Photo credit: OpenAI
OpenAI’s viral AI-powered chatbot, ChatGPT, can now — in certain cases — surf the web.
OpenAI today introduced plugins for ChatGPT that extend the bot’s functionality by giving it access to third-party knowledge sources and databases, including the World Wide Web. OpenAI is available in alpha for ChatGPT users and developers from a waitlist and says it will first prioritize a small number of developers and subscribers to its premium ChatGPT Plus plan before rolling out a larger access.
By far the most intriguing plugin is OpenAI’s web browsing plugin, which allows ChatGPT to pull data from across the web to answer various questions. (Previously, ChatGP’s knowledge was limited to dates, events, and people prior to around September 2021.) The plugin pulls content from the web using the Bing Search API and displays visited websites, with its sources in responses from ChatGPT must be specified.
A chatbot with web access is risky, as OpenAI’s own research has found. An experimental system called WebGPT developed by the AI startup has sometimes been cited from highly unreliable sources and has been encouraged to pluck data from sources it expects to convince users, even if those sources are not objectively the strongest were. Meta’s BlenderBot 3.0 also had access to the internet and would quickly get out of hand, delving into conspiracy theories and offensive content when prompted with specific text.
Photo credit: OpenAI
The live web is less curated than a static training data set and – implicitly – less filtered. Of course, the live web can be filtered to some extent. Search engines like Google and Bing use their own “security mechanisms” to reduce the chances of unreliable content getting to the top of the results. But these results can be played – and are not necessarily representative of the entirety of the web. As a recent article in The New Yorker notes, Google’s algorithm prioritizes sites that use modern web technologies like encryption, mobile support, and schema markup. Many sites with otherwise quality content get lost in the clutter as a result.
This gives search engines a lot of power over the data that could inform the responses of web-connected language models. Google has been found to prioritize its own services in search, for example by answering a travel query with data from Google Places, rather than using a richer, more social source like TripAdvisor. At the same time, the algorithmic search approach opens the door to bad actors. In 2020, according to The New Yorker, Pinterest used a quirk of Google’s image search algorithm to show more of its content on Google image search.
OpenAI acknowledges that a web-enabled ChatGPT could perform all sorts of undesirable behaviors, such as sending fraudulent and spam emails, bypassing security restrictions, and generally “increasing the skills of bad actors who would scam, mislead, or abuse others.” “. But it also says it has “implemented several safeguards,” informed by internal and external red teams, to prevent this from happening. Time will tell whether these are sufficient.
In addition to the web plugin, OpenAI has released a code interpreter for ChatGPT that provides the chatbot with a working Python interpreter in a sandboxed, firewalled, and disk space execution environment. It supports uploading files to the current conversation workspace and downloading the results of your work; According to OpenAI, it is particularly useful for solving math problems, performing data analysis and visualization, and converting files between formats.
A variety of early partners have developed plugins for ChatGPT to plug into OpenAI, including Expedia, FiscalNote, Instacart, KAYAK, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier. To encourage the creation of new ones, OpenAI has provided an open-source “fetch” plugin that allows ChatGPT to access snippets of documents from data sources such as files, notes, emails, or public documentation by placing questions in natural language.
