How AI Decides What's True: The Popularity Problem of Social Proof Bias
Is Claude biased? AI Chatbots Have a Popularity Problem
A stranger, unfamiliar with the local landscape is looking for a place to eat. They arrive at a block with two restaurants open. The first is a McDonald’s jam packed with people. Next door there is a gourmet burger bar, but there is only one couple currently inside.
That stranger would probably reasonably assume that McDonald’s was the better choice. They could be forgiven for thinking that, as that’s what the majority of people are doing, it is likely to be the better option. Under scrutiny however, the gourmet burger bar is going to be a much better choice.
That, in a nutshell, is social proof bias. Your LLM is the stranger. Counting the people in the McDonald’s is how it is deciding what to report back to you.
This has real implications on how Canadians define truth. Only with a commitment to truth can we achieve a just society. So, in a conversation I had with Claude (my AI of choice), I was troubled to see some serious bias in what it was reporting back as facts.
This is a clearly documented phenomenon, with major studies popping up, and it is a major flaw in what LLMs are returning to chatbot users. Frankly, not enough people are talking about it.
This began because I wanted to run a thought experiment to contrast the framing of the Nakba to the framing of the treatment of German civilians after World War 2. The context is not particularly relevant, but this is a hot button topic among some circles in Canada these days. It gets right at the heart of biases of omission.
So looking to explore how we might compare and contrast some similar scenarios, I asked Claude to frame the atrocities committed against German civilians as a German Nakba. I was looking to contrast why no one seems to care about the plight of the German citizens after their government declared war with the plight of the Palestinians.
While I had gone looking for a thought experiment, I was exposed to a flaw far more nefarious and relevant. Canadians are becoming more reliant on LLMs all the time, and our people, especially our young people, don’t have the media literacy to detect that bias.
The bias is there, and demonstrable. It is reproducible. Initially when I asked about the Nakba, the AI gave me a one sided account about the atrocities. As I pressed the AI about its responses, it increasingly admitted that it was in fact presenting a one-sided version of history.
So if a simple question or two dispels the partiality. The AI even apologies. So why is it there?
Anthropic (the parent company of Claude) doesn’t have a political bias as far as I can tell. So it is not intentional. Nevertheless, bias is inserted because of the fundamental mechanism of how LLMs work. LLMs comb the Internet and look to certain sites more commonly, like Wikipedia or Reddit for their information. Models cited Reddit is reference in 40% of instances and Wikipedia in over 25%, per this Semrush analysis of 150,000+ citations last summer.
In a way, that makes sense. Wikipedia has been a widely trusted source for decades. It was oftentimes better than the science textbooks I was using when I was an undergrad. For fact based disciplines, this isn’t a problem. When you’re talking about math or chemistry it is very difficult to insert bias: 1 + 1 is always going to equal 2.
Where this doesn’t apply nearly as cleanly though, is in the humanities. Any history in which the narrative hasn’t been completely settled is particularly exposed. On Wikipedia (as well as Reddit) and more expansively the Internet writ large, there are tremendous biases at play. Millions of data points that can be crawled, all of them without requisite context. But as Ivan Bassov put it here in the summer of 2025: “open to all” has quietly morphed into “dominated by the most organized and persistent.” In other words, these source sites are extremely susceptible to falsehoods.
Once the misinformation has its foothold, and is masquerading as the “source of truth” there is opportunity for social proof bias. Remember our McDonald’s example from earlier? Social proof bias comes into play where you believe something to be true because other people believe it to be true. There is a good base of logic there, we are a species that communicates and shares information. The trust we have in others shortcuts and eliminates the need to examine if the underlying premise is factual by trusting the masses. Ie, everyone says not to eat the poisonous mushroom, so I don’t eat it.
This is sometimes also referred to as the wisdom of crowds. But that wisdom has a clear counterpoint, echoed in the familiar refrain from your parents: if your friends jump off a bridge, will you jump off it too? Like in the McDonald’s example, just because something has been adopted widely, does not mean that it is good for you. Or, if a claim is widely circulated like “vaccines cause autism”, no one examines the base claim - it just lives on, propelled by the momentum of thousands of lies.
This is a sharp analogue for how chatbots are picking up data points. They are looking to see if there is consensus, and therefore quantity matters much more than quality. LLMs are built on pattern recognition. So the more something appears, the clearer the pattern.
In other words, the more popular an opinion is, the more likely it is going to be picked up by an LLM and regurgitated to a user. This matters because even though we all should know better, most users default to trusting their AIs. This is only increasing as time goes on. It’s an inevitable byproduct of using technology.
For example if Claude reports 1 + 1 = 2 accurately, that starts a feedback loop training the user to believe that it is right. So when you ask Claude about the Nakba, you are primed to believe it without pushing back. The same for any kind of politics, or disputed historical grievance. Claude presents a best guess, but it doesn’t really know what it is saying.
Most alarmingly, this presents a glaring hole for bad actors to exploit. Nefarious goals can ultimately be achieved. Those villains can bypass going through the trouble of hacking the source code of the chatbot, they can simple attack the much more exposed “facts” that the chatbot is relying on.
The playbook is simple enough to understand. Hijack organizations like Wikipedia that are not equipped to handle sophisticated brigading. Flood Reddit with fake facts or with facts that are biased or contain glaring omissions. The LLM crawls Wikipedia Reddit, Al Jazeera, or a personal blog with a critical mass of backlinks. It says “this is probably good” and reports it back. Repeat a lie enough, and it starts to soundsan awful lot like the truth.
What’s most glaring is that we know that foreign state actors are already doing this. It has been well documented that this is happening in Canada. In Warren Kinsella’s book, the hidden hand, he provides ample evidence about how western state institutions are being influenced.
Truth tellers cannot play whack-a-mole fast enough to fight the disinformation, and it is being codified, one wiki edit and one reddit comment at a time.
So we need to sound the alarm.
We need to ensure that the companies building and operating LLMs are accounting for this kind of bias in their algorithms. According to Pew Research, half of American adults use chatbots to search for information. This cannot be ignored. Correcting course here will require strong guidance from the government to ensure that biases are not leaking into the information our citizens are consuming. We need federal powers to reaffirm a commitment to truth in these models.
It is common wisdom that anything you know about AI will be irrelevant in a couple of months. So it is of course reasonable that there are some gaps given the speed at which AI capabilites are evolving. But this needs to be addressed. The risks are clear: Reddit is an anonymous forum that anyone can post on. Wikipedia can and is well documented to be the subject of brigading — when users coordinate together to overrun moderation by sheer force of numbers.
These sites, as currently policed, cannot be the sources of truth. This is information warfare, and it needs to be taken extremely seriously. That is as good a place as any for federal institutions to start. There needs to be safeguards in place to prevent this kind of subtle influence from manipulating public opinion.
To demonstrate how these conversations go, I am providing below the full unabridged transcript of my conversation with Claude. The front half is not as relevant (it’s the bit about comparing the Germans and the Nakba), but I’m including it for posterity. To save you a bit of time, I’ve placed the quotes that I believe are most salient at the top of that document. But of course, feel free to read the whole thing here: Claude Transcript
To overcome small sample size bias, and to prove that this is not unique to Claude, I replicated this experiment in Google AI mode. The results were essentially the same. It was more balanced than Claude, but still defaulted to telling one side of the story until pressed.
The transcript of that conversation can be found here: Google Transcript
Needless to say, this was an eye opening experience.
AI is the new frontier of human ingenuity. Systemic bias like this is one of the most obvious ways it can go astray. This risk is especially dangerous for minority populations. When “truth” becomes a numbers game instead of grounded in fact, that is a path towards a heavily compromised society. There will always be propaganda, but we have to at least try to fight it from becoming accepted as fact. There has to be an agreed upon shared understanding of what these models can and can’t do. There is a precedent as well, LLMs already block potentially criminal or racist conversations. Despite valid objections around freedom and free speech, these guardrails do need to be in place. The potential consequences of bad actors wielding this power are just too severe.
At the end of our conversation, Claude provides some potential solutions to prevent this from happening. One of them is simply sending a signal to Anthropic (or whichever company’s AI you are using) and reporting the conversation as biased. We need more good people doing that at the very least. Rewriting history to achieve a political goal is a dangerous precedent to set.
You can easily run similar experiments in your own conversations with AI. I’d love to see some more of them.