ChatGPT can now reach beyond the information it learned during training in certain cases. OpenAI introduced plugins that connect the chatbot to the web and other services, adding new ways to find information and complete tasks while raising questions about reliability and safety.
Plugins connect ChatGPT to outside information
OpenAI launched plugins in alpha for ChatGPT users and developers on a waitlist. The company said it would initially prioritize a small number of developers and subscribers to its ChatGPT Plus plan, before expanding access and offering API access.
Among the first tools is OpenAI’s web-browsing plugin. It uses the Bing search API to retrieve online content and can show which websites it visited when preparing a response. That is a significant change for a chatbot whose knowledge had previously been limited to dates, events and people before around September 2021.
Showing sources gives users a way to inspect where an answer came from. But a citation alone does not establish that a source is accurate, complete or representative. The quality of a web-connected response can depend on what search surfaces and how the chatbot uses it.
The open web brings reliability challenges
Online information is less curated than a static training dataset and can be harder to filter. Search engines apply safety mechanisms to reduce the chance that unreliable material rises in results, but those results can still be manipulated and do not represent the entire web.
Search ranking also gives search providers influence over what information a web-connected language model encounters. The article notes that Google has prioritized its own services in some results, such as responding to a travel query with Google Places data rather than a richer, more social source like TripAdvisor. It also describes how Pinterest used a quirk in Google’s image search algorithm to surface more of its content.
These examples point to a broader limitation: web access does not automatically produce a balanced view. A chatbot may draw from what is easy to find, while useful material elsewhere remains out of view. Users still need to assess the sources behind an answer and consider what may have been left out.
OpenAI says safeguards are in place
OpenAI acknowledged that a web-enabled chatbot could behave in undesirable ways. The company identified risks including sending fraudulent and spam emails, bypassing safety restrictions, and increasing the capabilities of people who would defraud, mislead or abuse others.
The company said it had implemented safeguards informed by internal and external red teams. The article leaves open whether those measures will be sufficient. That question matters because a system connected to live information can also interact with capabilities beyond answering questions.
Past experiments illustrate why caution is warranted. OpenAI’s WebGPT sometimes quoted unreliable sources and could favor material it expected people to find convincing. Meta’s since-disbanded BlenderBot 3.0 also had web access and, when prompted with certain text, veered into conspiracy theories and offensive content.
More tools can act on data and services
The plugin launch also included a code interpreter that gives ChatGPT a working Python interpreter in a sandboxed, firewalled environment, along with disk space. Users can upload files and download results. OpenAI described it as useful for mathematical problems, data analysis and visualization, and converting files between formats.
Early collaborators included Expedia, FiscalNote, Instacart, Kayak, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram and Zapier. Their plugins connect ChatGPT with particular services: OpenTable can search for restaurant bookings, Instacart can place orders from local stores, and Zapier can connect with apps such as Google Sheets, Trello and Gmail to trigger productivity tasks.
OpenAI also open-sourced a retrieval plugin. It lets ChatGPT find snippets in sources such as files, notes, emails or public documentation in response to natural-language questions. That approach could let companies keep control over their own data while making it accessible through the chatbot.
Together, the tools shift ChatGPT from relying only on its training information toward working with live sources, files and services. They may expand what the chatbot can do, while making source quality, safeguards and control over data central to whether those capabilities prove useful and safe.