Sorry, I cannot complete that request."
I enabled Browse with Bing under GPT-4 and asked it to create a list of EVs on the market, the same prompt I gave GPT-3.5. After two minutes of loading, which felt like an eternity compared with GPT-3.5’s instant results, it returned an ugly table of 11 randomly selected electric vehicles, most of which are not the most popular models. It didn’t even list the Tesla Model Y, the world’s best-selling car. Also, the prices were inaccurate even though it has web access.
What happened? Well, the Browse with Bing function lists the sites it visits for its answers, to its credit, so I could see where it got the information. It visited two Car & Driver articles, one with a full list of EVs available today, which it would have made sense to pull from and cite, and then an article about an Aston Martin EV that doesn't come out until 2025.
A week later, I asked the same question again. This time, it visited the same article on all the EVs available today, but ultimately couldn’t finish the task. It apologized and spat out random specs for the Hyundai Kona EV.
In comparison, Google Bard made a clean table with a handy Export to Sheets option. Its prices were still wrong, though. Bard added a column showing each car's range, a helpful and appropriate addition, though it would also require fact-checking. At the end, Bard suggested additional information, which would only be helpful to someone who's new to EVs.
Later that day, I asked ChatGPT to help me figure out how many Tesla Superchargers there are in the US. I gave it a link to the Tesla website, which lists them all but does not tally them up. ChatGPT claimed it could not do the task, but when I asked Google Bard it returned an answer: 3,478. Bard works with URLs you give it easily whereas with ChatGPT you have to enable a web browser, which is defunct as of this writing.
This ability to scrape a web page and return information from it is not something Google Search can do. Still, you likely should fact-check any information Bard delivers. In my case, I didn't want to spend an afternoon counting potentially 3,478 Superchargers.
Almost all my searches on GPT-4 with web browsing enabled were similarly unsuccessful, making me wonder if this feature even works at all. It often failed to return results and issued an apology, when Bing Chat and Google Bard could do the task for the same prompt.
I searched for concerts happening in Tokyo this weekend. Any basic online search can answer this question, so I expected something similar from ChatGPT. But instead it visited 13 websites over three minutes, then errored out. After that, it tried to placate me with a list of EVs, which it seemed to remember I was interested in previously. Finally, it recommended I visit a "reliable event website" to check for upcoming events.
For a simpler search, GPT-4 with web browsing worked well. I asked it to recommend laptops for college students, and it happily returned a bulleted list from PCMag—though I’d like to see OpenAI pay publications when it regurgitates reporting that costs time, money, and years of expertise to generate.
Is It Safe to Use ChatGPT?
When you sign up for a ChatGPT account, you agree to its terms and privacy policy. OpenAI says it may disclose your personal information to "affiliates, vendors and service providers, law enforcement," and states that anonymized and aggregated information may be shared with third parties and made generally available.
ChatGPT Uses Everything You Enter
ChatGPT uses anything you input to improve its model. That’s why several major companies block ChatGPT on internal networks—for fear of sharing intellectual property that could crop up in responses to users outside the company.
In addition to the data collection that happens automatically within the technology, OpenAI says human AI trainers may look at your conversations. ChatGPT's overview notes, “we review conversations to improve our systems and to ensure the content complies with our policies and safety requirements.”
OpenAI offers a few ways to opt out of it using your data, but it sends mixed messages in this area. The overview says, “We are not able to delete specific prompts from your history. Please don't share any sensitive information in your conversations.”
Can You Opt Out of Sharing Data?
For privacy-conscious users who don't want their conversations with ChatGPT being fed back into the product, OpenAI offers an “opt-out” request form, accessed via a curiously casual Google Sheet. It takes one week to go into effect, according to a representative.
“We understand that in some cases you may not want your data used to improve model performance,” OpenAI says. “You can opt out of having your data used to improve our models by filling out this form. Please note that in some cases this will limit the ability of our models to better address your specific use case.”
You can toggle off Save Chat History & Training in the settings, which goes into effect immediately. Disabling it renders the menu of previous chats on the left side blank.
Between turning off chat history and filing an opt-out request, you’re probably safe, but it’s still not wise to enter private data such as addresses, medical records, and your employer’s intellectual property or trade secrets.
Ethical Concerns
We also can’t forget that ChatGPT comes with a host of ethical concerns, as many past technologies have in their early stages. OpenAI CEO Sam Altman himself says it needs more regulatory oversight, given risks of spreading misinformation and automating jobs on a mass scale. For example, source citations should be table stakes for any responsible AI system, which ChatGPT currently lacks. Between these issues and shortcomings in the tool itself, some might want to wait for the kinks to be ironed out before becoming regular ChatGPT users.
ChatGPT Is Both a Marvel and a Work in Progress
ChatGPT’s conversational capabilities and ability to summarize volumes of source data in coherent paragraphs is why it has become one of the fastest-growing apps of all time. ChatGPT is like an assistant that can get projects going and hand them off to you in a way no other widely available tool can. But I still default to a standard Google search, even after months of reviewing ChatGPT, Google Bard, and Bing Chat. Their shallow, overview-style chatbot answers mean I end up on Google anyway, where the information is richer and easier to fact-check.
For the average user, the overpriced Plus account doesn't warrant $20 per month. Competitors Bard and Bing Chat are better options, as they offer up-to-date information from the web and are free. Bing Chat in particular is likely to be just as accurate and advanced as ChatGPT Plus since it runs on GPT-4 as well. Nevertheless, PCMag is not awarding an Editors' Choice winner in the category of generative AI chatbots at this time, as none are reliable enough to recommend without reservation and the technology still has progress to make.
0 Comments