You Autocomplete Me

Could Web-search algorithms make spelling a lost art?

Art for autocorrect.

Robert Neubecker.

Want to know more about Brittany Griner, the 6-foot-8 dunking phenom who helped Baylor’s women’s basketball team win a national championship on Tuesday night? How about the war in Afganistan, the morgage crisis, and American Idle? Go ahead and Google it.

“Did you mean: American Idol?” Google asks. Oops. Sure did. With each of the other misspellings above, Google is so confident it knows what you meant that it doesn’t even bother to ask. “Showing results for brittney griner,” it notes, before offering, in small text, an option to “search instead for brittany griner.” (Microsoft’s Bing search engine does the same thing, with slightly altered wording.)

Over the past five years, Web browsers have become better at spelling than most humans. In a 2008 Slate piece, Chris Wilson noted that browsers had surpassed traditional spell-check programs as well. Microsoft Word will turn morgage into mortgage, but it sees nothing wrong with American Idle, and it thinks the way to fix Brittany Griner is to make her a Grinner or a Grinder. And while Word’s spell-check function has stagnated, browsers get smarter by the day, assimilating new search data in real time. Search engines don’t just catch errors; they prevent you from making them in the first place. To find the latest on Afghanistan, just type “Afg” and select from the options in the drop-down autocomplete menu.

This functionality will soon be coming to a word processor near you: Last month, Google began applying similar algorithms to the spell-check feature on Google Docs, marking one of the first attempts to apply Web-based, contextual spell-checking to a word-processing program. Smartphone autocorrect tools are also improving, though embarrassing errors remain all too common. And mobile personal-assistant tools like Siri and Voice Actions for Android combine cutting-edge voice-recognition technology with search data to cut out the need for typing entirely.

With spelling becoming more and more optional, it’s easy to draw a parallel to the changing nature of arithmetic. Once upon a time, if you wanted to divide 154 by 19.6, you had to get out a pen and paper. Now, an electronic calculator—or the calculator on your computer or your watch or your iPhone—will do it for you. Just as long division has become all but obsolete, could autocomplete make spelling a lost art—something for kids to learn in elementary school, never to use again?

Probably not. Still, it’s a big improvement over what has come before.

Kids may struggle equally with spelling and arithmetic, but the latter is vastly easier for machines. Numerical calculations, which follow simple, unchanging rules, are right in a computer’s wheelhouse: Fourteen times eleven is 154 no matter what the context. By contrast, complement is a correct spelling if you’re talking about two things that go together but a spelling error if you’re doling out praise. To tell the difference, a computer needs to do something much harder than consult its internal dictionary. It needs to understand the meaning of your sentence. That’s what artificial intelligence researchers call a “hard problem.”

A “hard problem,” in AI, is more than just a problem that’s tough to figure out. Rather, it’s a problem that can’t be solved without also solving the fundamental problem of the field: how to make a computer think like a human.

Spell-check programs work best when they don’t have to think. Type typpo and any spell-checker will recognize that you probably meant typo. That’s what’s known as a “nonword” error—a spelling mistake that results in a combination of letters that doesn’t match any real word in the dictionary. Though many typos are of this sort, unfortunately these are the types of mistakes that humans are already adept at noticing. Both humans and spell-checkers struggle more with “word errors”—spelling mistakes that result in a real word but not the one that was intended. Homophones are one pitfall; proper names are another.

Far from making spelling obsolete, traditional spell-checkers often serve to reinforce its importance. A widely cited 2005 study found that students actually caught fewer spelling and grammar mistakes when their word processor’s language-checking program was turned on. The explanation: People placed undue confidence in the software, skipping over misspellings that the computer didn’t flag.

The same can happen with Web-based spell-checking, but it’s less common. That’s because search-based algorithms consider two things that spell-checkers don’t: context and human experience. Rather than checking a string of letters against a dictionary, Bing and Google check your phrase against the millions of other Web searches that people have conducted. The moment you type mor, a search engine will find the most popular search results that start with those three letters: mortgage calculator, morgan stanley, morgan freeman. It can work even if you start off on the wrong foot. Type tarmig and the first suggestion is ptarmigan, an oft-misspelled game bird. And while autocomplete isn’t perfect on homophones, it’s better than Word’s spell-checker. If you type here, it doesn’t know that you might have misspelled hear. If you type do you here what I here, however, it knows which Christmas carol you’re asking about.

Still, autocomplete is no substitute for human spelling skill, for several reasons. For one, you need to get somewhat close to a word’s proper spelling in order for it to be helpful. Second, it hasn’t yet been incorporated into most email and word-processing programs. The recent Google Docs upgrade is a significant improvement over traditional spell-check tools, but work remains to be done. Google Docs now flags the Brittany in “Brittany Griner Baylor” because that’s a popular search term. But when it comes to a phrase like “Brittany Griner of Baylor,” which you’d be more likely to use in an essay, Google Docs is mute. Spell-check programs will probably never make spelling as easy as a calculator makes arithmetic. But at least now they can put two and two together.