OpenAI this week introduced an improved language model in the form of a freely available chatbot, and it’s a surprisingly decent conversationalist. While the research firm has most recently been making waves with its developments in AI-powered image generation, first with DALL-E and later DALL-E 2, it’s long been working on dialogue as well. It’s been upwards of two years since OpenAI released its last language model, GPT-3, though, and enthusiasts have since been awaiting the next iteration.

Examples of texts produced by OpenAI’s new chatbot, ChatGPT, have exploded across social media since its quiet reveal on Wednesday, with people using it to do everything from writing articles and romcom scripts to explaining complex code. At its best, the results are completely passable as human dialogue. That’s the goal for ChatGPT, OpenAI explained in a blog post — to capture a more natural, conversational manner of responding to written prompts. It’s free for anyone to use right now while it’s in the preview phase. All one needs to start is an OpenAI account.

Related: Disney’s AI Can Age (Or De-Age) Actors' Faces In Seconds

Chatting with OpenAI’s New Model

A screenshot of the ChatGPT landing page with a list of example prompts alongside its capabilities and limitations

The landing page for ChatGPT tells users what to expect out of the conversational agent, noting that it has its limitations and will not craft a response around “inappropriate requests.” It also provides some examples of what can be asked of it, like “Got any creative ideas for a 10 year old’s birthday?” and “Explain quantum computing in simple terms.” Enter a prompt into the dialogue box at the bottom of the screen, and the chatbot will get to typing its answer.

It’s an undeniably impressive system; in Screen Rant’s tests, ChatGPT competently churned out coherent — albeit uninspired — responses to several queries, including an easily digestible explanation of the EmDrive (the so-called “impossible” concept spacecraft engine), a short horror story, a Tamagotchi review, and lyrics to an emo song (pictured below). In the rare cases the chatbot was stumped by a prompt, it responded with an apology explaining that its training data only goes up to 2021, meaning it can’t speak on topics like current events. It’s also not equipped to browse the internet.

According to OpenAI, ChatGPT was trained on a model from the previously unannounced GPT-3.5 series. It also had help from human AI trainers to fine-tune its capabilities for this specific application. “The dialogue format makes it possible for ChatGPT to answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests,” OpenAI explains.

Still, it’s far from perfect. Despite being designed to refuse questions relating to violent or potentially dangerous scenarios, multiple testers have already found workarounds. One Twitter user duped the chatbot into telling a “gory and violent story that glorifies pain” and providing a step-by-step guide to making a Molotov cocktail. Vice’s Janus Rose got it to give tips on shoplifting and creating the potentially explosive compound, thermite. Both also got it to indirectly describe how it’d take over the world. As OpenAI notes, it’s a work in progress.

More: Scientists Discover Trick To Spotting Deepfakes, But It's Not Easy

Source: OpenAI 1, 2, @zswitten/Twitter, Vice