SEO News You Can Use: Could Natural Language Models Like BERT Be Inherently Biased?

It’s inevitable that, with advancements in artificial intelligence (AI), come a host of new and unprecedented problems. One of them is bias.

In a fascinating Search Engine Land article, George Nguyen delves into the potential issues with BERT, Google’s natural language processing model that helps computers understand language in a similar way humans do.

But, with a respected AI researcher exiting Google – or being let go, depending on who you ask – amid an internal dispute, there may be more to the story.

AI researcher Timnit Gebru, a diversity advocate and Google’s first Black female researcher, recently co-authored a paper about biased language technology. No stranger to academic research, Gebru is the woman behind the widely shared 2018 study that found facial-analysis software shows an error rate of 34.7 percent for women of color and less than 1 percent for light-skinned men.

In her most recent paper, which is not yet publicly available, Gebru posits that language models like BERT, which are trained on large datasets, pose a number of risks. Among a vast array of potential issues, she argues that language models trained on existing data are likely to contain racist, sexist or otherwise bigoted language.

As it stands currently, this isn’t just a risk, but a reality: The article includes a visual example of the terms linked to the word “engineer” being first names most commonly associated with English-speaking men.

While this could have a negative effect on society at large, it also impacts marketing. If biased models are incorporated into future search algorithms, the search engine marketing industry will require optimization for models built on prejudice – and that paints a very bleak picture. The fact that Gebru argues her departure from Google was not her choice only adds to the concern. If BERT could indeed reinforce biases in Google Search and Google just lost the diversity-driven co-lead of its Ethical Artificial Intelligence team, what does that mean for the future of SEO?

More SEO News You Can Use

Passage Indexing Will Look Identical to Regular Search Result Snippets: With Google’s recent announcement of passage indexing (which, as we now know, is more of a “ranking thing”  than anything else), many SEOs assumed passage indexing would somehow look different in Google Search results – but that’s not the case. The confusion likely stems from the screenshot Google shared in its original announcement, which compared a regular snippet to a featured snippet. Google’s Danny Sullivan has written it off as a “bad illustration,” reiterating that passage indexing is purely about ranking. The only thing we can expect to change is that Google will gain a deeper understanding of a web page’s content and rank pages that previously went unnoticed.

The Nosnippet Tag Won’t Prevent Passages From Ranking: Here’s yet more evidence that passage indexing is to do with ranking, not featured snippets. Danny Sullivan said on Twitter that using the nosnippet tag won’t prevent your pages or passages from being ranked. The question arose because the nosnippet tag prevents pages from showing up as featured snippets, but Sullivan said that passage indexing is not a snippet – it’s just another search result. Google’s John Mueller also joined the conversation, calling the nosnippet tag “purely a snippet control mechanism.” If you don’t want your page to rank, as always, just block Google from crawling and ranking it.

The Washington Post Thinks Google Is Killing Local News: A recent Nature Human Behavior study, conducted for the Washington Post, has reached the worrisome conclusion that Google News directs readers – and much-needed advertising dollars – away from local, community outlets and towards national news outlets. The study found that, even for inherently community-driven search queries like “mayor” and “school board,” local news outlets rarely make the first page of Google News results. The researchers argue that Google’s algorithms favor large, well-funded national news outlets over local outlets reporting on the same stories – and local media, a necessity for a healthy democracy, is suffering.

Follow Moz’s Months-Long Foray Into Outreach-Based Link Building: It’s no secret that outreach-based link building is a soul-destroying undertaking that can bring little benefit. Moz agrees. That’s why, over the next few months, Russ Jones is offering readers a window into this “brutal slog,” complete with tips and tricks to make the mission worth it. Jones will update this diary with weekly progress reports tracking wins, losses and motivation level. It’s an interesting project that should provide plenty of useful information over the next few weeks and months, so be sure to check in every now and then for updates and insights. Who knows? It could make outreach-based link building feel a little less futile.

Google Is Testing a Feature That Gives Instagram and TikTok Videos Their Own Dedicated Carousel: It’s looking like, pretty soon, there won’t be a single reason to leave the Google app. Barry Schwartz of Search Engine Roundtable reported that Google’s latest test shows a “Short videos” carousel on mobile search engine results pages (SERPs) featuring related TikTok and Instagram videos. Users can watch the videos without going into the apps. Something similar to this test has been seen before, back in April 2020, but the initial carousel only showed YouTube videos. It should be interesting to see if anything will come of this, or if it’s just another test we’ll all have forgotten about by February.

Editor’s Note: “SEO News You Can Use” is a weekly blog post posted every Monday morning only on SEOblog.com, rounding up all the top SEO news from around the world. Our goal is to make SEOblog.com a one-stop-shop for everyone looking for SEO news, education and for hiring an SEO expert with our comprehensive SEO agency directory.