OpenAI’s viral AI-powered chatbot, ChatGPT, can now surf the web – in certain cases.
Open AI today launched plugins for ChatGPT, which extend the bot’s functionality by giving it access to third-party knowledge sources and databases, including the web. Available in alpha for ChatGPT users and developers on the waiting listOpenAI says it will initially prioritize a small number of developers and subscribers of its premium ChatGPT Plus plan before rolling out large-scale and API access.
The most intriguing plugin is definitely OpenAI’s first-party web browsing plugin, which allows ChatGPT to pull data from anywhere on the internet to answer the various questions. (Previously, ChatGPT’s knowledge was limited to dates, events, and people prior to approximately September 2021.) The plugin pulls content from the web using the Bing search API and shows all visited websites when composing an answer, citing the sources in ChatGPT’s answers.
A web-enabled chatbot is a risky prospect, as OpenAI’s own research has found. An experimental system built in 2021 by the AI startup called WebGPT, sometimes quoted from untrustworthy sources, and was incentivized to select data from sites it expected users would find convincing – even if those sources weren’t objectively the strongest. Meta’s now defunct BlenderBot 3.0 also had internet access, and fast went off the railsdelve into conspiracy theories and offensive content when prompted with certain text.
Image Credits: Open AI
The live web is less curated than a static training dataset and – implicitly – of course less filtered. Search engines such as Google and Bing use their own safety mechanisms to reduce the chance of untrustworthy content appearing at the top of their results, but these results can be exploited. Nor are they necessarily representative of the totality of the web. As a piece in The New Yorker points out, Google’s algorithm prioritizes websites that use modern web technologies, such as encryption, mobile support, and schema format. Many websites with otherwise high-quality content get lost in the shuffle as a result.
This gives search engines a lot of power over the data that can inform the answers of Internet-connected language models. Google has been found to prioritize its own services in Search by, for example to answer a travel question using data from Google Places rather than a richer, more social source like TripAdvisor. At the same time, the algorithmic approach to search opens the door for bad actors. In 2020, Pinterest took advantage of a quirk of Google’s image search algorithm to retrieve more of its content in Google Image Search, according to to The New Yorker.
OpenAI admits that a web-enabled ChatGPT can exhibit all kinds of unwanted behavior, such as sending fraudulent and spam emails, bypassing security restrictions, and generally “increasing the capabilities of bad actors who would defraud others, mislead or misuse”. But the company also says it has “implemented various security measures” informed by internal and external red teams to prevent this from happening. Time will tell if they are enough.
In addition to the web plugin, OpenAI has released a code interpreter for ChatGPT that provides the chatbot with a working Python interpreter in an environment with a sandbox and a firewall along with disk space. It supports upload files to ChatGPT and download the results; OpenAI says it’s especially useful for solving math problems, performing data analysis and visualization, and converting files between formats.
Image Credits: Open AI
A host of early contributors built plugins for ChatGPT to join OpenAI, including Expedia, FiscalNote, Instacart, Kayak, Klarna, Milo, OpenTable, Shopify, Slack, Speak, Wolfram, and Zapier.
They largely speak for themselves. For example, the OpenTable plugin allows the chatbot to search restaurants for available bookings, while the Instacart plugin allows ChatGPT to place orders at local stores. Zapier is by far the most extensible of the bunch, connecting to apps like Google Sheets, Trello, and Gmail to launch a range of productivity tasks.
To facilitate the creation of new plugins, OpenAI has open sourced a “fetch” plugin that allows ChatGPT to access snippets of documents from data sources such as files, notes, emails or public documentation by asking questions in natural language.
“We are working to develop plugins and bring them to a wider audience,” OpenAI wrote in a blog post. “We still have a lot to learn and with everyone’s help we hope to build something that is both useful and safe.”
Plugins are a curious addition to ChatGPT’s development timeline. Once limited to the information within the training data, ChatGPT, with plugins, is suddenly much more capable – and perhaps with less legal risk. Some experts accuse OpenAI of profiting from the unlicensed work on which ChatGPT is trained; ChatGPT’s dataset includes a wide variety of public websites. But plugins may address that problem by letting companies maintain full control over their data.
