Windows Copilot AI Makes Your PC Feel Incompetent
▼ Summary
– Microsoft envisions AI-powered computers that understand natural language and perform tasks for users, as demonstrated in Copilot PC ads.
– The current reality of Copilot in Windows 11 is frustratingly inaccurate, with frequent errors, fabricated responses, and slow performance.
– Testing Copilot’s advertised capabilities revealed failures in identifying objects like microphones and rockets, plus incorrect location guidance for images.
– Copilot struggles with practical tasks, providing generic advice, misreading data, and lacking promised features like system control or reliable assistance.
– Despite Microsoft’s ambitious AI vision, Copilot currently functions as an incomplete solution that makes powerful computers seem incompetent.
The vision Microsoft is selling for its AI-powered future feels almost magical, a world where your computer understands your voice and effortlessly handles tasks on your behalf. Microsoft’s Copilot AI promises to transform how we interact with our devices, letting users simply speak to get things done. In advertisements, people chat with their laptops, which respond intelligently and take action, all under the tagline, “The computer you can talk to.” Yusuf Mehdi of Microsoft has described this as a future where your PC comprehends your requests and “magic happens” as a result. Even CEO Satya Nadella has shared an ambitious outlook, suggesting AI models could eventually operate computers as capably as humans, with Microsoft’s entire software ecosystem redesigned to support these advanced AI agents.
But the reality of using Copilot in Windows 11 today is far from magical. Instead of seamless assistance, engaging with the AI often becomes an exercise in frustration. During a week of testing, Copilot repeatedly delivered incorrect answers, invented details, and addressed the user in an oddly patronizing tone. Activating the Vision feature requires granting screen access every single time, much like joining a Teams call. Once enabled, responses are painfully slow, and the assistant insists on using your name with every interaction, creating an experience that feels both intrusive and inefficient.
Testing Copilot against the scenarios shown in Microsoft’s own ads reveals a stark gap between marketing and performance. In one commercial, Copilot Vision scans a YouTube video and correctly names a HyperX QuadCast 2S microphone. In practice, the assistant first offered generic facts about dynamic microphones, then began speaking as if the user were the person on screen, and finally misidentified the microphone as an earlier HyperX model. When asked where to buy the item, it provided a broken Amazon link and a working link to the wrong product at Best Buy.
Another advertisement shows Copilot identifying a Saturn V rocket in a PowerPoint slide and calculating its thrust. During testing, the AI failed to recognize the rocket even with “Saturn V” visible on screen. After being told what it was looking at, Copilot gave a thrust estimate but could not run simulations as shown in the ad, instead directing the user to Matlab. A third ad scenario involves identifying a watery cave location and explaining how to visit. While a longer version of the commercial correctly named Rio Secreto in Mexico, the shorter ad did not. When tested, Copilot’s answers were wildly inconsistent, sometimes offering directions in File Explorer, other times explaining how to open Google Chrome, and occasionally suggesting travel to Belize or the Cayman Islands, despite the actual location being in Mexico. Renaming the image file caused Copilot to confidently invent new locations, showing it relied on filenames rather than visual analysis.
Microsoft also presents Copilot as capable of performing tasks, such as turning a portfolio into a short biography. In the ad, it generates a sentence about an artist being inspired by their cat. When pointed to a real Instagram account, however, it produced cliché-ridden, meaningless text that failed to capture the person’s actual interests or mention their pets. Beyond ad-inspired tests, practical uses for Copilot Vision proved hard to find. It cannot yet toggle system settings like dark mode, and a Microsoft spokesperson clarified that local file actions are still experimental and not publicly available. In third-party apps like Adobe Lightroom, it offered generic tips delivered in a rapid, tutorial-like monologue. When asked to analyze a benchmark table in Google Sheets, it made calculation errors and misread clearly labeled scores, undermining trust in its accuracy.
Even in gaming, an area Microsoft promotes for Copilot, the assistance was superficial. For Hollow Knight: Silksong, it gave shallow advice resembling a book report written from glancing at the cover. In Balatro, it misidentified cards in hand and offered irrelevant details about other games. While there is potential value, especially for accessibility, once Copilot can fully control Windows, the current experience makes even powerful PCs seem clumsy. Generative AI like Copilot often feels like a solution in search of a problem, and the distance between Microsoft’s visionary promises and the tool’s actual performance is vast. For now, talking to Copilot highlights how much work remains before AI can deliver on the futuristic computing experience being advertised.
(Source: The Verge)
