Skip to main content

Discord starts testing ChatGPT-powered Clyde chatbot and other AI features

Discord starts testing ChatGPT-powered Clyde chatbot and other AI features

/

Discord is using OpenAI’s technology to improve its Clyde bot, moderation tools, and platform features.

Share this story

Illustration of Discord logo
Illustration: Alex Castro / The Verge

Discord is now using OpenAI’s ChatGPT technology to transform its existing Clyde bot into a talkative chatbot. Clyde is being upgraded next week to answer questions and have conversations with users, much like OpenAI’s ChatGPT or Microsoft’s Bing chat feature. It’s part of a broader push for AI in Discord, which also includes AI-generated conversation summaries and the ability for Discord admins to leverage AI technology to moderate servers.

The Clyde chatbot will be free for Discord users during a public experiment in a very limited number of servers with Discord alpha users next week. Discord admins will eventually be able to add the Clyde chatbot to servers to allow users to summon it into conversations and have it respond to queries to send GIFs to a channel, recommend music, or much more.

The AI-powered Clyde chatbot responding in Discord.
The AI-powered Clyde chatbot responding in Discord.
Image: Discord

Discord users can direct message Clyde to ask questions, and the chatbot can even create new threads in channels to facilitate conversations between groups of friends. Unlike the ChatGPT integration in Slack, you won’t be able to use Clyde to talk to your friends for you by drafting messages. Think of it as more of a chatbot you summon when you’re arguing about the most popular songs with friends, discussing how many decades it has been since Tottenham Hotspur won a major trophy, or trying to find out useful facts like the time zone your friend is in.

Discord plans to make this a fundamental part of its chat and communities app, but there’s lots of testing, learning, and iterating ahead before it’s ready to roll out to all Discord users. Like many previous Discord experiments, the upgraded version of Clyde will be available to a limited set of alpha testers initially before the chatbot is made available more broadly. Discord wouldn’t confirm the number of users or servers getting initial access.

Discord is starting off slow, with a limited number of users getting chatbot access

“Clyde has been alive on Discord for years and years, and we’ve always been interested in figuring out how to make interacting with Clyde smarter, better, and more fun,” says Anjney Midha, head of Discord’s platform ecosystem, in an interview with The Verge. “You often have a vision of how to do that from a product perspective, but you often have to wait for technology to catch up to deliver what your expectations are.”

Midha and the Discord team are confident large language models have caught up to their vision over the past six months amid the emergence of ChatGPT and Microsoft launching its AI-powered Bing. “Really in the last 90 days it became clear that the large language models... the base of deployment has picked up where these are now ready to have really engaging and fun interactions with consumers and not just researchers,” says Midha.

It’s a risky bet, though. Clyde is currently based on the same technology that powers ChatGPT. We’ve already seen ChatGPT hallucinate and generate false information, and Microsoft’s new Bing chatbot has shown what happens when you roll out this emerging technology to the masses: wild and rude responses. Discord’s chatbot could easily do the same. Discord is fully aware of this and even has a warning when you opt in to use Clyde for the first time:

Don’t tell Clyde secrets or rely on it for advice, Discord support, or safety issues. Clyde is experimental and not perfect so you might come across information that could be considered biased, misleading, harmful, or inaccurate even with safeguards in place.

Beyond the dangers of false information, the presence of an AI-powered Discord chatbot in servers that often include very tech savvy and younger users could present some challenges to Discord’s AI ambitions. There will be many who will ignore the warning and attempt to jailbreak the Discord chatbot to say bad things.

“We watch what’s happening with open models and I’m very familiar with people trying to jailbreak them and trying to induce models to say things that might be unsavory or break terms of service,” says Midha. “The only real ways to protect against those is to build with your real users, so we’re starting off slow by rolling it out to some alpha folks and asking them for feedback that allows us to tweak and tune the models.”

Midha also says that “no user data that is collected or surfaced to Clyde is being used to train any models or improve models by OpenAI.”

New AI conversation summaries in Discord.
New AI conversation summaries in Discord.
Image: Discord

Discord is also experimenting with AI-generated conversation summaries. It’s an opt-in experience within a server that’s not turned on by default and will allow Discord users to catch up on conversations they might have missed. This makes particular sense for Discord, where tight-knit communities keep conversations rolling endlessly across time zones, making it harder — but also more important — for users to catch up with chatter.

Where the power of AI could get interesting for Discord admins is with an upgrade to AutoMod. Discord rolled out its autonomous moderation tool widely last year to fight spam and slurs, and now, it’s experimenting with an AI-powered version of AutoMod.

“I’ve never seen anything like it.”

While Discord has been using a variety of large-scale deep learning models for AutoMod previously, it’s now planning to use language models from OpenAI to allow AutoMod to more accurately process the context of a message and capture where there’s clear intent to break server rules. “The large language models are unbelievably effective at inferring the context and intent of conversations,” says Midha. “I’ve never seen anything like it.”

AutoMod has already blocked more than 45 million unwanted messages based on Discord server rules, and AutoMod AI will be available in a limited number of servers starting today.

Discord’s prototype whiteboard feature.
Discord’s prototype whiteboard feature.
Image: Discord

Discord is also experimenting with some AI-powered prototype features that might eventually launch in the future. The company has built a whiteboard for Discord, which has been a highly requested feature. “We’re exploring a shared visual space to collaborate with friends and colleagues that includes an AI-powered text-to-image generator you can iterate and experiment with together,” says Midha in a blog post. It’s not clear when this feature might launch more broadly, though.

There’s also an avatar remix app that lets Discord users remix each other’s avatars using generative image models. Discord has made this app open source on GitHub, allowing users to fork, remix, and extend it freely using whichever models they want.

Discord is also trying to position its platform as the home of AI developers by dedicating its own ecosystem fund and resources to AI incubator programs that are designed to help developers build AI apps for Discord. Cash grants are available to support development, alongside access to discord development teams, cloud compute credits, and even early access to Discord platform features.

Discord is already home to many AI developers, including Midjourney, which has created an AI image generator accessed through its own Discord chat server. Midjourney has more than 13 million members in its Discord server, and it’s a big part of the rise of AI image generators we’ve seen over the past year.