Can Microsoft Copilot actually do the things in its Super Bowl ad?

In life, there will be people who tell you that you can’t do things like open a business or get a degree. But with Microsoft’s AI helper here to design your logos and quiz you in chemistry, you can now tell those people, “Watch me.”

At least that’s the premise of the Microsoft Copilot ad that aired in front of over a hundred million viewers during this year’s Super Bowl. The minute-long commercial backs up its claims with a montage of rapid-fire examples in which users tap it to move their dreams forward—presumably making their doubting haters seethe.

But can Microsoft’s new AI Copilot actually perform the tasks as they were demonstrated on television screens across America? Armed with a test subscription to the service, Tech Brew attempted to find out.

First, some quick background: Microsoft Copilot is the company’s latest packaging of it and OpenAI’s generative models for enterprise and consumer users. This iteration of the service, which debuted last November and opened to smaller businesses and consumers last month, weaves through the various programs of Microsoft’s productivity software suite like Word, Excel, PowerPoint, Outlook, and Teams. It also features a ChatGPT-style interface where users can give it instructions in conversational terms.

Microsoft is aiming to make Copilot ubiquitous in offices and homes in the same way it has its workplace software, and the Super Bowl ad was an attempt to frame how that could look. The ad seems to have piqued at least some viewers’ curiosity; it propelled the Copilot app to a top-three spot on the App Store the morning after it aired, according to NBC News.

In an effort to make AI appear accessible to a larger audience, Microsoft centers the ad on a series of “‘watch-me’ moments”—examples of the AI “enabling people to do things previously unattainable,” according to a company blog post.

“In the ad, we wanted to show a variety of great examples of what Copilot can do, and the first sentence/phrase of the actual prompt was shown,” Turnbaugh said. “All of the prompts have an ellipsis (…) to signify there was more context. The more specific your prompt is in Copilot, the better the results.”

Storyboarding scenarios

The first of these moments we see is an aspiring filmmaker asking Copilot to “generate storyboard images for the dragon scene in my script.” Typing that into Copilot indeed yields some storyboard-format art of dragons ravaging the countryside in the same visual style as the output in the ad.

But the user in question specifically asks about “my script,” implying she is trying to have the AI visualize a specific scene she’s written—not just any dragon content. So I next had Copilot generate a script snippet for a cinematic sequence involving a dragon lair and a knight seeking wisdom. Could the AI pump out a storyboard specifically for that?

The resulting images didn’t have much to do with the text of the scene. But if I zeroed in on the first bit of setting description—a dragon in a lair lit by lava, piled with treasure, and marked with ancient runes—it ably churned out that image. One could imagine building the rest of the storyboard bit by bit that way; it just might take a bit longer than the ad suggests.

Coding context

The next “‘watch-me’ moment” is a bit trickier—the user here asks the Copilot to “write code for my 3D open-world game.” Having never written code for a video game myself, I had a harder time judging this one.

Typing that simple prompt into Copilot produces a 57-word code snippet for the Unity platform, with an addendum that it might take more user guidance to produce any more. It doesn’t include any of the kind of detail shown in the ad, but again, we don’t really know what kind of previous back-and-forth the character had with the Copilot. Asking for next steps beyond the initial code snippet yields an extensive list of advice.

Chemistry quiz

This task was probably the most straightforward of the bunch. The next user asks Copilot to quiz her in organic chemistry. When I tried it, Copilot adeptly cooked up some questions similar to those in the ad. Given that I didn’t study chemistry beyond high school, I didn’t know the answers, but fact-checking them online seemed to confirm their accuracy.

Logo looniness

The final user shown in the ad is an entrepreneur looking to “design a sign for my classic truck repair garage Mike’s.” I was skeptical off the bat with this one given AI’s generally dismal track record with text in images. And as it turns out, those concerns were well-founded. Rather than the sleek signs shown in the ad, the ones I got were chock-full of garbled text.

A close analysis of the ad, though, seems to suggest that the garage founder (Mike himself?) may have taken a few steps in between that prompt and the results shown, the latter of which includes the prompt text: “A realistic image of an expansive truck repair shop interior for a…” I tried to finish the thought and generate something from that but it mostly just yielded some generic garage interiors.

Eventually, through some more description and finagling, I was able to get something closer to what appeared in the ad, a few warped letters notwithstanding.

That was pretty much the story with most of these examples: perhaps doable, but not without additional effort and prompting. While they do seem to be technically achievable through the platform, many of them definitely take more finesse and back-and-forth than the ad suggests, and many of the creative prompts would require substantial editing. One must hope that their doubters have a bit more patience before they’re proven wrong.