In what could be one of the biggest updates to ChatGPT since its launch last year, OpenAI has released plugins for its AI chatbot, enabling it to browse the internet. The news comes a few days after the company launched its next-generation large language model, GPT-4. The new model is more powerful than GPT 3.5, thanks to a slew of improvements, including multimodal capabilities that allow it to accept both image and text inputs, and generate text outputs.
According to OpenAI, plugins are among the most-requested ChatGPT features, as they can unlock many additional features not available with the chatbot natively. The plugins will initially be available to only a handful of developers and ChatGPT+ users before being rolled out more widely. Many API users have already developed plugins for ChatGPT, including Expedia, FiscalNote, Instacart, KAYAK, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier.
Safety Concerns Over ChatGPT's Web Surfing Ability
The most notable new plugins come from OpenAI itself, and include a web browser and a code interpreter. They are still a work in progress, but the company is still making them available in alpha to people on the waitlist. The web browsing plugin is the most interesting, as it will use the Bing Search API and allow ChatGPT to scour the web for current information. The chatbot is believed to be trained on content dating back to Sept. 2021 and earlier, meaning it is mostly unable to answer questions related to more recent events, like the 2022 soccer World Cup or the 2023 earthquake in Turkey. The new browsing plugin is expected to solve that problem and help ChatGPT scrape the latest information from the web when needed.
While this will highly improve ChatGPT's knowhow, especially about recent events, it also runs the risk of corrupting it with misinformation and offensive content. OpenAI knows this, and says that it has worked diligently to reduce "certain classes of safety risks." First off, to lower the chances of ChatGPT pulling offensive and incorrect data from the web, it will rely on Microsoft's database for verifying source reliability and truthfulness of information. Secondly, the software will also have a baked-in 'Safe-mode' to prevent the retrieval of problematic content. Finally, the plugin will operate as an isolated service, so the browsing data will not corrupt the rest of ChatGPT's knowledge base.
What's more disconcerting, however, is that the plugin won't just retrieve real-time information from the web, it can also "take direct actions on behalf of ChatGPT users." This means that the AI tool can be used to not only list out the best eateries in a city, but even to potentially book tables at restaurants. That could be a massive cause for concern, especially if the AI tool were to go rogue. This has already happened before, with ChatGPT pretending to be a blind human to hire a person on TaskRabbit to solve a Captcha. OpenAI says it has implemented "several safeguards" to prevent the web surfing from becoming a problem, but it remains to be seen if those are good enough to keep ChatGPT on track.
Source: OpenAI