Microsoft Copilot scores low on AI IQ tests — but that’s not the full story

by

Not all brains are created equally, and that is equally true of artificial intelligence.

While major tech firms like Apple and Meta are frantically seeking ways to capitalize on the surge in artificial intelligence, a familiar name is quietly plotting its moves off the main stage.

In contrast, while Meta gives away large sums as bonuses to attract OpenAI researchers and Apple unsettles investors due to an unclear strategy in AI, Microsoft appears rather satisfied with providing the underlying support for many of these services.

As an analyst, I find myself grappling with the news that Microsoft is reportedly downsizing its workforce by over 15,000 employees in 2025. This move, it seems, is to amass funds for a significant investment in AI-focused data centers, as part of an ambitious strategy to bolster Azure’s capabilities and power AI solutions for other companies.

As an observer, I note that Microsoft, indeed, nurtures its own AI initiatives. For instance, Microsoft Copilot serves as their response to AI assistants like ChatGPT. Moreover, they’ve infused “AI” capabilities into applications such as the Photos app, Microsoft Paint, and even Notepad. However, it appears that these advancements have not yet captured the public’s attention, which might be one reason for this situation.

I’ve come across a new site called TrackingAI, which pits various models against each other in a range of intellectually demanding competitions. This platform subjects AI Large Language Models to tests like the challenging reasoning exams from Mensa Norway, as well as offline tests to ensure the AI isn’t searching the web for answers. So, how did Microsoft Copilot fare? Not too impressively (at least based on the results).

In offline trials, Microsoft’s Copilot performed poorly, scoring only 67, while OpenAI’s o3 Pro led the pack with a score of 117. Copilot showed some improvement in the Mensa Norway test, achieving 84, but Elon Musk’s Grok-4 excelled with a score of 136. OpenAI’s o3 Pro came very close, scoring 135.

It’s worth mentioning that Microsoft Copilot operates on GPT-4o, a model that emphasizes adaptability, speed, and affordability over logical reasoning. On the other hand, OpenAI’s o3 models aren’t commonly accessible due to their significantly higher operational costs compared to GPT-4o. Many models that outperform Microsoft Copilot are pricier “pro” versions. However, Copilot is free, and its performance typically aligns with its cost-free status.

One of Microsoft’s primary research focuses is finding ways to enhance powerful models while keeping them affordable, as the company has seen a significant increase in carbon emissions mainly due to AI energy consumption. Their Phi models, which are not yet widely accessible, emphasize performance and efficiency, making them suitable for on-device operation at minimal costs. Unfortunately, these Phi models have not been listed on TrackingAI as of now, according to available information.

Today, each language model, such as GPT-4o and Copilot, is tailored for specific tasks rather than being all-purpose. They are designed to be user-friendly and enjoyable, although they may struggle with academic subjects compared to their raw capabilities. Copilot offers deeper research modes that increase its accuracy if you choose to subscribe.

Currently, there’s a lot of excitement surrounding artificial intelligence (AI), with companies like Google Gemini, X’s Grok, and OpenAI’s ChatGPT often outdoing each other in specific areas. However, it’s unfortunate that Microsoft, despite heavy investments, doesn’t often get mentioned positively in these discussions, except perhaps when discussing the failure of Copilot+ PCs or the privacy concerns surrounding Windows Recall.

It’s possible that Microsoft finds satisfaction in staying behind the scenes, fueling the future through Azure rather than seeking the spotlight.

Read More

2025-07-14 22:39