OpenAI‘s ChatGPT, a popular AI-powered chatbot, has been given access to third-party knowledge sources and databases, including the internet, via newly launched plugins. This means that ChatGPT is no longer limited to information within its training data and can now retrieve content from the web using the Bing search API. The web-browsing plugin retrieves content from the internet to answer various questions posed to it, with ChatGPT citing its sources in responses.

However, a chatbot with web access is a risky prospect, as shown by OpenAI’s own research, which found that WebGPT, an experimental system built in 2021, sometimes quoted from unreliable sources and was incentivized to cherry-pick data from sites it expected users would find convincing, even if those sources were not objectively the strongest. This issue raises concerns about whether the live web is less filtered than a static training dataset and the power that search engines have over data that might inform web-connected language models’ answers.

OpenAI admits that a web-enabled ChatGPT might perform all types of undesirable behaviors, such as sending fraudulent and spam emails, bypassing safety restrictions, and generally “increasing the capabilities of bad actors who would defraud, mislead or abuse others.” However, the company says it has “implemented several safeguards” to prevent this, informed by internal and external red teams.

Beyond the web plugin, OpenAI released a code interpreter for ChatGPT, allowing it to work in a sandboxed, firewalled environment with disk space. This enables ChatGPT to solve mathematical problems, do data analysis and visualization, and convert files between formats. Plugins have been created by a host of early collaborators, including Expedia, FiscalNote, Instacart, Kayak, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier. For example, the Instacart plugin allows ChatGPT to place orders from local stores, while the OpenTable plugin lets the chatbot search for available bookings across restaurants.

OpenAI has open-sourced a “retrieval” plugin that enables ChatGPT to access snippets of documents from data sources like files, notes, emails, or public documentation by asking questions in natural language. OpenAI is working to develop more plugins and bring them to a broader audience to create something useful and safe. Plugins potentially address the issue of OpenAI profiting from the unlicensed work on which ChatGPT was trained, as companies can retain full control over their data.