Confirmation bias in search

when technology panders to an existing weakness of human psychology

Confirmation bias is the tendency to seek information which will bolster an existing belief.

There’s a related idea, which is the tendency to dismiss contradictory evidence. As a purist, I insist on calling this “biased assimilation”, although I recognise that making the distinction between a bias in what information you choose to accept (biased assimilation) versus what information you actively choose to go out and look for (confirmation bias) is a) possibly too much for most people b) a similar kind of lost cause to my insisting that “beg the question” does not mean “raise the issue”, or that the Chatham House Rule is the singular Rule, not the plural Rules.

Confirmation Bias is a threat to reasonable belief updating. If you search for information that confirms what you believe you lower your chances of discovering the contradictory facts. Some tasks routinely evoke what looks like a confirmation bias in most people, although there is debate about the extent to which this may reflect an adaptive response to seeking information in a confusing world. What is clear is that technology, and in particular the ubiquitous search function of the information age, is capable of interacting with any native tendency we have towards confirmation bias.

Before all search was juiced by LLMs and LLM summaries, confirmation bias may have been a mechanical product of overlap between the query and the documents returned. Search for “health benefits of coriander” and pages containing the phrase “health benefits” will necessarily have a higher overlap with the query than pages that don’t. Now most major search engines use LLM technology both in their search (e.g. to find synonyms) and to provide a summary of results, you may get more balanced results. On the other hand, the tendency of commercial AI companies to build sycophancy into model behaviour may over-rule any balance from an AI summary — there’s a short line between being nice to the user and delivering answers that confirm what it appears they want to hear.

§

I thought I’d try a direct test of confirmation bias in search and fired up Google to ask the same question, framed in both negative and positive ways,

Asking "Is Sheffield an unsafe city?" gave me this:

“CrimeRate ranks Sheffield as the most dangerous major city in South Yorkshire”. For context, non-UK readers should know that Sheffield is the only major city in South Yorkshire, but still the answer is framed in the terms of the question. Asking if Sheffield is unsafe confirms that it tops a list of most dangerous cities.

Asking “Is Sheffield a safe city?” gave me this:

“Sheffield is generally considered one of the safest major cities in the UK”

This seemed to clearly demonstrate the confirmation bias I assumed I’d get from a search engine, so I quickly posted it to social media and immediately got replies from people who tried the same thing and didn’t get different answers, either in the AI summaries or in the results returned for topics different from Sheffield and safety.

Had I, ironically, fallen prey to confirmation bias about confirmation bias? I had stopped after one example, seeing that it confirmed my expectation.

§

I think a few things could be going on here.

First, it may matter if you do a search which provides personalised results (e.g. you search Google while logged in to your Google account). I have no reason to think personalisation will definitely affect the results, but I think the less a search engine knows about you, the more chance of getting answer that are most heavily dependent on ambiguities in the question.

Second, I don’t think all topics will provoke search results which are so different. Famously, the tech companies intervened during the COVID-19 pandemic to make sure that leading questions returned balanced information about vaccine safety, so that topic is out. Other topics will be so unambiguous (“Is the earth a planet?”) there is no raw material for a biased result.

Firing up Tor to obfuscate my identity, and using the Brave search engine, I did a few more searches:

Politics first,

“Is Keir Starmer failing?”, gave me an answer with simple start: “Yes”

“Is Keir Starmer succeeding?”, gave me a more ambiguous answer: “Keir Starmer is currently still the UK Prime Minister, but his leadership is under severe threat”

Health:

“are the[re] dangers of getting vaccinated?” gave me an an answer discussing side effects and the rate of extreme reactions:

“is it safe to get vaccinated?” gave me an answer which starts with (bolded) “Yes, getting vaccinated is safe”

§

At this point, I think I confirmed that the confirmation bias in search can happen. Obviously it doesn’t always happen, and there’s a programme of work to figure out which topics and phrasing make it more likely (as well as how different search engines are affected).

Part of the freedom of writing a newsletter is to allow myself to publish before committing to a multi-year research programme to get to the bottom of a topic. If you are inspired to do your own opposite searches, please do hit reply or let us know in the comments how it goes and if you find any interesting patterns.

§

Obviously if I search for “ways in which X is great” I will, and should, get different results than if I search for “ways in which X is terrible”. A simple “Is X a Y?” question feels like it shouldn’t give different results to the “Is X not a Y”?” question, but is that really so? There’s information in how a question is asked, so it isn’t clear to me that the potential for opposite questions to give different results should be removed from search, even if it could.

From a research and media literacy perspective, however, the lesson is simple: a great deal of the bias in information technology is a product of how it interacts with our own biases. When we’re looking for something we should check we’re not just looking in ways that risk confirming what we’re asking. Asking if the opposite is true is a good search strategy, as well as a useful tool for thought.

---

This newsletter is free for everyone to read and always will be. If you can afford it, feel free to chip in to help me keep writing for everyone: upgrade to a paid subscription (more on why here).

See below for references and other things I’ve noticed.

---

References

The classic review of confirmation bias is:

• Nickerson, R. S. (1998). Confirmation bias: A ubiquitous phenomenon in many guises. Review of general psychology, 2(2), 175-220.

Other stuff, but sticking with politics…

Podcast: What did ‘Nudge’ get wrong? With Nick Chater

Nick Chater is a behavioural scientist big cheese, and was on the Climate Change Committee after being involved with the UK’s nudge unit. In this pod he describes how that experience was part of his movement from being a proponent of nudge to recognising the limitations of individual framing of societal problems. Interesting!

Link: What did ‘Nudge’ get wrong? With Nick Chater

Screaming expressive responding batman!

This is a graph of “level of trust” in Putin as president by Russians (from Why the Russian regime is not on the brink of collapse). Although this is besides the main point of the article, the result seems to me like a screaming example of expressive responding. I don’t believe 25% of Russians changed their view of Putin overnight when he starts a war, but I do believe that a time of war might change the people’s feelings about expressing support, or not, for their leader.

Book: Why We Think What We Think: The Unexpected Origins of Our Deepest Beliefs:

Turi Munthe’s new book was published on Thursday: Why We Think What We Think: The Unexpected Origins of Our Deepest Beliefs:

It’s an engaging, humane, tour through a rich seam of research into factors that appear to affect our beliefs, from genetics, upbringing, personality, climate and our social context. Ultimately, the book is a plea for humility and pluralism. By unmooring the certainty that our thoughts are ours alone, without origin, the book encourages us to be a bit more hesitant that we are completely right, even about our strongest beliefs, and a bit more disposed to entertain the strong beliefs of others.

Full disclosure: Turi sent me a pre-publication copy to read, and in the context of non-rational influences on beliefs it seems only fair that I acknowledge that although I don’t think that this influenced my judgement that this is a good book, it may have!

Link: Why We Think What We Think: The Unexpected Origins of Our Deepest Beliefs

Catch-up

Recent posts:

• When trusted signals collapse. A case study in what can happen when AI levels the field

• The 1,000 neuron challenge. A competition to design small, efficient neural models might provide new insight into real brains—and perhaps unite disparate modeling efforts.

• The better algorithms of our nature. Engagement, bridging, and the design of digital platforms which don’t pander to our weaknesses.

• The mythical median voter. Most people have an above average number of legs, and what that means for our political imagination

… And finally

“Don’t let anyone else ruin your day for you”

Comments? Feedback? Search results? I am tom@idiolect.org.uk and on Mastodon at @tomstafford@mastodon.online

AI declaration: I write all the words and think all the thoughts myself. I asked Gemini to check for spelling and grammar.