Why did Microsoft admit that AI is making us dumb?

By Parmy Olson

Microsoft Corp. plans to spend $80 billion on artificial intelligence this year, reinforcing its position as a leading vendor. So why did it recently publish a research paper showing an erosion of critical-thinking skills among workers using generative AI tools such as ChatGPT? If we were being generous, we could say it was genuine scientific enquiry. More likely, it wants to keep ahead of the curve as AI disrupts certain jobs, and ensure that its tools remain useful to businesses. At a time when Big Tech is racing to make AI models bigger, that’s a refreshingly thoughtful approach both to the industry’s business model and its social outcomes.

The study, carried out in conjunction with researchers at Carnegie Mellon University, surveyed 319 knowledge workers about how they used AI, including a teacher generating images for a presentation about hand-washing to her students using DALL-E 2 and a commodities trader generating strategies using ChatGPT.

The researchers found a striking pattern: The more participants trusted AI for certain tasks, the less they practiced those skills themselves, such as writing, analysis and critical evaluations. As a result, they self-reported an atrophying of skills in those areas. Several respondents said they started to doubt their abilities to perform tasks such as verifying grammar in text or composing legal letters, which led them to automatically accept whatever generative AI gave them.

And they were even less likely to practice their skills when there was time pressure. “In sales, I must reach a certain quota daily or risk losing my job,” one anonymized study participant said. “Ergo, I use AI to save time and don’t have much room to ponder over the result.”

A similar recent study by Anthropic, which looked at how people were using its AI model Claude, found that the top skill exhibited by the chatbot in conversations was “critical thinking.”

This paints the picture of a future where professional workers ultimately become managers of AI’s output, rather than the originators of new ideas and content, particularly as AI models get better. OpenAI’s latest “Deep Research” model, which costs $200 a month, can conduct research across the internet, scouring images, PDFs and text, to produce detailed reports with citations.  

One result is that cognitive work is going to transform, and quickly, according to a Feb. 12 note to investors from Deutsche Bank AG. “Humans will be rewarded for asking their AI agent the right questions, in the right way, and then using their judgment to assess and iterate on the answers,” research analyst Adrian Cox writes. “Much of the rest of the cognitive process will be offloaded.”

As frightening as that sounds, consider that Socrates once worried that writing would lead to the erosion of memory, that calculators were once expected to kill our mathematical skills and that GPS navigation would leave us hopelessly lost without our phones. That last one might be somewhat true, but by and large humans have managed to other uses for their brains when they outsource their thinking, even if our math and navigating skills become lazier.

What’s different with AI is that it encroaches on a much broader part of our everyday cognition. We’re put in positions to think critically far more often than we are to calculate sums or chart routes—whether crafting a sensitive email or deciding what to flag to our boss in a report. That could leave us less able to do core professional work, or more vulnerable to propaganda. And it leads back to the question of why Microsoft—which makes money from sales of OpenAI’s GPT models—published these findings.

There’s a clue in the report itself, where the authors note that they risk creating products “that do not address workers’ real needs,” if they don’t know how knowledge workers use AI, and how their brains work when they do. If a sales manager’s thinking skills go downhill when they use Microsoft’s AI products, the quality of their work might decline too.

A fascinating finding in Microsoft’s study was that the more people were confident in the abilities of their AI tool, the less likely they were to doublecheck its ouput . Given that AI still has a tendency to hallucinate, that raises the risk of poor-quality work. What happens when employers start noticing a decline in performance? They might blame it on the worker — but they might also blame it on the AI, which would be bad for Microsoft.

Tech companies have loudly marketed AI as a tool that will “augment” our intelligence, not replace it, as this study seems to suggest. So the lesson for Microsoft is in how it aims future products, not in making them more powerful but in somehow designing them to enhance rather than erode human capabilities. Perhaps, for instance, ChatGPT and its ilk can prod its users to come up with their own original thoughts once in a while. If they don’t, businesses could end up with workforces that can do more with less, but also can’t spot when their newfound efficiency is sending them in the wrong directions.