November 2, 2026
Last week, I was reminded of one of the internet’s oldest magic tricks: write it on Wikipedia, and eventually people start treating it like scripture.
Now we’ve upgraded the trick. It’s no longer just journalists, students and that one bloke in the office who speaks with suspicious certainty. It’s AI too.
That matters more than most people think.
If you are a founder, operator, or leader hiring into data, cyber, infrastructure or product, you are already living in a world where AI tools summarise people, companies and markets before a human has had their second coffee. And if those tools are leaning on weak or outdated public information, bad inputs become polished nonsense at scale.
A recent piece in Smart Data Collective digs into this exact issue: why LLMs still treat Wikipedia as a source of truth, despite all the chest-thumping about broader training data and more advanced reasoning. Source: Smart Data Collective, 'Wikipedia Editing Services and AI Training Data: Why LLMs Still Treat Wikipedia as a Source of Truth'.
My view is simple: Wikipedia still punches far above its weight in AI outputs. And if you think that is just a trivia problem, you are missing the operational risk.
Let’s call this what it is.
Wikipedia is the internet’s default credibility layer. It is the page people scan when they want the quick story. It is structured, highly linked, constantly referenced, and written in a tone that sounds authoritative even when it is having a bit of a wobble.
That makes it catnip for machine learning systems.
The Smart Data Collective article points to training papers and current model behaviour suggesting that Wikipedia remains a meaningful source in model development and response generation. Even where models are trained on broad corpora, Wikipedia has three advantages:
It is neatly structured, which machines love.
It is widely mirrored and cited, which boosts its footprint.
It carries an aura of legitimacy, which means its claims often get repeated elsewhere.
So even if a model is not directly quoting Wikipedia, it may be absorbing versions of the same claim from a dozen other places that copied or paraphrased it.
That is the important bit. Bad information does not need to stay on one page to become influential. It just needs to spread early, sit still, and look tidy.
There is a slightly annoying myth doing the rounds that newer AI systems are somehow floating above the messy old web, doing pure thought in a vacuum like a philosopher king with GPUs.
Come off it.
LLMs are pattern machines. Brilliant ones, yes. But they are still shaped by the material they absorb and the confidence signals around that material.
Wikipedia is one of those confidence signals.
The article highlights a point many leaders should pay attention to: when users ask models about disputed or outdated facts, the answers can still reflect Wikipedia-shaped assumptions. Sometimes directly. Sometimes through the wider information ecosystem it influenced. Either way, the end user gets an answer that sounds clean and settled.
That is where the trouble starts.
AI does not need to be wrong in a dramatic way to cause damage. It just needs to be slightly off, at scale, in contexts where people move quickly.
Think about where that shows up:
Candidate background summaries
Company research before interviews or pitches
Executive reputation checks
Market mapping and competitor scans
Automated internal briefings
If the underlying public data is skewed, stale or manipulated, AI can package it beautifully and deliver it with the confidence of a bloke explaining wine after one vineyard tour.
I run a recruitment business. So naturally I am going to bring this back to hiring, because that is where consequences get expensive.
Leaders love speed. Fair enough. Hiring can be painfully slow, and everyone wants leverage. AI promises leverage.
But if you are using AI-assisted research to assess candidates, benchmark leadership profiles, or build shortlists of target companies, you need to ask a very unsexy question:
What is this answer standing on?
Because polished misinformation is still misinformation.
I have seen versions of this problem for years, long before ChatGPT made everyone an instant prompt engineer. A half-true public profile becomes accepted narrative. Then it gets repeated in blogs, bios, databases and introductions. Eventually it hardens into reputation.
Now AI can accelerate that process.
That creates three risks for employers:
Bad filtering
You may dismiss good candidates because a public narrative is incomplete or wrong.
Weak diligence
You may trust AI-generated summaries without checking the supporting sources.
False confidence
Your team may assume a smooth answer equals a verified answer. It does not.
This is especially dangerous in specialist functions like cyber, cloud, data and BI, where nuance matters and surface-level interpretation can completely miss capability.
Put bluntly, if you rely on scraped internet narrative to judge specialist talent, you are hiring with one eye closed.
Let me be fair to Wikipedia for a second.
The problem is not that Wikipedia exists. The problem is that people, and now AI systems, often treat it as the end of the research process rather than the beginning.
That is a leadership issue as much as a tech issue.
Good operators know the difference between:
A signal and a fact
A summary and a source
A likely answer and a defendable answer
The Smart Data Collective article also points to a practical angle: if you want corrections reflected in AI outputs, unsupported complaints will not cut it. You need evidence-backed updates, published in places models or retrieval systems can actually see and trust. Source: Smart Data Collective, 'Wikipedia Editing Services and AI Training Data: Why LLMs Still Treat Wikipedia as a Source of Truth'.
That is the bit many people miss. You cannot simply shout, 'the AI is wrong'. You need to improve the underlying evidence trail.
Which leads to a useful principle for leaders:
If the public record matters, treat it as infrastructure.
Not vanity. Not fluff. Infrastructure.
If a model keeps surfacing inaccurate claims about your company, leadership team, sector, or public narrative, here is the grown-up response.
Find where the claim appears repeatedly.
Wikipedia
Company profiles
Press coverage
Old bios
Third-party databases
If it is everywhere, the model is not hallucinating. It is inheriting.
Do not try to muscle reality with opinion.
Use:
Reliable published sources
Updated official company pages
Credible press coverage
Public filings or institutional references where relevant
If the issue involves Wikipedia, any correction needs to meet Wikipedia’s sourcing and neutrality standards. That is not glamorous, but it is how the thing works.
Your website, leadership bios, newsroom, and company descriptions should be current, specific and consistent.
A shocking amount of public company information is vague, outdated or written like it was approved by a committee that feared nouns.
Clear source material matters because AI systems, users, journalists and candidates all bounce off it.
This one is vital.
If your talent team, exec assistants, researchers, marketing staff or ops leads are using AI for briefing and research, give them a simple rule:
Never make an important decision from a summary you cannot source.
Yes, that slows things down slightly. It also avoids making daft decisions with great formatting.
If you want a quick way to test whether an AI-generated claim is trustworthy, use this:
Where does the claim appear first or most authoritatively?
If you cannot identify a credible origin, be cautious.
How has the claim been framed?
Has nuance been stripped out? Has a contested point been turned into a neat fact? Has biography become mythology?
How widely has it propagated?
If dozens of sites repeat the same thing, that does not prove it is true. It may just mean one bad source had a very good day.
This framework is useful in hiring, due diligence, vendor research and comms. Frankly, it is useful anywhere someone says, 'AI told me'.
If this topic is hitting a nerve, good. Here are a few questions worth asking your team this week:
What public sources most shape how AI describes our company?
Are our leadership bios and company profiles current and evidence-based?
Do we know which inaccurate claims about us are repeated online?
Are teams using AI summaries for hiring or diligence without source checks?
Who owns correction of public narrative when it affects hiring, reputation or sales?
If nobody owns this, that is your answer. You have a gap.
For all the talk of frontier models, synthetic reasoning and AI’s brave new future, a lot of machine confidence still rests on very old internet scaffolding.
Wikipedia is part of that scaffolding.
So no, this is not just a nerdy quarrel about online editing. It is about how reputations are formed, how decisions are accelerated, and how easily bad information gets dressed up as insight.
The smartest leaders will not panic about that. They will do something more useful.
They will tighten their public source trail, teach their teams to verify before trusting, and stop confusing smooth AI output with truth.
If you are hiring the people who build and check these systems, see our data science, AI and machine learning recruitment.
Because if you leave your public narrative unattended, do not be surprised when the robots start freelancing.