OpenAI Unleashes ChatGPT, a Conversational Model That Admits When It's Wrong

OpenAI has dropped ChatGPT, a new conversational model that’s free to use during its research preview. It’s a sibling to the existing InstructGPT, but the dialogue format is the star here. You can ask it follow-up questions, and the company says the model will push back on incorrect premises, admit its own mistakes, and even reject inappropriate requests. The whole point of this public release, which you can access at chatgpt.com, is to gather user feedback and suss out the model’s real-world strengths and weaknesses.

The training recipe is where things get interesting. OpenAI used Reinforcement Learning from Human Feedback (RLHF), the same core method behind InstructGPT, but with a conversational twist. Human trainers played both sides of a chat—user and AI assistant—to generate a new dialogue dataset. The team then built a reward model by having those trainers rank different AI-generated responses, fine-tuning the whole thing with Proximal Policy Optimization. Under the hood, it’s a model from the GPT‑3.5 series, trained on Azure’s AI supercomputing infrastructure.

OpenAI is refreshingly blunt about the model’s flaws. ChatGPT will confidently spout nonsense, an issue they blame partly on the lack of a single source of truth during reinforcement learning. Making it more cautious backfires too, causing it to clam up on questions it could actually answer. The system is also hypersensitive to how you phrase a prompt—one wording gets a refusal, a slight rephrase gets the correct answer. It’s overly verbose, a quirk inherited from human trainers who preferred longer, seemingly more thorough responses. And rather than asking for clarification on a vague query, the model usually just takes a wild guess at what you meant.

Safety mitigations from earlier deployments like GPT‑3 and Codex are baked in, and the company is using its Moderation API to flag unsafe content, though they admit it’s still prone to errors. Despite the laundry list of limitations, the iterative deployment strategy is all about gathering data to make the system safer and more useful. The real test isn’t the slick demo samples—it’s what happens when millions of users start trying to trip it up. That feedback loop is what will determine if this thing gets genuinely better, or just better at sounding confident.