For the first time, ChatGPT is now capable of browsing the web. OpenAI rolled out new plugins for its popular chatbot, allowing it to source information from more sources. Currently, there’s a waitlist for the plugins as an alpha feature. OpenAI announced that initially, only a small number of developers and users with the ChatGPT Plus subscription would be prioritized.
The new plugins allow ChatGPT to access various third-party knowledge sources and databases.
However, it’s the web-browsing plugin that draws the most attention. So far, ChatGPT could only use the information it had been pre-trained on.
The introduction of the new first-party plugins gives ChatGPT a groundbreaking boost in its capabilities. Although the viral chatbot was trained on a vast array of resources from around the web, it was limited to information prior to 2021. This restricted its ability to answer questions accurately, and the chatbot often failed to provide the latest information.
The web-browsing plugin uses the Bing search API to retrieve content from the internet. Answers crafted from such information include citations for the sources. The plugins will be later released on a larger scale and with API access.
Potential Pitfalls
While the release of the web-browsing plugin is a huge upgrade to ChatGPT, it’s not without risks. Information curated from the web isn’t always accurate and unbiased.
Back in 2021, OpenAI built an experimental system known as WebGPT that used a text-based web browser to answer open-ended questions.
However, it was found to cherry-pick information from sources it believed to be more convincing for the users. That is, even if the sources weren’t the most reliable.
Meta came up with an AI chatbot with web access, too – BlenderBot 3.0. However, it was quickly disbanded due to similar issues. When prompted with certain queries, BlenderBot also presented answers with offensive content and conspiracy theories.
The dataset used to initially train the AI behind ChatGPT is far more curated and thus contains verified information.
The main reason why information gathered from the live web is less reliable than a static training dataset is that the former is less filtered.
AI chatbots sourcing information from the web must access it via search engines, such as Bing in the case of ChatGPT. In contrast, search engines have their own automated mechanisms to filter out unreliable content.
However, website owners can game these systems to push certain results to the top, even if they contain inaccurate content. For instance, Google’s algorithm prioritizes websites with modern web technologies, as a result of which credible sources often get buried under other results.
OpenAI’s Admission to Potential Misuse of the New Plugins
The San Francisco-based company has admitted that with access to the internet, ChatGPT might perform a host of undesirable behaviors. As stated by OpenAI, the web-browsing plugin might end up “increasing the capabilities of bad actors who would defraud, mislead or abuse others” by sending spam/fraudulent emails and bypassing safety restrictions.
However, OpenAI also assured that it had implemented several safeguards to prevent such activities. The effectiveness of these measures remains to be seen. ChatGPT has received several new plugins in recent times.
In addition to the ones released by OpenAI, early collaborators have also developed plugins like Zapier, Shopify, Slack, Instacart, etc.
The company also rolled out a Python code interpreter that works in a sandboxed and firewalled environment. The chatbot continues to grow in popularity with the constant addition of new capabilities.
Read More
Zonia Motsinger