The Internet Is Playing Telephone With Itself
The Internet has always repeated itself. Generative AI changes the scale, speed and cost. One source can become twenty publications and fifty citations - until repetition begins to look like independent confirmation.
Most of us played some version of the telephone game as children.
A sentence is whispered to one person. That person whispers it to the next. The message travels through the room until the last participant says aloud what they heard.
The result is usually ridiculous.
A sentence about a dog, a bicycle and a rainy afternoon somehow becomes a story about a horse, a broken car and someone's grandmother.
Everyone laughs.
The game works because everyone knows something important:
there was an original sentence.
It existed before the chain began.
And when the game is over, we can go back to the first person and ask:
What did you actually say?
That small detail is what makes the game harmless.
The Internet is beginning to play the same game.
Except the line now contains billions of people, websites, search engines, social networks, crawlers, aggregators and generative AI systems.
And increasingly, nobody knows who whispered the first sentence.
We have always repeated things
None of this started with artificial intelligence.
Humans have always passed information from one person to another.
Someone tells a friend something they heard from a colleague. A journalist quotes another newspaper. A blog summarizes a television interview. A forum user links to the blog. Someone posts a screenshot of the forum discussion on social media.
Eventually somebody asks:
Is this actually true?
And someone else answers:
It must be. Everyone is talking about it.
That answer has always contained a logical problem.
Ten people repeating one claim are not necessarily ten independent confirmations of that claim.
They may simply be ten links in the same chain.
The difference is easy to describe and surprisingly difficult to preserve:
publication count is not source count.
Five newspapers can represent five independent investigations.
They can also represent one wire-service report copied five times.
Those are very different information structures even though a search engine may initially see five documents.
Five sources confirm it
Imagine that Source A publishes a claim.
Source B reads A and rewrites it.
Source C reads A and writes another version.
Source D reads B.
Source E reads C.
A search returns all five.
The naive interpretation looks excellent:
Five sources confirm the claim.
But the information graph looks like this:
A
/ \
B C
| |
D E
There are five documents.
There are not five independent sources.
There is one root.
If A is wrong, the entire tree can be wrong.
Nothing particularly mysterious is happening here. The computer may even count everything perfectly.
The problem is not in counting.
The problem is in what we decided to count.
A very small bug with very large consequences
We can reproduce the problem with a few lines of Python.
reports = [
{"source": "A", "claim": "X", "origin": "original"},
{"source": "B", "claim": "X", "origin": "A"},
{"source": "C", "claim": "X", "origin": "A"},
{"source": "D", "claim": "X", "origin": "B"},
{"source": "E", "claim": "X", "origin": "C"},
]
confirmations = sum(
report["claim"] == "X"
for report in reports
)
print(f"{confirmations} sources confirm X")
The program answers:
5 sources confirm X
And technically, it counted correctly.
Now let us ask a different question.
def root(source, reports):
by_name = {
report["source"]: report
for report in reports
}
while by_name[source]["origin"] != "original":
source = by_name[source]["origin"]
return source
roots = {
root(report["source"], reports)
for report in reports
}
print(f"{len(roots)} independent source confirms X")
Now the answer is:
1 independent source confirms X
Same documents.
Same claim.
Same Internet.
Completely different conclusion.
The missing variable was not truth.
It was provenance.
Search engines have seen a version of this movie before
The web has spent decades dealing with a related problem.
Links were once an extraordinarily useful signal.
If many independent websites pointed to a page, that could indicate that the page was useful, important or authoritative.
Then people learned that algorithms valued links.
An industry emerged around producing them.
Link exchanges, artificial networks, doorway pages, automated content, expired domains, reputation exploitation and entire ecosystems of sites created primarily to influence ranking systems followed.
The underlying incentive was simple:
if an algorithm treats a naturally occurring signal as evidence of quality, there is economic value in manufacturing that signal.
Google has spent years adapting its systems and spam policies to this problem. Its current policies explicitly address scaled content abuse, including large amounts of low-value content created through automation or generative AI, as well as scraping, doorway abuse and attempts to exploit the ranking reputation of other sites.
The important part is not that SEO is bad.
It isn't.
The important part is that the history of search provides a useful lesson:
once a machine learns to trust a signal, somebody will eventually learn to manufacture the signal.
Links were one such signal.
Source diversity may become another.
AI changes the economics
Until recently, manufacturing apparent information diversity required work.
If you wanted twenty articles to look genuinely different, somebody had to write them.
If you wanted different tones, different structures, different vocabulary and different audiences, that required additional effort.
Translation required more effort.
Localization required more effort.
Maintaining multiple identities required more effort.
Generative AI changes that equation.
One source can now become ten summaries.
Ten summaries can become fifty rewrites.
The same claim can be expressed as a technical analysis, a local news story, a personal blog post, a social media thread, an expert commentary, a question-and-answer page and versions in twenty languages.
The wording can be different.
The structure can be different.
The examples can be different.
Even the apparent writing style can be different.
And all of them can still have the same informational ancestor.
This matters because one of the old shortcuts for detecting copied information was textual similarity.
Identical sentences were suspicious.
Identical ordering of facts was suspicious.
The same unusual phrase appearing in five articles suggested that one had copied another or that all five were working from the same material.
That heuristic becomes much weaker when paraphrasing costs almost nothing.
The machine does not need to copy the sentence.
It only needs to preserve the claim.
One source can now sound like a crowd
This gives us a simple progression:
one source → one interpretation → twenty publications → fifty citations → "multiple sources confirm"
Read that last phrase again.
Multiple sources confirm.
Do they?
Or did one source simply reproduce?
This distinction matters even when nobody is trying to deceive anyone.
A model summarizes a report.
Someone publishes the summary.
Another model discovers the publication and produces a shorter explanation.
A newsletter incorporates that explanation.
A search engine indexes the newsletter.
Another AI system finds the report, the summary, the explanation and the newsletter.
Four documents now appear to agree.
They should.
They are relatives.
This is the telephone game without children sitting in a circle.
And unlike the childhood version, the output of one round can become the input of the next.
The line has become a loop
The traditional telephone game has a useful property.
It ends.
A → B → C → D → E
The modern information system does not necessarily end.
It can look more like this:
source
↓
AI summary
↓
publication
↓
search index
↓
AI answer
↓
another publication
↓
another search index
↓
another AI answer
↘
back into the web
Machines read material produced by humans.
Machines increasingly read material produced with the assistance of other machines.
Machines summarize it.
Machines rank it.
Machines cite it.
Machines monitor what other machines cite.
And the resulting analyses can themselves become web pages that machines later discover.
The Internet is no longer merely a library being read by machines.
Parts of it are becoming a conversation between machines whose output is continuously returned to the library.
That does not automatically make the information false.
But it makes its ancestry increasingly important.
Errors can acquire descendants
Suppose an original report contains an important qualification:
The effect was observed in this sample, but the evidence is insufficient to generalize the result.
A summary becomes:
Researchers observed the effect.
Another summary becomes:
Research shows the effect.
A headline becomes:
Study confirms the effect.
A social media post says:
Scientists have confirmed it.
None of these transformations requires an explicit lie.
Each can be produced by ordinary compression.
The telephone game is working exactly as expected.
But now something new happens.
The later AI system searches for confirmation and discovers the headline, the social media discussion, two rewritten articles and the original summary.
Several sources appear to agree that scientists have confirmed the effect.
The lost qualification does not merely disappear.
Its absence reproduces.
Eventually the transformed version may become easier to find than the careful original.
The echo becomes louder than the voice.
And someone can do it deliberately
So far, nobody needed malicious intent.
That is important.
An information feedback loop can emerge naturally from summarization, aggregation, syndication and republication.
But once a naturally occurring mechanism exists, it can also be exploited.
This part is not new either.
Astroturfing, sockpuppet networks, coordinated inauthentic behavior, circular reporting and information laundering existed long before modern language models.
What changes is scale.
Research into AI-enabled influence operations has already identified the obvious economic shift: language models can reduce the cost of producing persuasive material, increase the scale of operations and make previously expensive forms of customized content cheaper.
Now consider a slightly different objective.
Do not begin with the sensational claim.
Begin with its future "confirmation."
Prepare a collection of apparently unrelated pages.
Give them different styles.
Different histories.
Different framing.
Different wording.
Perhaps different languages.
Let search engines discover them.
Then release the sensational claim.
A skeptical person does exactly what we have spent years telling people to do:
Don't trust one post. Check other sources.
They search.
And there they are.
Five sources.
Ten sources.
Twenty sources.
The verification procedure has been anticipated.
The evidence was waiting for the question.
Synthetic independence
This may be the more useful concept.
Not synthetic content.
Synthetic independence.
Whether a particular paragraph was written by a human or generated by a language model is often not the most interesting question.
A perfectly human-written article can repeat a synthetic claim.
An AI-generated article can accurately summarize a primary source.
Authorship alone tells us surprisingly little about informational independence.
The harder question is:
Did these sources arrive at the claim independently?
That cannot be answered by vocabulary.
It cannot reliably be answered by writing style.
It cannot be answered by counting domains.
It may not even be answered by checking whether the articles explicitly cite one another.
The dependency can be hidden several generations back.
Consider:
X
/ \
A C
| |
B D
\ /
E
E appears to rely on B and D.
Two sources.
Except B and D ultimately descend from X.
The apparent diversity exists at the document layer.
It disappears at the provenance layer.
Then reality joins the loop
There is an even harder version.
A manufactured story does not need to remain inside a manufactured network.
Suppose the initial synthetic sources attract real attention.
Real people discuss them.
Real journalists report that the claim is spreading.
Real experts are asked to comment.
Real institutions issue denials.
Real politicians react.
At that point, the original manipulation has produced genuine events.
A newspaper article describing the controversy may be completely authentic journalism.
A researcher studying the spread may produce completely authentic research.
A government statement denying the claim is a completely authentic primary document.
The information ecosystem now contains genuinely independent sources.
But they exist because the original claim caused them to exist.
The false signal has generated real evidence of its own social existence.
An AI system searching the web later may encounter a remarkably rich information environment:
news reports, commentary, official statements, social posts, analysis and academic discussion.
Everything looks substantial.
And it is substantial.
But that does not tell us whether the original proposition was true.
We have moved from a problem of misinformation to a problem of genealogy.
Google had PageRank. What comes after source count?
The early web gave search engines a graph.
Pages linked to pages.
That graph could be analyzed.
Not every link had equal value. Not every cluster of links represented independent endorsement. Search engines spent decades learning that the topology matters.
The next graph may be considerably harder.
It is not simply:
Page A → Page B
It is:
Claim X on Page A
↓
interpreted as Claim X1 on Page B
↓
summarized as Claim X2 on Page C
↓
cited as evidence for Claim X3 on Page D
The nodes are no longer merely documents.
They are claims, observations, interpretations and transformations.
And the edges are not merely hyperlinks.
They represent informational dependence.
The question is no longer only:
Who links to whom?
It becomes:
Who actually knows this independently?
That is a much harder graph to build.
Provenance helps, but provenance is not truth
There are already serious attempts to improve content provenance.
Standards such as C2PA can attach cryptographically verifiable information about the origin and editing history of digital assets. That can help establish whether provenance assertions belong to a particular asset and whether they have been tampered with.
That is valuable.
But even provenance systems deliberately stop short of declaring whether the information itself is true.
And that distinction is essential.
Knowing who created an image is not the same as knowing whether the claim attached to the image is correct.
Knowing that a document has not been altered does not prove that its author was right.
Knowing that five documents are authentic does not prove that they are independent.
We need provenance.
But provenance is infrastructure for asking better questions.
It is not an oracle.
The original source still matters
There is an old instinct that becomes increasingly valuable in this environment:
go upstream.
If ten articles discuss a scientific paper, read the paper.
If twenty sites describe a government announcement, find the announcement.
If several aggregators reproduce a public procurement notice, find the official notice.
If a quotation appears everywhere, find the original interview, transcript or recording.
Every step upstream removes a transformation.
Sometimes the chain is short.
Sometimes it disappears into a screenshot of a screenshot of a deleted post quoting an unnamed person who supposedly heard something from somebody else.
That absence is itself information.
If nobody can find the first sentence, perhaps we should be more careful about repeating the last one.
"Everyone says so" was never evidence
The uncomfortable part is that none of this is fundamentally a machine problem.
Humans are extraordinarily good at social confirmation.
We trust repetition.
We notice consensus.
We use the apparent agreement of independent observers as evidence because, under many conditions, that is entirely rational.
If five independent weather stations report the same temperature, confidence increases.
If five witnesses who did not communicate describe the same event, that matters.
If five laboratories independently reproduce a result, that matters enormously.
Independence is doing the epistemic work.
Not the number five.
Remove independence and the mathematics changes.
Five copies of one measurement are still one measurement.
A thousand retweets are not a thousand witnesses.
A hundred rewritten articles are not a hundred investigations.
And a hundred AI-generated variations of one source do not create ninety-nine additional sources.
The cost of sounding independent is approaching zero
This may be the fundamental change.
The Internet has always contained repetition.
It has always contained propaganda.
It has always contained spam.
It has always contained people pretending to be more numerous than they really are.
But producing convincing variation once had a meaningful marginal cost.
Generative systems push that cost downward.
The cost of another paragraph approaches zero.
The cost of another style approaches zero.
The cost of another translation approaches zero.
The cost of another interpretation approaches zero.
And therefore the cost of appearing independent can approach zero as well.
That changes the value of one of our most basic verification heuristics.
"Check several sources" remains good advice.
It is simply incomplete.
The better advice is becoming:
Check several independent sources.
Which immediately raises the difficult question:
How do you know they are independent?
The Internet may need family trees
Perhaps future information systems will need to think less like search engines and more like genealogists.
Not merely:
- how many documents contain this claim?
- how authoritative are their domains?
- how recently were they published?
- how similar is their language?
But:
- where did the claim first appear?
- which sources observed something directly?
- which sources merely reported another report?
- which documents share an informational ancestor?
- where did qualifications disappear?
- where did interpretation become assertion?
- which branches are genuinely independent?
In other words, the system needs a family tree.
Not of pages.
Of claims.
That is technically difficult, socially messy and probably impossible to solve perfectly.
Sources disappear.
People lie about their sources.
Private conversations leave no public trace.
AI transformations may not disclose their inputs.
The same fact can genuinely be discovered independently.
And sometimes nobody knows where an idea originated.
But imperfect provenance may still be much better than confident counting.
Back to the telephone game
Imagine playing the childhood game again.
The last person announces:
The blue elephant stole the mayor's bicycle.
Everyone laughs.
Then somebody asks the first player what the original sentence was.
Nobody knows.
The first player left.
Half the participants joined halfway through.
Some heard the sentence from two different people.
Several wrote their version on pieces of paper.
Other players copied those notes.
A machine generated twenty alternative phrasings.
Another machine grouped them as independent confirmations.
Someone uploaded the result.
A search engine indexed it.
Three articles cited the search result.
An AI summarized the articles.
And now the next player is about to whisper the sentence again.
That is no longer a children's game.
That is an information system.
So how many sources are there?
This may become one of the defining questions of the machine-readable web.
Not:
How many pages say this?
Not:
How many citations can we find?
Not even:
How many different organizations published it?
But:
How many independent paths lead back to reality?
Sometimes the answer will be five.
Sometimes it will be one.
Sometimes there will be no identifiable path at all.
And sometimes one source will have been transformed, translated, summarized, amplified, indexed, cited and rewritten so many times that it appears to be an entire crowd.
The Internet has always had echoes.
Generative AI did not invent them.
It gave them a printing press.
Now the echoes can write.
They can search.
They can cite each other.
And they can whisper the sentence back into the line.
The Internet is playing telephone with itself.
The important question is no longer whether the message changed.
It is whether anyone can still find the person who spoke first.
Ready