Artificial Intelligence and Machine Learning

Giving Users More than They Can Handle

Earlier this year, Microsoft incorporated artificial intelligence tools into its suite of Microsoft now creates performance issues.

In February 2023, Microsoft customers began complaining in a Microsoft Answers forum thread that new artificial intelligence (AI) files showing up in Microsoft Windows were slowing or crashing their computers ,and making some tasks difficult to impossible.

User "Denise Heap" posted: "Windows just crashed on me as I attempted to convert a one-page (!) [sic] doc to PDF. When I pulled up Task Manager, it was at 100% CPU, 90% of which were five (!) [sic] instances of AI.EXE."

Microsoft Word can convert the documents it creates into PDF files.

Regarding AI.exe, user "Giuseppe Carafa" posted: "We have started getting complaints from our customers over this "feature" as it appears to be quite the resource hog, and of course, they are blaming our add-ins."

User "Giuseppe Carafa" continues: "I need to know how we can disable it. I'd prefer to give an instruction to our users on how they can switch off a feature they do not want, they feel they do not need, and that causes their old PCs to start sounding like a Harrier Jet taking off."

Users, including "Giuseppe Carafa," complained they have trouble permanently deleting or disabling AI files such as AI.exe. The other AI files include aimgr.exe, mlg.dll, ai.dll, and aitrx.dll. User "Chandy – MSFT Microsoft Agent | Moderator" posted that future updates will probably reinstall AI.exe and that there is no current option to turn AI off on Microsoft Windows.

According to the user "Joshua Burkholder (MSFT) Microsoft Employee," who posted in a related Microsoft Answers thread, the functionality in AI.exe isn't new. The Microsoft AI in reference here once lived inside individual applications like Word and Outlook, Microsoft Office applications that are installed on the desktop computer.

Now AI lives outside the applications, according to the "Burkholder" post. One purpose of the change was to address memory limits on these x86 Office desktop applications. As the post continues, moving AI allows Office applications to use memory for other tasks.

"Now this separate executable can consume more resources," said Chris Pratt, a former Microsoft product manager now CEO and Founder of Silatus, Inc., an AI research company.

While it was a great way to get around those memory limits per application, they ended up causing problems for many people as a result, according to Pratt, referring to Microsoft.

One expert shared a big-picture explanation for changes to Microsoft software, such as added AI files.

"Microsoft must keep up with new mobile technologies," said Richard Luna, CEO of Protected Harbor, an managed service provider (MSP) managing applications and services provided by Microsoft.

"Older technologies run on physical devices. Recent mobile devices are really cloud relay devices where the application processing takes place in the cloud and not on the device itself," Luna said.

So, when Microsoft adds an AI capability in the cloud, additional processing happens in the cloud. But when adding AI capabilities to older computers where the Microsoft Office applications are installed and operating on the device, added processing occurs on the device, which may not have adequate resources.

"It means Microsoft must either redesign Windows to keep up and break support for the millions of previous devices and programs, or glue one new functionality after another onto the existing structure until it all falls apart, said Luna.

There is the question of cybersecurity regarding the availability of systems and data with slowing, crashing systems and applications. The CIA triad, a model for cybersecurity, stands for the Confidentiality, Integrity, and Availability of systems and data, which requires adequate security to maintain. Systems and data must be available to keep the organization going.

According to Nizel Adams, CEO and principal engineer of Nizel Co., a Microsoft partner, "Availability would only be a concern if a user has specific access to information and needs it to do their job (i.e., a payroll specialist needing to use their machine to access a payroll database server that is secured with role-based access)."

"If the issue only affects a handful of users, there are redundant functioning workstations, or the user can ask a co-worker for temporary assistance," said Adams. So, a user can use another computer workstation that is not experiencing these issues.

"Microsoft licenses Office out to a minimum of 345 million clients, versus the handful of people in that particular thread, so if there were genuine issues, then we would notice," said Adams.

A Microsoft spokesperson (whom the company declined to name) commenting on system load concerns with AI.exe said, "There should not be any significant load on system resources for the pre-trained models that we currently ship."

A developer page from Google provides some insight into the training of AI models.

The Microsoft spokesperson continued, "If customers are experiencing a significant load on system resources using a computer that meets or exceeds Office's minimum specifications, please reach out to customer support so that they can get more information and investigate." The spokesperson provided a link to customer support at Microsoft 365 for business.

According to Pratt, "AI.exe runs various AI models. It's not clear which models it runs." Pratt suggested the models could include background blurring for MS Teams, predictive text, or Copilot (when/where available) for Outlook and Word.

Pratt recalls his time at Microsoft working on the resource consumption of AI models. "When I was at Microsoft, we found that seemingly simple AI models—like the background blurring model for Teams—could consume far too many resources, depending on the resolution of the input (camera) and specifications of the host machine," Pratt said.

"Microsoft has made considerable efforts to optimize for these resource constraints, but newer models –like Copilot models—will likely be more resource intensive until the data science teams have time to optimize them. A tradeoff decision is made for every AI model between moving to market quickly and being resource efficient," Pratt explains.


David Geer is a journalist who focuses on issues related to cybersecurity. He writes from Cleveland, OH, USA.

Join the Discussion (0)

Become a Member or Sign In to Post a Comment

The Latest from CACM

Shape the Future of Computing

ACM encourages its members to take a direct hand in shaping the future of the association. There are more ways than ever to get involved.

Get Involved

Communications of the ACM (CACM) is now a fully Open Access publication.

By opening CACM to the world, we hope to increase engagement among the broader computer science community and encourage non-members to discover the rich resources ACM has to offer.

Learn More