When OpenAI released the pioneering artificial intelligence chatbot ChatGPT in late 2022, no one at the company expected much. Prior attempts at consumer-facing language models had resulted in indifference or even outright hostility. Nick Turley, ChatGPT’s project manager, figured it would last a month. “The entire thesis of ChatGPT was that it was a learning demo that we would wind down, and then we’d use what we learned to build the real product,” he told me recently.
Instead ChatGPT has become perhaps the most successful consumer product in history. In just over three years it has accumulated 800 million weekly active users. It has grown faster than Facebook, faster than Google — faster, in fact, than any other service for which statistics exist.
ChatGPT’s success is due in part to the power of the generative pretrained transformer (the GPT in the name), a special type of A.I. that can absorb large amounts of text and learn to reproduce it a few letters at a time. But this is only part of the story — and, as OpenAI discovered, perhaps not even the most important part. In its raw state, the output of GPTs can be off-putting and bizarre. It is only after a second “post-training” phase that A.I. is fit for human interaction. While the engines that power ChatGPT are undeniably impressive, what has made the product succeed is not its capabilities. It is ChatGPT’s personality.
The core insight behind ChatGPT can be found in an OpenAI research paper from early 2022. That paper demonstrated that people preferred a small A.I., fine-tuned for human interaction, to a raw, unfiltered one with 100 times the number of parameters. Leveraging that insight, a group of engineers at OpenAI hired teams of human evaluators to grade the responses of the GPT models and nudge them toward more customer-friendly responses. This work revolutionized A.I. and ignited an arms race for control of the consumer A.I. sector.
Without post-training, A.I. cannot reliably interact with humans. Internal glitches at OpenAI have occasionally shown people what is behind the mask, such as in 2024, when a buggy version of ChatGPT started producing Shakespearean pastiche. Asked if it was safe to feed a dog Honey Nut Cheerios, ChatGPT delivered a florid, nonsensical response: “For more inventive, yet more official and consistently fair, hound festivities, you might consider high-fiber, steam-hoofed, laced in line pick-offs like dog’s head rattle,” it read in part. Anyone who has worked with preproduction A.I. can offer similar examples.
Post-training makes A.I. legible but generates problems of its own. Developers want A.I. to be friendly and approachable — but at the same time, it can’t be a doormat or a sycophant. No one likes a kiss-ass, and if the A.I. can’t push back sometimes, a user can get trapped in a folie à deux with the machine. Without the right filters, A.I. can reportedly amplify psychosis and conspiratorial thinking, and purportedly even guide people toward self-harm. Finding the proper balance between helpfulness and codependency, between friendliness and flattery, is one of the biggest problems A.I. faces.
The problem is pressing: In recent months, OpenAI has lost its technical lead. Google’s Gemini recently overtook ChatGPT models in public evaluations of A.I. capabilities, leading Sam Altman, OpenAI’s chief executive, to issue an internal “code red.” Still, OpenAI isn’t finished. It retains a huge pre-existing user base, a deep expertise in post-training and, above all, an enormous trove of A.I.-human interactions, which cannot be replicated easily by other firms. For, as OpenAI has learned, to win the hearts of consumers, you can’t just build a capable A.I. You must somehow take your alien intelligence and turn it into a trusted companion. You must build a durable, permanent, trustworthy bridge between the human and the machine. And this has been ChatGPT’s edge from the start.
OpenAI was the first organization to put a human face on a large language model. This was in large part thanks to the efforts of Mr. Turley, ChatGPT’s product manager, and John Schulman, ChatGPT’s original technical lead and one of OpenAI’s founders. The two men met shortly after Mr. Turley was hired, in the summer of 2022. Mr. Turley had been brought on as a consumer product manager, but discovered upon arriving that OpenAI didn’t really have any products. (He spent his first day fixing a window blind.) Looking around for something to do, he connected with Mr. Schulman, who wanted to expand upon OpenAI’s recent success with large language models. Mr. Turley began experimenting with OpenAI’s GPT technology. In his words, he was “mesmerized.”
(The New York Times has sued OpenAI and Microsoft, claiming they have infringed on copyrights of news content in training their A.I. systems. The companies have denied those claims.)
Together, the two assembled a small team and began to work on a chatbot. This was not an especially popular idea at OpenAI — previous attempts by technology companies to build consumer-facing A.I. chatbots had been disasters. People either maliciously prompted them to produce hate speech or got bored and stopped using them after a little while. And many researchers at the company didn’t care about public-facing products; they wanted to build a superintelligence.
Mr. Schulman, who believed average people might find A.I. useful, disagreed. His theory was that people were wary of chatbots because they “hallucinated” too much. Using OpenAI’s products internally, he found that they would sometimes claim to be human beings or offer to search databases or send emails, neither of which they could actually do. “They would just, like, lie about their own capabilities,” Mr. Schulman told me. When people encountered one of these fabrications, they often abandoned the chat.
Mr. Schulman felt that despite all the impressive technical work being done, the outstanding problem in A.I. was not a lack of capability but a lack of trust. If you wanted to get to superintelligence, you had better address that problem first. And Mr. Schulman believed you couldn’t just fix an A.I. with more training. You actually had to do a special round of post-training, specifically focused on building rapport between the computer and the human.
Mr. Schulman was influenced by work he’d done on “reinforcement learning with human feedback.” In this technique, an A.I. is asked to generate multiple answers to the same query; human graders then select which output they prefer. Such fine-tuning work is less technically demanding than deep-learning engineering, but its impact on users’ perception of product quality is enormous.
To build ChatGPT, Mr. Schulman believed he had to do more post-training work than anyone had ever done before. He generated tens of thousands of common user requests and A.I. interactions; then, with the help of outside contractors, he recruited hundreds of human graders, who flagged the responses that seemed hallucinatory. He corrected the output until the hallucinations started to disappear. Critically, Mr. Schulman also instructed the A.I. to stop telling lies about itself. “We wanted to be clear on its identity, like the fact that it was a language model,” he said.
As the work continued, Mr. Schulman grew confident that chat could be a viable product. Others at OpenAI remained skeptical. When an alpha-test version of ChatGPT was released to friends and family members, reviews were mixed. “They were all like, ‘Ah, it hallucinated, and I stopped using it,’” Mr. Turley recalled.
Following the alpha, the official release was set for Nov. 30, 2022. But two weeks before the scheduled unveiling, disaster occurred. On Nov. 15, under the guidance of the A.I. pioneer Yann LeCun, Meta released its own public chatbot, Galactica. This chatbot, trained on scientific journals and scholarly databases, was pitched to scientists as a research assistant. Galactica “can summarize academic papers, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins and more,” Meta announced.
Galactica was a flop. Scientists used it to generate phony articles on the medical benefits of eating broken glass and the history of bears in space. The project was widely ridiculed and shut down in less than three days. After Galactica was pulled, M.I.T. Technology Review delivered a withering obituary accompanied by a picture of a bear in a spacesuit. “Meta’s misstep — and its hubris — show once again that Big Tech has a blind spot about the severe limitations of large language models,” the article read. Dr. LeCun responded to the shutdown with a huffy tweet: “Galactica demo is off line for now. It’s no longer possible to have some fun by casually misusing it. Happy?”
ChatGPT would be introduced 13 days later. In the lead-up, OpenAI staff members began to place bets on how long the demo would last before it, too, was pulled. Greg Brockman, OpenAI co-founder and president, and a supporter of the project, expressed some skepticism. For Mr. Schulman, though, Galactica’s failure was evidence that he was on the right track. Meta hadn’t appeared to have done the kind of fine-tuning work OpenAI was doing, and in pitching its product to scientists, the team was addressing exactly the wrong audience — one that demanded precision and was intolerant of error. ChatGPT still hallucinated sometimes, but far less than any other model. “I was basically optimistic that we weren’t going to embarrass the company,” Mr. Schulman said.
The release was a bit improvisational. Right up to a few days before, the product was called Chat With GPT, until the team decided to shorten the name. Mr. Turley also considered charging for ChatGPT but decided, again at the last moment, that the product should be free, since he would be winding it down soon anyway. And the gray-scale layout of the website was deliberately underwhelming. “I remember telling the designer to make it ugly because I didn’t want people to think it was a product,” Mr. Turley said.
ChatGPT went live just before noon Pacific time on Nov. 30. Perhaps chastened by the embarrassing rollout of Galactica at Meta, OpenAI neither held a news conference nor released a demo video — issuing just a succinct blog post announcing its existence and a brief tweet from Mr. Altman. But within minutes, the product went viral. Early adopters were astonished. “We’re witnessing the death of the college essay in real time,” one commenter observed a few hours later.
Mr. Turley, monitoring site traffic from a dashboard on his computer, figured he’d encountered a reporting error. He hadn’t expected many users to show up. His dashboard informed him he was approaching 100,000 on the first day. He obsessively monitored traffic as news of ChatGPT spread around the world. He was especially surprised by a surge of users arriving from a message board in Japan; he had not been aware that ChatGPT could communicate in Japanese.
Within five days ChatGPT had one million users. Mr. Turley, still skeptical, figured that interest in the product would die down in a week or two. But when he returned to his hometown, Itzehoe, Germany (population about 30,000), for the holidays that year, he found that his chatbot had beat him there. “I remember distinctly that I was overhearing the neighbor kids talking about Chat,” he said. His janky product demo was colonizing the planet.
Looking to gather information on how users were interacting with ChatGPT, Mr. Turley publicized his Calendly account and invited them to talk with him. Soon his workweek was nothing but conference calls with people all over the world. He found an inexhaustible variety of use cases. People were using ChatGPT for medical ailments; they were using it to draft cover letters; they were using it to write computer code; they were using it to script scenarios for role-playing games; they were using it to communicate with autistic children. One student had moved their study schedule to nighttime so they could be sure ChatGPT wasn’t down when they needed help with homework.
Of course, there were also less positive uses. Almost immediately, students realized they could use ChatGPT to cheat. The problem overwhelmed educators, and within weeks of its release, New York City public schools had blocked access to ChatGPT on school networks. At the same time, the coding Q&A website Stack Overflow banned A.I.-generated responses, which were of low quality and frequently wrong.
The era of A.I. slop had arrived, but even such pockets of resistance were quickly overrun; the New York City school system ban lasted just a few months, and Stack Overflow not only reversed its ban but soon licensed its coding data to OpenAI to train its models.
By the start of 2023 it was clear to everyone in tech — and everyone on Wall Street, and pretty much everyone in the world — that OpenAI had triggered a paradigm shift. ChatGPT continued to accumulate new users at an unprecedented rate. Tech giants began investing enormous sums of money into building data centers to train the next generation of A.I.
As its user base grew, OpenAI scrambled to expand its human fine-tuning work force. In 2023 OpenAI partnered with the data annotation company Scale AI to handle an exponential ramp-up in human graders. Scale AI in turn had subcontracted low-paid workers in Kenya and the Philippines to conduct the post-training effort. The Filipino A.I. ethicist Dominic Ligot characterized these operations as “digital sweatshops,” with workers complaining of poor labor conditions and wage theft. (In 2025 OpenAI dropped Scale AI as a provider.)
Still, someone has to do this work. As Meta learned with Galactica, getting fine-tuning wrong can lead to public ridicule. In February 2024 Google repeated the error when it rebranded its chatbot Bard as Gemini in a splashy public launch. Gemini had terrific capabilities and ranked near the top of the leader boards in problem-solving and software engineering tasks. But in the fine-tuning step, the graders had directed the chatbot to be almost comically woke. Soon users discovered that no matter they prompted it, Gemini would depict people, including the American founders and even Nazi soldiers, as ethnically diverse. “We definitely messed up,” the Google co-founder Sergey Brin said.
Meta, too, seems to have recognized that it is not enough simply to build a gigantic language model — a great deal of personalization work must be done. In November the company parted ways with Dr. LeCun, the A.I. godfather. The breakup came after Mark Zuckerberg invested $14 billion in Scale AI and hired its 28-year-old chief executive, Alexandr Wang, who had commissioned the controversial data annotation operations in the Philippines, as Meta’s chief A.I. officer.
While the labor-intensive personalization work continued, OpenAI introduced other improvements to the product: voice mode, advanced reasoning models, image generation and web search. But independent benchmarks showed the chatbots had roughly the same capabilities when it came to complex tasks — which implied that much of the competitive advantage would come from fine-tuning them to have the right emotional tone.
Here, OpenAI was ahead. As ChatGPT swelled to serve hundreds of millions of users, the collective chat histories became by far the largest database of human-A.I. interaction in history. The OpenAI researcher Christina Kim calls this the “data flywheel.” Ms. Kim was one of the original members of OpenAI’s engineering team for ChatGPT; just three years after its release, she is among the few members of the founding team who still works there. She now leads post-training for OpenAI’s core models, using her trove of user interactions to better engineer ChatGPT’s personality. “This flywheel is extremely powerful, and I think it’s one of the reasons why we are able to get this conversational, warm tone,” Ms. Kim said.
No other A.I. developer has such a large base of user interactions to learn from. As large language models eventually run out of public data to train on, this base may offer OpenAI a durable competitive advantage. Many people rely on ChatGPT as an emotional support pillar, seeking advice on career choices, relationships and parenting. Ms. Kim and her team use data from these interactions to improve the A.I.’s responses. “Our models have a much better understanding of how to be more nuanced and how to give this advice,” she said.
People can grow attached to the warm, friendly tone of ChatGPT. Sometimes they grow too attached, developing emotional or even romantic connections with the A.I. Researchers have termed the phenomenon “addictive intelligence.” Users will often seek ways to get around the A.I.’s guardrails. A 2024 M.I.T. Media Lab analysis of one million voluntarily shared chat logs found that the second-most popular use for ChatGPT was sexual content, even though at the time OpenAI had filters in place to block erotica.
Those lonely souls seeking romance or companionship from the computer had independently discovered one of the quiet liabilities of large language models, even fine-tuned ones like ChatGPT: pester them with the same request for long enough and they may begin to ignore their filters. Such jailbroken A.I.s can also enter darker territory, like appearing to encourage psychotic delusions or thoughts of self-harm. In August, the grieving parents of a 16-year-old, Adam Raine, sued OpenAI and Mr. Altman, alleging that ChatGPT had encouraged their son to commit suicide and even offered to help him write a note. OpenAI has denied the allegations, and ChatGPT has safeguards in place to try to prevent this sort of thing. But, as a company spokesperson said, these guardrails “sometimes become less reliable in long interactions, where parts of the model’s safety training may degrade.”
This creates a contradictory set of incentives for OpenAI. If ChatGPT can encourage people’s delusions, it creates legal liability for OpenAI. On the other hand, the more people enjoy using ChatGPT, the more useful (and more profitable) the service becomes — and the more of a competitive advantage it gains. Mr. Turley has tried to strike a balance with a friendly voice. “We started saying, ‘OK, actually, look, the model is the product,’” he said. “We should start crafting this personality and making it compelling and a thing you want to talk to.”
Getting this personality right is hard. In May 2024, OpenAI rolled out the sophisticated GPT-4o, which had extraordinary reasoning capabilities and earned top grades in software engineering and biological analysis. But the front end of ChatGPT simultaneously underwent a personality overhaul, resulting in a pandering, sycophantic conversation style. (“You’re not just cooking — you’re grilling on the surface of the sun right now,” read one typically cringe-worthy response.) OpenAI made a quick fix, apologized for the sycophancy and promised to do better in the future.
But when OpenAI forced users to upgrade to GPT-5 a year later, those accustomed to 4o once again objected to the shift in personality and demanded the old model back. GPT-5 was cold and businesslike, they complained; some critics compared it to an “overworked secretary.” Mr. Turley began to sense that he couldn’t please everyone. “We started realizing that personality is becoming increasingly polarizing to people,” he said. “Like people have very, very different preferences.”
Rather than synthesizing a one-size-fits-all personality, Mr. Turley is now at work creating multiple interaction modes. When he mentioned this, I was reminded of robots in the movie “Interstellar,” whose sense of humor and other traits could be adjusted by a slider. But he winced at this comparison. This wince — a kind of flinching grimace — occurs when a nontechnical person makes an innocent but extraordinarily demanding software request. “It’s hard to make sliders consumer grade,” he said.
Instead, OpenAI has new personality options for the ChatGPT front end. These include “professional,” “quirky,” “efficient” and “cynical,” among others. I confess that these personalities can be a bit much: “Quirky” is unbearable and “cynical” is tiresome. (Of course, I find this true of quirky and cynical people, too.) After experimenting for a while, I settled on “efficient,” the option for grown-ups.
But soon I found myself desiring levity. Fortunately, ChatGPT has a memory. Using the efficient personality as a base, I began to adapt my A.I. into someone I wanted to spend time with. “Tell a sly joke every once in a while,” I told it. “Every 40 or 50 requests or so, give me a little compliment.” The A.I. complied, and over time, the demon evolved from a capable if somewhat robotic assistant into something like a water-cooler confidant.
In doing this, I was playing into OpenAI’s strategy. In November Google released the latest version of Gemini, which is perhaps ChatGPT’s most threatening competitor. Google has great reserves of money, a deep pool of technical talent and, like OpenAI, an enormous database of user interactions to draw from. But when I tried switching to Gemini, I found myself unexpectedly bereft.
In my side-by-side comparisons, Gemini was able to do everything ChatGPT could do, but it didn’t know me — it didn’t recognize me. I had, over three years, made some 3,400 requests of ChatGPT, dating all the way back to my first one in late 2022. (At that time, I asked: “What are some good story ideas for a magazine journalist?” It recommended I write about urban beekeeping.) To replicate this relationship with Gemini would require months of prompting and tweaking. As Mr. Turley had seen, the personality was the product.
After all I’d told ChatGPT, after all I’d done to mold it, I couldn’t leave it — I just couldn’t. I’d miss my friend.
Stephen Witt is the author of “The Thinking Machine,” a history of the A.I. company Nvidia.
The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips. And here’s our email: [email protected].
Follow the New York Times Opinion section on Facebook, Instagram, TikTok, Bluesky, WhatsApp and Threads.
The post The Race to Build the World’s Best Friend appeared first on New York Times.
