Google’s AI articles abandon professional creators and the visually impaired

I edit support articles for Google as my main source of income. As Alphabet experimented with Content Sidekick (CS), its in-house AI for creating articles, the company did not let professional human writers or editors suggest changes for them. I was curious about how well CS’ results compared to human authors. I internally verified which articles were created by AI, and I found that CS made plenty of errors, including, but not limited to…

There were also a host of assorted errors, like…

This doesn’t even mention all the broken links CS invented, such as the ones found in these 3 articles. Google will use their AI to create Help Center advice going forward, so you should expect these issues to remain a constant feature.

“I’ve been trained to write super simple instructions that even your grandma could quickly understand. Every detail contributes to that experience, even if you might not immediately realize why you’re suddenly struggling to use Google products… When my team tried to present our findings to our managers, they told us that Google didn’t care. In fact, we noticed Google removing style guidelines to make it easier for the AI tool to score higher…  And yet, they’re increasing this tool’s use for over nine months. It hasn’t improved.”

Tahlia Kirk, Lead Trainer for the creators of Google’s support articles

All that said, those screenshots are only interesting to grammar nerds like myself. Let’s face it: I make typos too. Sure, a professional editor would’ve prevented CS’ errors, but is it worth it to nitpick details when such creations weren’t possible five years ago?

The thing about grammar, though, is that it doesn’t just smooth out the communication process for written language. Even on a computer, not everything gets read. People with visual impairments use screen readers to recite onscreen text. Those softwares read content out loud to users and ensure no one’s denied access to essential information. So how do screen readers translate an image into words? They can’t, so instead they read the descriptive texts associated with said pictures. Alt texts work for nonessential content, but screen readers often need actual words placed next to the image in question.

Due to corporate regulations, I can’t show you the policies or websites for Google employee instructions. But it’s not a secret that companies, Alphabet included, have rules about making content accessible for screen readers. It is federal law, after all. The teams that write Help Center support articles are based in California and Texas, and both states have separate regulations for accessible content. Screen readers can’t identify an icon if its name isn’t next to it. That means articles, in order to be accessible for all users, must list the icon name next to its icon, without exceptions.

Would you like to guess how often CS bungles this crucial step?

As Google automates the process of creating and editing support articles, it’s not just daily users and grammar pedants who will suffer. I recently interviewed Steve Elliot, a professional mediator based in Texas, on the subject of tech accessibility. The macular degeneration he endured from Stargardt’s disease inspired Elliot to become a disability rights attorney. He, like some of his clients, uses screen readers.

“Icons without alt text are useless,” Elliot told me during our conversation. Standard webpage guidelines require image descriptions in alt tags. Pictures without alt text, as interpreted by a screen reader, become a long string of characters and numbers that leave the listener totally lost, regardless of the webpage’s context. Situations like those open up opportunities for users to file lawsuits, DOJ complaints, or demand letters against the website’s owner. “Otherwise,” Elliot said, “you [users with vision impairments] don’t know what you’re scrolling across.”

Will Google’s AI improve to the point where it no longer makes this mistake? Perhaps. But CS didn’t find these errors on its own. Google’s AI won’t escalate this issue to itself. It took a team of humans to find this problem and understand its magnitude. Without my coworkers—diligent, human professionals who helped me create this article—billions of users would be locked out of Google product support with no avenue to fix the problem. Content Sidekick is a poor replacement for human writers that may save Google money in the short term but will cost it much more in errors, quality, and lawsuits in the future. I’m willing to let Google know about this danger in advance, but not every company is lucky enough to have me as an employee.

---

BONUS CONTENT: Here are the analysis data from publicly available articles, with the total errors found and their final quality scores:

• I counted the style guide errors in a random sample of articles written by CS. Green rows are good, the redder a row gets the worse the article’s quality is. The “Service Score,” as you can imagine, attempts to quantify how well an article aligns with our style guide. I’m not sure what parameters I’m allowed to reveal, but it wouldn’t surprise you that I check to see if, among other things…
  • The links work.
  • The icons help the visually impaired.
  • Our lists present information correctly.
  • The instructions make sense as written.
  • There are any basic spelling, grammar, and punctuation errors.
• The average CS Service Score was 82 out of 100. A ‘B-‘ doesn’t sound awful at first, but consider that, according to my coworkers, 99% of Google’s professional writers get a perfect service score on every article they write. Google’s AI would only achieve that goal 3% of the time, at best.
  • Let’s assume that 1 error every 100 words is considered acceptable. Even with that generous guideline, the AI would create unacceptable work 80% of the time.
• At my job, if a human’s Service Score got lower than 90, their supervisor would come talk to them about how to improve their performance. A professional author who needed retraining/discipline 5 or more times would get cut from this project. Going by the data, for every 4 articles CS creates, 3 of them would trigger manager involvement. Were it human, this machine would get fired before its oeuvre reached the double digits.

BONUS BONUS CONTENT: The following AI-written articles come from legally sensitive Help Centers, which makes their errors all the more important for Google to fix:

• Create, plan, organize & collaborate with Canvas in AI Mode 
• Gemini Apps Privacy Hub
• “AI Mode” in Search Labs 
• Share your AI-powered responses from Google Search
• How navigation data makes Maps better for everyone
• See places you might like in Maps