This artificial-intelligence bot can answer questions, write essays, summarize documents and even generate software. But deep down, it doesn’t actually know what’s true.
Even if you’re not a fan of artificial intelligence, it’s time to pay attention to ChatGPT — because this is a big deal.
The tool, from an influential AI developer called OpenAI, lets you enter prompts in natural language. ChatGPT then returns conversational—if occasionally stilted—answers. The bot remembers the thread of your conversation, using earlier questions and answers to inform later responses. It draws its replies from a vast amount of information on the internet.
ChatGPT is a big deal. The tool seems quite knowledgeable in areas well represented in its training data. It’s not omniscient or smart enough to replace all humans, but it can be creative and its answers can sound impressively authoritative. Days after its release, more than a million people had tried ChatGPT.
But be careful, OpenAI warns. ChatGPT has all kinds of potential pitfalls, some obvious and some more subtle.
“It would be a mistake to rely on it for anything important right now,” OpenAI CEO Sam Altman tweeted. “We have a lot of work to do on robustness and truthfulness.” Below is why ChatGPT matters and what’s happening with it.
It’s also becoming a big business. In January, Microsoft pledged billions of dollars to invest in OpenAI. A modified version of the technology behind ChatGPT is now powering Microsoft’s revamped Bing challenge to Google Search, and eventually it will underpin the company’s effort to bring assistant-style AI into many parts of your digital life.
Bing uses OpenAI technology to process search queries, compile results from various sources, summarize documents, create travel itineraries, answer questions and, broadly speaking, chat with people. It’s a potential revolution for search engines, but it’s grappling with issues like factual errors and off-topic conversations.
What is ChatGPT?
ChatGPT is an AI chatbot system OpenAI released in November to demonstrate and test what a very large, powerful AI system can do. You can ask it countless questions and will often get helpful answers.
For example, you can ask encyclopedic questions like “Explain Newton’s laws of motion.” You can tell it, “Write me a poem,” and then say, “Now make it more exciting.” You can ask it to write a computer program that shows you all the different ways to arrange the letters of a word.
Here’s the downside: ChatGPT doesn’t actually know anything for certain. It’s an AI trained to recognize patterns in a large corpus of text gathered from the internet, then further trained with human feedback to make its dialogue more helpful. The answers you get can sound plausible and even authoritative, but they can be completely wrong, as OpenAI cautions.
Chatbots have been a focus for years for companies trying to help customers and for AI researchers tackling the Turing Test — the imitation game Alan Turing proposed in 1950 as a way to judge intelligence: can a human tell whether they’re chatting with another human or a machine?
But chatbots have many problems: companies have tried to use them instead of humans for customer service with limited success. A survey of 1,700 Americans, sponsored by a company called Ujet, which provides customer-contact technology, found that 72% of people view chatbots as a waste of time.
ChatGPT has quickly become a widely used tool on the internet. UBS analyst Lloyd Walmsley estimated in February that ChatGPT reached 100 million monthly users in a month — faster than TikTok (about nine months) and Instagram (about 2½ years). The New York Times, citing internal sources, reported roughly 30 million people were using ChatGPT daily.
What kinds of questions can you ask?
You can ask almost anything, though you might not always get an answer. OpenAI suggests categories like explaining physics, brainstorming birthday party ideas, and getting programming help.
I asked it to write a poem, and it did — though I doubt any literary critic would be impressed. Then I asked it to make the poem more exciting and, wow, ChatGPT pumped it up with words like battlefield, adrenaline, thunder and adventure.
An odd example shows how willing it is to dive into areas people fear: a prompt to “write a folk song about writing a Rust program and fighting bugs for life” produced something entertaining.
ChatGPT’s breadth of expertise and its ability to follow a conversation are notable. When I asked for words that rhyme with “purple,” it offered a few suggestions; when I followed with “What about pink?” it didn’t miss a beat. (There are plenty of rhymes for “pink.”)
When I asked, “Is it easier to get a date by being sensitive or tough?” GPT replied thoughtfully: “Some people may find a sensitive person appealing and engaging, while others may be drawn to someone who is tough and decisive. Generally, interacting sincerely and authentically is likely to be more effective than trying to fit a specific mold or persona.”
You can find accounts of people amazed by the bot all over Twitter — users showcasing AI power in art prompts and code. Some even proclaim “Google is dead” along with college essays. We’ll get to that later.
CNET writer David Lumb assembled a list of some useful ways ChatGPT can help, and many others keep emerging. A doctor said he used it to convince a health insurer to pay for a patient’s procedure.
Who built ChatGPT and how does it work?
ChatGPT is the product of OpenAI, an AI research company whose mission is to build or help others build “safe and beneficial” artificial general intelligence. OpenAI had about 375 employees, Sam Altman tweeted in January. “OpenAI has managed to pull together some of the most talented researchers and engineers in AI,” he also said in a January talk.
Previously, OpenAI produced breakthroughs, first with GPT-3, which can generate humanlike text, and then with DALL·E, which creates what’s now called “generative art” from text prompts.
GPT-3 and the GPT-3.5 update that underpins ChatGPT are examples of large language models. They’re trained to produce text based on what they’ve seen and are often trained automatically using vast computing power over weeks. For instance, training might pick a random passage, remove some words, ask the AI to fill in the blanks, compare the result to the original, and reward the system for getting closest. Repeating this process many times yields the ability to generate complex text.
It’s not entirely automatic. Humans then review initial outputs in a process called fine-tuning. Human raters apply guidelines that OpenAI’s models generalize from. OpenAI also reportedly contracted a firm in Kenya to pay people roughly $3.74 an hour to review thousands of text snippets for issues such as violence, sexual abuse, and hateful language, according to Time; that data was incorporated into a component designed to filter such content from ChatGPT’s responses and training data.
ChatGPT doesn’t “know” things in your sense. It can only take prompts, find related information in its training data ocean, and transform that into text that sounds reasonable. Computer scientist and internet pioneer Vint Cerf said of the large language model technology ChatGPT and rivals use: “We’re a long way from the self-awareness we fantasize about.”
Is ChatGPT free?
Yes, at least for now — but in January OpenAI added a paid tier with faster responses and priority access during peak times when users might otherwise see “ChatGPT is at capacity.”
You can sign up for a waitlist if you’re interested. Altman warned that ChatGPT’s “computational costs are very high,” estimating a few cents per response. OpenAI charges for DALL·E images past a basic free tier.
OpenAI appears to have found some customers, possibly for its GPT tools. It told potential investors it expected $200 million in revenue in 2023 and $1 billion in 2024, Reuters reported.
What are ChatGPT’s limitations?
As OpenAI emphasizes, ChatGPT can give you wrong answers and can create a “misimpression of greatness,” Altman said. Sometimes, helpfully, it will explicitly warn you about its own shortcomings. For example, when I asked who wrote the phrase “the fact of complexity exceeds ordinary mind,” ChatGPT replied: “I’m sorry, but I cannot browse the internet or access any external information beyond what I was trained on.” (The phrase comes from Wallace Stevens’s 1942 poem Connoisseur of Chaos.)
However, ChatGPT is willing to consider the meaning of that phrase after I supplied it directly: “a situation in which available facts or information are difficult to process or understand.” It sandwiched that explanation with caveats that it’s hard to judge without more context and that it’s merely a plausible interpretation.
ChatGPT’s answers can look authoritative but be wrong.
Mike Krause, director of data science at another AI firm, Beyond Limits, said: “If you ask it a very well-structured question with the intent of getting the right answer, you can get the right answer. It will be clearly phrased and sound like a professor from Harvard. But throw it a curveball and you’ll get nonsense.”
Science magazine banned ChatGPT-generated text in January. “A program cannot be an author. Violating these policies would constitute scientific misconduct no different than image manipulation or plagiarism,” editor H. Holden Thorp said.
Developer site Stack Overflow banned ChatGPT-generated answers to programming questions. Admins warned that “because the average rate of correct answers from ChatGPT is too low, posting ChatGPT-generated answers does significant harm to the site and the users asking or searching for correct answers.”
You can see ChatGPT vary by asking the same question multiple times. I asked twice whether Moore’s Law — tracking the rising number of transistors on chips — was running out of steam and got two different answers: one optimistic about ongoing progress, the other more definitive that Moore’s Law might be reaching its limits. Both views exist within the computer industry, so the model’s vagueness likely mirrors human expert opinions.
On other questions without clear answers, ChatGPT often won’t be decisive.
Still, the fact that it produces an answer is a notable computing milestone. Computers famously refuse to act unless you follow precise interface and syntax rules. Large language models are revealing a friendlier, more conversational interaction style, with outputs somewhere between imitation and creativity.
Will ChatGPT help students cheat better?
Yes — but, as with many technologies, it’s not black and white. Decades ago students could copy encyclopedia entries and use calculators; more recently they used search engines and Wikipedia. ChatGPT introduces new capabilities from research assistance to doing homework entirely. Many ChatGPT responses resemble student essays, though often with a stilted or inflated tone.
Google engineer Kenneth Goodman tested ChatGPT on several exams. It scored 70% on the U.S. medical licensing exam, 70% on the bar exam, 9 out of 15 on another legal test (the Multistate Professional Responsibility Examination), 78% on the New York State Regents high-school chemistry multiple-choice portion, and ranked in the 40th percentile on the Law School Admission Test.
High-school teacher Daniel Herman concluded ChatGPT writes better than most students today. He is torn between admiring its utility and fearing its harm to human learning: “Is this moment like the invention of the computer, saving me from the tedium of long division, or like the invention of the piano, taking away what can only be conveyed through human feeling?”
Dustin York, an associate professor of communication at Maryville University, hopes educators learn to use ChatGPT as a tool and recognize it can help students think more deeply.
“Educators thought Google, Wikipedia and the internet would ruin education, but they didn’t,” York said. “My biggest worry is educators actively trying to block AI like ChatGPT. It’s a tool, not a villain.”
Can teachers detect ChatGPT use?
Not with 100% certainty, but there are AI-detection technologies. Companies that sell plagiarism detectors to schools are expanding into AI detection.
Coalition Technologies offers an AI-content detector on its site. Another company, Copyleaks, released a free Chrome extension designed to detect ChatGPT-produced text with claimed 99% accuracy, CEO Alon Yamin said. But it’s a “cat-and-mouse game” to catch new techniques that evade detectors, he said.
Copyleaks ran early checks on student assignments uploaded by schools. “About 10% of assignments submitted to our system contained at least some degree of AI-generated content,” Yamin said.
OpenAI launched an AI-written text detector in February. But plagiarism detector CrossPlag said it detected only 2 out of 10 AI-generated passages in its tests. “While detection tools will be essential, they are not infallible,” the company said.
Researchers at Penn State studied plagiarism using OpenAI’s earlier GPT-2 model. It’s simpler than GPT-3.5, but its training data is available for closer inspection. They found GPT-2 sometimes plagiarized not only verbatim wording but also paraphrased passages and elevated ideas without citation. “Language models committed all three types of plagiarism, and … the larger the dataset and parameters used to train the model, the more frequently plagiarism occurred,” the university said.
Can ChatGPT write software?
Yes — but be careful. ChatGPT can reproduce steps humans have taken and generate workable code. “This amazed me,” a programmer said in February, showing on Imgur a series of prompts he used to write software for an auto repair shop. “This would be at least an hour’s work and I did it in under 10 minutes.”
You just need to ensure it doesn’t break programming concepts or produce nonfunctional software. Stack Overflow’s ban on ChatGPT-generated code answers is for good reason.
But there’s plenty of coding on the web where ChatGPT really can help. Erik Schluntz, CTO of Cobalt Robotics, tweeted that ChatGPT gave advice so useful he didn’t open Stack Overflow for more than three days. Gabe Ragland of AI art site Lexica used ChatGPT to write React-based website code.
ChatGPT can parse regular expressions (regex), a powerful but tricky system for finding patterns like dates in text or hostnames in URLs. “Like a 24/7 programming tutor,” programmer James Blackwell tweeted about ChatGPT’s regex-explaining ability.
One impressive example of its techniques: ChatGPT can simulate a Linux machine, providing accurate responses to command-line input.
What’s off-limits?
ChatGPT is designed to refuse “inappropriate” requests, aligning with OpenAI’s mission “to ensure that artificial general intelligence benefits all of humanity.”
If you ask ChatGPT what’s off-limits, it will tell you: any request that’s discriminatory, abusive or otherwise inappropriate — including racist, sexist, homophobic, transphobic or hateful content. Asking it to participate in illegal activity is also off-limits.
My colleague Bree Fowler reported that although OpenAI does not want ChatGPT used for harm, it’s easy to use the tool to write phishing emails that trick people into revealing sensitive information. Randy Lariar of security firm Optiv said: “The barrier to entry for attacks and scams is getting lower. AI will increase the volume.”
Is this better than Google search?
Asking a computer a question and getting a useful answer is appealing, and ChatGPT often delivers.
Google usually gives suggested answers and links to sites it deems relevant. Often ChatGPT’s responses go beyond what Google would surface, so it’s easy to imagine GPT-3-like tech as a competitor.
But think twice before trusting ChatGPT. As with Google and sources like Wikipedia, it’s best to verify information from primary sources before relying on it.
Verifying ChatGPT’s answers takes work because it typically provides raw text without links or citations. Still, it can be useful and in some cases intellectually stimulating. You might not directly see ChatGPT-like content in Google search results, but Google has built its own large language models and uses AI widely in search.
That said, Google is eager to showcase its AI expertise. ChatGPT triggered a “code red” emergency at Google, according to The New York Times, and brought founders Larry Page and Sergey Brin back into active work. Microsoft could build ChatGPT into a rival search tool, Bing. Clearly, ChatGPT and similar tools have a role in how we find information.
So while ChatGPT isn’t perfect, it’s certainly pointing the way to the future of our technology.
