Most people who have ever written a blog post, left a comment on a news article, or posted an opinion on a forum did not sign a release form. They were just talking. That talking — billions of words of it, stretching back decades — is what large AI language models were built on. The systems now producing text at industrial scale learned to do that by absorbing what ordinary people wrote, mostly without those people knowing, and certainly without paying them. That is only part of the problem. The other part is what happens next.
The basic loop most people have not noticed
AI language models are trained on enormous collections of text scraped from the public internet: journalism, opinion columns, Reddit threads, product reviews, comment sections, Wikipedia edits. The people who wrote that material were writing for their own reasons — to inform, to vent, to connect. They were not contributing to a training dataset.
Now those systems are producing new text, and that text is flowing back into the same information environment the original writers live in. News summaries, social media posts, blog articles — a growing share of what people read was not written by anyone who observed or verified anything. It was generated by a system that learned to sound like someone who did.
That loop — human writing goes in, AI output comes back out — is the foundation of nearly every concern discussed in this article.
Why this is a disinformation problem, not just a labour problem
There is a reasonable economic argument about AI and writers: people are losing paid work. That argument is real and worth having. But there is a separate problem that gets less attention, and it is about readers rather than writers.
When AI-generated content displaces human-reported content, the new material has no independent source. It recombines what already existed. A claim that originated in a single report, or a single misreading of a report, can get restated, smoothed, and repeated until its origin disappears entirely.
Call this source laundering. The claim looks like common knowledge because it appears in many places, but all of those places drew from the same upstream error. A careful reader’s first instinct — where did this actually come from? — becomes much harder to satisfy. That difficulty is not accidental. It is a structural feature of how AI-generated misinformation spreads.
The confidence problem: why AI text sounds authoritative
Read a paragraph of competently generated AI prose and you will notice something: it almost never sounds unsure. The tone is steady, the sentences are complete, the logic appears to follow. That consistency is a feature of how the systems are built, not a reflection of how well-grounded the claims are.
Human writers signal uncertainty all the time, often without thinking about it. They write “according to,” “officials said,” “it is not yet clear,” or “this reporter was unable to confirm.” Those phrases are not just hedges — they are calibration tools. They tell the reader how much weight to put on a claim.
AI output often strips those signals away. The result is prose that reads as confident regardless of whether the underlying claim is solid or invented. Confidence signals credibility to most readers, even when the confident voice has no special knowledge at all. That is a known feature of persuasion, and it works whether the confident speaker is human or not.
Feedback loops: when the machine trains on its own output
Here is where the problem compounds. AI-generated articles and summaries are now appearing on public websites, in news aggregators, and across social platforms. Some of that content will inevitably be swept up into the next round of training data for future models.
Think of photocopying a photocopy. Each generation loses a little resolution and introduces new artefacts. Errors that appeared in early AI output — wrong framings, missing context, confident misstatements — can become normalised across later output because the later models learned from the earlier ones.
This matters for narrative control. The organisations or individuals who shape what the first wave of AI output says have an outsized influence on what later systems treat as normal or true. That is a significant amount of power, and it requires no particular technical skill to use.
How this tactic looks from the receiving end
You do not need to understand influence operations to recognise when one might be touching you. The experience from a reader’s perspective tends to feel like this: suddenly there is a lot of content making the same point, from sources you have not heard of, and it all sounds plausible and polished.
The practical appeal of AI for anyone running an influence operation is volume. Producing large amounts of varied, plausible-sounding content used to require many writers. It no longer does. When there is more content than anyone can check, readers tend to default to pattern recognition: if I am seeing this everywhere, it must be true.
AI-generated misinformation can also be tuned to mimic the style and formatting of credible outlets, which makes a quick surface-level scan less useful than it used to be. The concrete signal to watch for: a story with no named reporter, no dateline, and no link to a primary source. Any one of those absences is worth noticing. All three together is a clear prompt to stop and look harder.
What you can actually do when you encounter this
One question cuts through a lot of noise: who originally observed or verified this, and can I reach that source directly? Not “what website is this on” but “what human being, with a name and a job, confirmed this happened?”
Treat confident, sourceless prose as a yellow flag. Not automatic distrust — sometimes accurate information is written plainly — but a prompt to look one level deeper before you share it.
If a claim feels important, try searching for it using older, archived sources. Tools like the Wayback Machine let you check whether a story existed before a certain date. If a claim appeared suddenly and at scale with no trail behind it, that pattern is itself informative.
Pay attention to your own emotional state while reading. Influence operations — regardless of who runs them or what political direction they push — tend to rely on urgency and outrage. If a piece of content is making you feel that you must share it immediately, that feeling is worth pausing on. Slowing down is a defence. Most of these tactics depend on you acting before you have thought twice.
The bigger picture: who benefits from blurring the source
Step back and ask a simple question: who gains when readers cannot tell whether content was made by a person with accountability or a system with none?
Source opacity has always been a tool of propaganda, across every political tradition and in every era. When the author is removed, the motive disappears with them. So does the funder, the agenda, and the track record. A claim with no accountable source behind it carries no reputational cost if it turns out to be wrong.
This is not a reason to distrust everything you read. It is a reason to treat authorship and sourcing as meaningful rather than decorative. The two questions that matter are not complicated: who made this, and what did they have to gain? Those questions do not require specialist training. They require only the habit of asking.
Frequently asked questions
How can I tell if an article was written by an AI rather than a human journalist?
There is no single reliable test, but several signs point in that direction. Look for the absence of a named author, a dateline (the place and date at the top of a news story), and links to primary sources like official documents or named witnesses. AI-generated prose also tends to be unusually even in tone — no rough edges, no admissions of uncertainty, no “this reporter could not confirm.” Human journalism, especially on complex stories, usually shows its working. If an article reads as though every question has already been answered, ask who answered them.
Does AI-generated content always spread misinformation, or is it sometimes accurate?
AI-generated content can be accurate. Systems trained on reliable sources can produce summaries and explanations that are broadly correct. The problem is not that AI always gets things wrong — it is that you often cannot tell when it has. A human journalist who makes an error can be contacted, corrected, and held accountable. An AI system has no byline to hold accountable and no memory of why it said what it said. Accuracy and accountability are different things, and both matter.
Why does it matter that my old online writing was used to train AI systems?
It matters for a few reasons. Your writing reflected your voice, your reasoning, and sometimes personal details about your life and views. That material now contributes, without your knowledge or consent, to systems that produce content other people read as factual. There is also a subtler point: the values, assumptions, and blind spots present in the training data shape what the system treats as normal. Your words — and everyone else’s — are part of what these systems learned to call ordinary. That is a form of influence over public knowledge that happened without any public conversation about it.
Scams, fraud, bots and manufactured noise keep spreading because the internet was built with no reliable way to know