AI has slipped almost seamlessly into our daily lives. It is fast, useful and often remarkably good at what it does.
But beneath those polished answers is a problem we need to understand: Synthetic Consensus.
Most people will never have heard the term. The simplest way I can describe it is as a childhood game of Chinese whispers, mixed with the power of a rumour.
Let me explain.
The Rainbow Eggs
Imagine telling a friend that you have just been to the local market and bought a box of unusual eggs, painted in all the colours of the rainbow.
As adults, we know chickens do not naturally lay rainbow-coloured eggs. But a small child overhears the conversation and assumes that they do. Their parents decide to keep the fantasy going.
The child tells their friends at school the next day. Those children tell their parents, and some of the parents visit the same shop and buy the rainbow eggs themselves.
To the children, this appears to confirm the story. More people are talking about the eggs, and the eggs really do exist. What they do not understand is that the eggs were painted after they were laid.
In just a few small steps, a misunderstanding begins to look like an accepted fact.
Now change the characters.
The parent becomes the original source of information. The child becomes an AI system looking for an answer. The classroom becomes the wider network of AI-generated content spreading across the internet.
The first AI repeats the claim. Someone publishes that answer online, where it is indexed, rewritten and discovered by other systems. Eventually, another user asks the same question and receives the rumour as fact - apparently supported by several sources.
But every source leads back to the same unverified claim.
This is Synthetic Consensus.
When Several Sources Are Really One
Synthetic Consensus is the appearance of independent agreement created when multiple sources repeat information derived from the same unverified origin.
An AI produces an answer. Someone publishes it. Other websites repeat it. Search engines index those pages, and other AI systems use them to produce new answers.
Each repetition makes the claim appear more established.
An AI may find ten websites making the same statement and treat that as supporting evidence. But those websites may not represent ten independent sources. They may simply be ten digital echoes of the same original answer.
Repetition is not corroboration.
When machines repeat one another, agreement is no longer proof.
The Loop Has Already Begun
This is not entirely theoretical.
In 2024, a Forbes investigation reported that Perplexity was citing AI-generated blog posts containing inaccurate, outdated and sometimes contradictory information. Searches involving cultural festivals in Kyoto, healthcare, Bangkok street food and young tennis players returned answers supported by AI-generated material.
The information had already travelled from an AI system into published webpages, before being found and presented by another AI system as supporting evidence.
A separate 2025 study of Google AI Overviews examined 29,000 searches involving health, finance, law and politics. It found that 10.4 per cent of the pages cited by Google’s AI answers were apparently AI-generated. More than half of all citations came from outside Google’s top 100 ordinary search results.
Neither example proves that a single false claim has already spread across the internet and become accepted history. But both demonstrate that the underlying loop now exists: AI-generated material is being published, indexed and returned to users by other AI systems as supporting information.
The Danger of Trust
AI systems are genuinely useful. They streamline our workloads, support creativity and help us solve problems efficiently.
Their usefulness is also what makes this issue dangerous.
Because the answers are usually helpful, we rarely question where they came from. Trust gradually becomes a habit. In many cases, we have outsourced our scepticism without noticing.
But what happens when something important - a medical fact, legal position or historical event - is delivered with the same confidence and turns out to be completely wrong?
The AI system is doing what it usually does, and the user is responding as they usually do.
The system provides an answer. The user accepts it.
So who is responsible when that answer is wrong: the system that presented it, the user who trusted it, or the wider information chain that made a rumour appear true?
There is no simple answer, and that is part of the problem.
From Searching to Receiving
Misinformation and scams have existed since the beginning of the internet. But we once had to search for information ourselves, choose which websites to open and decide which ones to trust.
After decades online, most users have developed at least some instinct for navigating the web. We notice suspicious addresses, poor spelling, excessive advertising and websites that simply do not feel genuine. We compare results and make our own judgements.
AI changes that relationship.
Instead of searching through competing sources, we increasingly receive one polished answer. The searching, comparing and summarising happen out of sight.
We may not know whether the answer originated from an official document, an experienced professional, an anonymous blog, an AI-generated article or several websites repeating one another.
The final response can look equally convincing in every case.
The problem is therefore not simply whether an AI provides citations. It is whether those citations represent independent evidence or several retellings of the same claim.
When Errors Become History
This is where I suspect we could be heading.
The barriers to publishing online are disappearing, and convincing websites are becoming easier to create. At the same time, AI systems depend on being able to find and distinguish reliable information.
Even when a system checks several sources, the original evidence could eventually become buried beneath years of repetition.
Small errors can creep in with every retelling. A detail changes, context disappears or an assumption becomes a statement of fact. That altered version is then repeated again.
Eventually, the most visible version may no longer be the most accurate. It may simply be the version reproduced most often.
A mistake repeated often enough can begin to look like consensus. Given enough time, that consensus could begin to look like history.
So What Is the Solution?
The companies developing AI have a responsibility to improve source tracing, recognise repeated information and distinguish original evidence from recycled content.
But no system will be perfect.
As users, we must understand that a confident answer is not necessarily a verified answer - and several citations do not necessarily represent several independent sources.
Most everyday questions will not require an investigation. But when an answer could materially affect our health, finances, legal position, work or understanding of an important event, we should pause and ask one additional question:
Where did this information originally come from?
Synthetic Consensus does not mean AI cannot be trusted. It means trust should not become automatic simply because an answer is polished, immediate and widely repeated.
The internet may soon contain more answers than ever before.
The danger is that one day it could contain more answers than evidence.