A few weeks ago, Microsoft renamed its AI powered search engine, Bing Copilot to just Copilot, which can now be accessed by all MSU accounts on the Microsoft 365 browser.
“To improve the user experience and streamline our tools that empower creativity, Bing Image Creator is now Designer and Bing Chat is now Copilot. Create wow-worthy images with your words and AI with Designer, and try Copilot, your AI-powered search assistant for the web.”
Besides making fun of the constant nomenclature issues that generative AI applications seem to face, I want to walk through how our access to Copilot works, what it can be used for, and how it holds up to other AI tools.
MSU Access to Copilot
Instructors, staff, and students can access Copilot by signing on to Microsoft 365 on the browser. Once logged in using your single sign on credentials, you can select “more apps” to find the Copilot app, or visit https://copilot.cloud.microsoft/
As of today, we do not have access to Copilot features in other Microsoft applications, like Word or Excel.
We are limited to the use of (Bing) Copilot, which, as I’ll explain shortly, is not as advanced as other AI tools.
In terms of security and data privacy, Copilot is covered under the same enterprise data protection as the rest of our Microsoft subscription. And while the university advises against putting sensitive data into AI tools, Copilot falls under Microsoft’s FERPA compliance, so it can handle student data.
Copilot is not “training” on any of your data, because it’s not an LLM, however Microsoft can use anonymized data for product design and improvements, so it’s a grey area when AI tools say they are not using your data for training. Copilot even asked me if I wanted to give it feedback:
Using Copilot
Copilot functions much like other generative AI chats: You can prompt it for information, ask it to solve problems, and get feedback on your work. At the end of each response, the app includes footnotes with links to websites where it sourced the information from, and suggested follow up prompts.
There is a 30-message limit per conversation, which is probably to optimize the context window. Copilot also provides access to Designer, an image generator. I was most impressed with the image generator, which provides two options for generated images to use from your prompt.
Copilot Vs Other AI Tools
I find Copilot to be less robust than a Large Language Model like ChatGPT or Claude. One of the main issues is that Copilot is actually Microsoft’s integration of its search engine, Bing, with an API connection to an LLM.*
When you use Copilot, you are receiving responses that combine real-time search results from Bing with text generation from an LLM. This layering of search and AI-generated content provides a mix of up-to-date web data and language model outputs. There are also layers of guardrails that actively work to prevent what Microsoft might deem as misuse. And those guardrails don’t always work the way we expect.
Moreover, Bing itself has only 5% of US Search traffic, making its results worse than other engines, like Google. And the converse is true: when you ask ChatGPT 4.o to search the web, it uses an API of Bing to do so. I’ve noticed that ChatGPT is especially bad at providing relevant search results, because Bing is bad at providing relevant search results. So, when you use Copilot, you’re getting a watered-down version of an old LLM, plus the mediocre search results of Bing, and a layer of undisclosed guardrails that majorly limit usage.
So, how does Copilot fair against other AI tools?
Badly. On the one hand, if you are looking to introduce students to AI tools, they already have access to Copilot, so it is less of a barrier. But on the other hand, an LLM tool like ChatGPT or Claude can provide a broader range of capabilities, including complex problem-solving, deeper contextual understanding, and the ability to engage in more nuanced conversations.
If you’ve been using Copilot, or have other insights to share, as always, I’d love to hear from you.
Freyesaur out.
* While it’s generally understood that OpenAI’s GPT-4 architecture powers the system, the exact LLM and the parameters governing its use are not always transparent. In the past, Microsoft advertised that Copilot in “Creative Mode” used ChatGPT 4.0. However, the new Copilot does not have the “Creative Mode” option, and I couldn’t find any documentation that it still indeed uses ChatGPT 4.0.