Autocomplete and bias: search as a mirror, and how we filter it

A game built on search predictions has to face an uncomfortable fact: autocomplete reflects what people type, and people type all sorts of things. Some of it is charming, some of it is baffling, and some of it is ugly. This article looks at how researchers and campaigners have used autocomplete to expose bias, what Google says it does about harmful predictions, and exactly how GuessTheSearch decides what reaches a button in our Stereotypes theme.

A mirror, not a window

It is tempting to think of a search engine as a window onto the world: type a question, see what is out there. Autocomplete is better described as a mirror. Its predictions are drawn from what people have searched before (we cover the mechanics in how Google autocomplete works), so they reflect the curiosity, confusion and prejudice of the people doing the searching.

A mirror can be useful precisely because it does not flatter. Over the past decade and a half, two pieces of work in particular made the public look hard at what that mirror shows.

“The Autocomplete Truth”

In 2013, UN Women ran a print campaign called “The Autocomplete Truth”, created by the advertising agency Memac Ogilvy & Mather Dubai. The idea was very simple. Each ad showed a close-up portrait of a woman, with a search bar placed over her mouth. In the bar was the start of a question about women, and underneath it were real autocomplete predictions that the team had captured.

The predictions were stark: variations on the idea that women should stay at home, should not have rights, or needed to be kept in their place. The campaign did not claim that search engines had invented those views. Its point was the opposite — that the predictions were a snapshot of attitudes held by enough people, typed often enough, to be predicted. Sexism was not a fringe opinion hidden in a dark corner; it was showing up in the most ordinary tool on the internet.

The ads were widely shared and discussed, and they did something important for the conversation about search: they framed autocomplete as evidence about society, not just as a convenience feature.

Algorithms of Oppression

In 2018, the scholar Safiya Umoja Noble published Algorithms of Oppression: How Search Engines Reinforce Racism. The book grew out of her research into what search engines returned for queries about Black women and girls, which for a period surfaced sexualised and pornographic material rather than anything a girl searching for herself might want to find.

Noble’s argument goes further than “the mirror shows ugly things”. She contends that search results are not neutral reflections at all. They are shaped by commercial incentives, by advertising, and by design choices, and they can reinforce the stereotypes they reflect — especially for groups with less power to shape how they are represented. A prediction or a result that appears at the top of a trusted tool can look like a fact, even when it is simply popular.

We take two lessons from her work. First, “it is just what people search” is not a complete defence: whoever builds on search data decides what to amplify. Second, the harm is uneven. A joke prediction about a star sign and a demeaning prediction about a racial group are not the same kind of thing, and should not be treated the same way.

What Google says it removes

Google publishes policies for autocomplete, and they have been tightened over the years, often after journalists and researchers pointed out offensive predictions. In broad terms, Google says it does not allow predictions that are:

• violent, or that promote violence;
• sexually explicit, vulgar or profane;
• hateful towards groups based on characteristics such as race, religion or gender;
• harassing or disparaging towards individuals;
• promoting dangerous or harmful activities.

Google has also described specific rules for sensitive areas such as elections, where it aims not to show predictions that could be read as claims for or against a candidate or party, and it removes some predictions for legal reasons, such as court orders in particular countries.

According to Google, enforcement works in layers. Automated systems are designed to stop policy-violating predictions from appearing in the first place; when some slip through, dedicated teams remove them; and users can report an inappropriate prediction directly from the search box. Google has also said that for some kinds of queries it may show fewer predictions, or none.

How GuessTheSearch filters predictions

Google’s policies are designed for a search box, where someone is actively looking for information. A party game is a different context. An answer on a GuessTheSearch button is something players are invited to laugh at and share, so we apply much stricter rules, in layers:

1. A blocklist of slurs and violent phrases, applied to every theme.
2. Platform noise — predictions that just point at forums or video sites — dropped everywhere, because “where people asked” is not an answer.
3. Identity claims about religion and sexuality are dropped in every theme. “Is [country] a [religion] country?” is a real search, but it is not the game we are making.
4. Name collisions. Many subjects share a name with a character, celebrity or club, and autocomplete happily answers about the wrong one. A question about Berlin can turn out to be about a television character, and “Cancers” can mean the star sign or the disease. We filter known collisions so a light question about a zodiac sign does not turn into a medical one.
5. No feuds about places and people. For countries, cities and jobs, we drop “why does X hate Y” predictions, along with racial terms and phrases like “third world”. Animals and gadgets keep their “hate” — cats and water are fair game.

The Stereotypes theme gets by far the strictest set, because it is the one theme that is explicitly about groups of people. On top of everything above, it excludes:

• race, skin colour and ancestry;
• religion;
• party politics and political labels;
• sex and sexual content;
• health, illness and self-harm;
• put-downs of looks, intelligence, class or hygiene;
• borders, territorial disputes and immigration.

Unlike our other themes, Stereotypes has no hand-written answers. Every answer is a captured search, because a game about stereotypes only means something if the answers are real. That makes the final safeguard the most important one: every recapture is reviewed by a person before it ships. Filters are good at catching words; they are bad at catching meaning, and new completions appear constantly. Our editorial standards set out the rules in full.

What survives

One honest consequence: filtering changes the ranking. The #1 answer in GuessTheSearch is the highest-ranked prediction that survives our filters, which is not always Google’s raw #1. We think that trade-off is worth it. Here is what a typical Stereotypes list looks like after filtering:

These are questions about language, habits and road rules. (“Put x” is people wondering why British texters sign off with a kiss.) Most of our Stereotypes data looks like this, and it is the kind of thing groups tease themselves about. Plenty of predictions about families and life stages are similar:

Not everything that survives is gentle, though, and we think it would be dishonest to pretend otherwise. Some predictions encode a stereotype that research has questioned for decades:

The “selfish and spoiled only child” is a long-standing cliché, and studies comparing only children with children who have siblings have generally found far smaller differences than the stereotype suggests. Notice, too, that the list is phrased as questions — many of the people searching this may well be parents of only children, checking a worry. Keeping a list like this in the game is a judgement call, and we make it because the question is common, the stakes are low, and seeing it can prompt exactly that kind of second look.

Laughing at the question is not endorsing the answer

This is the heart of why we think a game about stereotypes can be done responsibly. GuessTheSearch never asserts that any answer is true. It asks something narrower: what do people search? To play, you have to recognise the stereotype as a stereotype — a shared picture — and guess how common it is. That is a different mental act from believing it.

A few principles guide where we draw the line:

• The joke is on the question. When “Why are teenagers always played by adults?” turns up, the humour is in what people wonder about, not in mocking teenagers.
• No punching at protected characteristics. If the laugh would come at the expense of someone’s race, faith, body, health or sexuality, it does not get a button, however popular the search.
• Context travels with the data. Articles like where stereotypes come from, European stereotypes in autocomplete, generations in autocomplete and what Google thinks of every star sign exist so that the questions come with some explanation of why people ask them.
• We will get some wrong. Filters miss things and judgement varies. If you see an answer you think crosses the line, please tell us and we will review it.

The UN Women campaign used autocomplete to hold up a mirror to something ugly. We use it mostly to hold up a mirror to something silly — why Brits say “bloody”, why teenagers are always tired. Both depend on the same idea: seeing what everyone else is typing can make you look twice at the pictures in your own head.

If you are new to the game, how to play covers the rules, and our strategy guide explains how to think like a searcher.

▶ Play today’s stereotypes round

GuessTheSearch is an independent game and is not affiliated with Google. All predictions shown were captured in September 2026 in US English and will change over time.