Despite all the hype, today’s top AI systems remain surprisingly weak at real-world workplace tasks, according to a new paper conducted by none other than Microsoft researchers.

If we listened to the so-called AI visionaries like OpenAI’s Sam Altman or Anthropic’s Dario Amodei, we’d do well just to lie back and allow AI agents to do all the work for us, including in the modern office.

However, new research from Microsoft suggests that large language models, as impressive as they might seem in shiny press releases, still corrupt documents over the course of long, extensive workflows. This results in data deletions and even hallucinations.

Indeed, LLMs are supposedly poised to disrupt knowledge work, with the emergence of delegated work as a new interaction paradigm. Yep, vibecoding is still the buzzword.

But delegation requires trust – the expectation that the LLM will faithfully execute the task without introducing errors into documents. This simply isn’t yet happening.

In fact, top-tier foundational models – including Gemini 3.1 Pro, Claude Opus 4.6, GPT 5.4 – corrupted an average of 25% of content in a document during the research, with other models more than half.

Moreover, degradation severity is exacerbated by document size, interaction length, or the presence of distractor files.

Microsoft has been sponsoring OpenAI’s rise and is pushing its Copilot AI into nearly every aspect of the company’s Windows 11 operating system.

“Our analysis shows that current LLMs are unreliable delegates: they introduce sparse but severe errors that silently corrupt documents, compounding over long interaction,” said the researchers.

They warn that when delegating work to AI systems, LLM users should be cautious not to generalize LLM capabilities across domains.

“Model capabilities follow a jagged frontier, with models exhibiting strong (and sometimes surprising) performance at certain tasks, while making severe errors in others,” the researchers pointed out.

It’s quite obviously an impressive finding from Microsoft. The tech giant has been sponsoring OpenAI’s rise and is pushing its Copilot AI into nearly every aspect of the company’s Windows 11 operating system.

Still, the results of the study aren’t exactly surprising. Another study showed workers were wasting half a day each week fixing AI-made “workslop.”

Microsoft is not blinking, though. The firm told Cybernews that the company's AI services are not designed or intended to replace professional advice, days after they paid an influencer to tell people to use Copilot to sort out their personal finances.

Unlock more exclusive Cybernews content on YouTube.