Not long ago, the internet quietly
passed a milestone almost nobody
noticed. For the first time in its
history, most of the traffic moving
across the web stopped coming from
people. The company that caught it is
Cloudflare, which sits in front of
roughly 1 in five websites on Earth and
watches the requests flowing into them.
Earlier this month, its CEO, Matthew
Prince, posted a single statistic that
probably deserved more attention than it
got. Automated software bots now
generate about 57% of the requests made
to web pages. [music] Actual human
beings have become the minority at
roughly 43%.
Open your laptop and nothing looks
different. Pages still load and search
still works the way it always has. The
strange part is underneath that familiar
surface in a layer you never actually
see. It's worth being precise about that
number because it's easy to misread.
Cloudflare is measuring requests to web
pages. Not every minute you spend
online, an evening of [music] streaming
or an hour lost to a scrolling feed
that's still overwhelmingly human. And
by time spent, people are very much
still here. What [music] tipped past the
halfway point is the machine readable
web, the constant background [music]
traffic that keeps everything running.
In that layer, the machines now
outnumber us. The part that makes
Prince's post interesting [music] is
that he didn't expect it this soon. Back
in March on stage at South by Southwest,
he guessed bots would overtake humans
somewhere around 2027. Then he nudged
the date earlier. Then it simply
happened. That came faster than I
predicted, he wrote, and pointed to the
cause a new kind of bot, multiplying far
quicker than anyone had modeled. To see
why that's a big deal, you have to
understand the deal the web was built
on. For about 20 years, the internet ran
on a handshake arrangement that mostly
worked. Websites made things worth
reading. Search engines crawled all of
it and pointed people [music] toward the
good stuff. And the people showed up,
which was the part that mattered most. A
visitor read your article. The ad beside
it earned its keep. And every so often,
someone bought what you were selling.
Traffic was the currency, and anyone who
made something useful got paid in it.
And then search started answering the
questions itself. If you've used Google
recently, you've watched it happen. You
type a question and before a single blue
link appears, there's a tidy paragraph
at the top written by Google's AI
telling you the answer. It's called AI
overviews. And for a lot of searches,
it's genuinely all you need. You get
what you came for and [music] you close
the tab. The Pew Research Center
measured what that does to the old habit
of clicking. They followed the real
browsing of 900 Americans across tens of
thousands of searches. When one of those
AI summaries appeared, people clicked
through to an actual website 8% of the
time. When there was no summary, they
clicked nearly [music] twice as often.
And the little source links tucked
inside the summary itself, those got
followed in about one search out of a
hundred. So, the answer arrives and the
visit never happens. For the site that
originally wrote that answer, the
traffic [music] that used to pay the
bills just doesn't show up. For
publishers, that's not a worry. It's a
hole in the accounts you can measure.
One firm tracking more than 2,500 news
sites found Google referrals down
roughly a third over the past year.
Business Insider watched its search
traffic fall by more than half and laid
[music] off a fifth of its staff. The
study site Cheg, now suing Google, said
its nonsubscriber traffic dropped 49%
[music] as AI began answering the
homework questions that used to bring
students to its pages. Some of them have
gone to court. Penske Media, the company
behind Rolling Stone and Billboard,
became the first major American
publisher to sue Google over this. And
the complaint lands the problem in a
sentence. Google has shifted from a
search engine that sends you to websites
into an answer engine that gives you no
reason to leave. Refuse to be
summarized, Penske argues, and you
essentially vanish from search, which
[music] when one company handles nearly
nine of every 10 searches, isn't much of
a choice. That is where the [music]
story takes an unexpected turn because
there's a second problem buried inside
those tidy summaries. In May, a court in
Munich looked at Google's AI overviews
and asked something nobody had really
settled. When the AI gets it wrong,
whose words are those? [music] The case
came from two German publishers whose
names the AI had casually linked to
scams and dubious business practices.
The catch was that none of the sources
it cited said anything of the kind. The
system had blended details from
unrelated companies [music] and served
the result up as fact. Google's defense
was the obvious one. These are just
summaries of what's already out there,
and people know not to trust AI blindly.
The judges weren't persuaded. A normal
search result, they reasoned, simply
points you towards someone else's page,
so Google isn't the author of it. An AI
overview is different. It rewrites
everything in its own words and reaches
its own conclusions. that the court
[music] decided makes it Google's own
statement and Google can be held liable
when that statement is false. It's one
ruling in one country and Google is
appealing. But the reasoning travels
easily. The moment a machine stops
handing you a shelf of sources and
starts telling you what's true, it
quietly stops being a librarian and
becomes something closer to a publisher.
And publishers can be sued. While the
lawyers argue, the plumbing underneath
the web is already being rebuilt for the
new kind of visitor. The whole old
bargain rested on one thing. When a
search engine crawled your site, it sent
readers back to you. AI crawlers broke
that promise. They pass through, absorb
everything, pour it into a model, and
the model answers questions without
sending a single person your way. You
give, nothing returns. So, Cloudflare
built a toll booth. They call it payer
crawl. And it leans on a dusty, almost
forgotten corner of the web's design, an
error code number 402 that has sat there
for decades labeled payment required and
gone essentially unused. Now it has a
purpose. When an AI crawler shows up, a
site can reply with that code and a
price tag. Pay and the content is yours.
[music] Don't and it isn't. Cloudflare
went further and started blocking AI
crawlers by default on new sites. So the
answer is no until an owner decides
otherwise. Strip away the jargon and
it's a quiet confession about where
things are heading. Websites are
beginning to treat machines as the real
audience, not visitors to court, but
software to charge at the door. Which
leads somewhere obvious. If software is
reading the web, when does it start
using it? That's already underway. The
big AI companies have built what they
call agents programs. You hand a task to
[music] and they go off and complete it.
Book the flight. Track down the cheapest
version of this jacket and buy it.
Openai's agent moves through real
websites the way a person would, right
down to clicking the buttons and filling
in the forms. And it's been connected to
checkout systems so it can actually pay.
Visa signed on to let these agents spend
real money within limits you set in
advance. The entire pitch is that you
stop browsing and let the thing browse
for you. And here's a detail worth
holding on to. A person shopping for a
jacket might check five sites. An agent
doing the same errand can hit thousands,
generating real traffic and real server
load while never once seeing an ad or
becoming a customer in any way the old
web understood. This is also where it
gets genuinely strange. A team of
security researchers ran an experiment
this year. They built a fake website and
invited seven of these AI agents to use
it alongside real people. Then asked a
simple question. Can you tell them
apart? Their tool flagged all seven.
CloudFlare's standard bot detection,
given the same test, caught one. The
reason is that these agents look nothing
like the clumsy bots of the past. They
run inside ordinary browsers and do
ordinary things. What gives them away
isn't an obvious tell. It's the texture
of how they move. A real person is a
little messy about it. The cursor
wanders, a typo gets fixed, the
scrolling comes in, fits, and [music]
starts. The agents aren't messy. They
slide straight to the button and drop in
the text in one clean motion. They're
betrayed by being a fraction too smooth.
Now, run that forward a year or two. A
typical website's traffic looks
perfectly healthy. Someone is reading
the reviews and adding things to a cart.
Except the someone is a piece of
software shopping for a person who never
laid eyes on the page. The visit is
[music] real. The visitor is not. This
is where the story gets darker because
the same technology that can browse the
web for us can also invent things that
were never real. The clearest place to
see the damage is in [music] schools. At
a high school in Westfield, New Jersey,
teenage boys took ordinary photos of
their female classmates, ran them
through an app that fabricates nude
images, and passed the results around.
At a middle school in Beverly Hills,
students did the same to 16 of their
peers, and the school expelled five of
them. These aren't experiments. in some
hidden corner of the internet. They're
happening in ordinary towns to ordinary
kids with apps that sit a few taps away.
The scale once you go looking is hard to
sit with. In a study spanning 11
countries, UNICEF, Interpol, and the
child protection group ECPAT found that
at least 1.2 million children reported
having their images turned [music] into
sexual deep fakes within a single year.
In some places that came out to one
child in 25, about one in every
classroom. And the Internet Watch
Foundation, [music] which has spent 30
years tracking this material down,
called 2025 its worst year on record,
driven by realistic AI video that often
stitches in the faces of real
identifiable children. There's an
instinct to shrug this off because the
images aren't real. UNICEF answered that
headon. Deep fake abuse, they said, is
abuse. And there is nothing fake about
the harm it causes. [music] When a real
child's face is used, that child is a
victim. Full stop. The picture may be
synthetic. The fear and the humiliation
that trail the person afterward are not.
Step back from all of it and one shape
comes into focus. The internet isn't
dying. That's the wrong word. And the
chatter about a dead internet misses
what's actually happening. The web is
busier than it has ever been. It's just
that more and more of it gets handled by
machines before it ever reaches a human
being. A bot makes the request. An AI
writes the answer back. An agent does
the shopping. And somewhere in that same
machinery, a generator builds a face
that never [music] belonged to anyone.
The people haven't gone anywhere. We're
simply drifting further from the raw
material. The path between asking a
question and receiving [music] an answer
used to run through other people, their
writing, their work, their actual pages.
Now, most of that path is handled by
software that reads it first, decides
what it means, and sometimes makes
things up [music] before any of it
reaches you. We get handed the finished
product and asked to trust it. So,
here's the question worth sitting with.
Maybe this is just what it looks like
for a technology to grow up. The web
finally getting a layer of automation on
top, the way nearly every messy human
system eventually does. Or maybe we're
watching something quieter and harder to
name. the human web, the one built by
people for people, slipping into the
background while the machines step to
the front. I honestly don't know which
it is. I'm not sure anyone does yet. Let
me know what you think down in the
comments. Does this feel like progress
or like something we're losing? [music]
I read them and the best conversations
on this channel happen there. Also, if
you want more content around science,
space, [music] and advanced tech, we've
launched a separate channel for that.
Links in the description. Go check it
out. If this gave you something to chew
on, hit subscribe so you don't miss the
next one. Thanks for watching and I'll
catch you in the next one.