- No ads, no sponsors, no affiliate links
- No paywall, no email gate, no account
- Free to share and translate, with credit, under CC BY-NC-ND 4.0
The Internet works because we trust it to work.
We trust that new web address takes us to the intended site.
But someone has to keep the map and tell the truth about it.
We trust that the Chrome lock icon means we reached the real website.
But that confidence depends on a chain of entities that don't all account to one another the same way.
We trust that the information we send will travel where it is supposed to go.
But many eyes can view many parts while it travels. And what they see is more than we admit.
Most days, most trust holds. When it does not, the failure can spread far and fast. Headlines don't make that clear and news cycles are short.
That feeling that more is happening behind the scenes is appropriate. This system looks more capable, more coordinated, and more certain than it really is. It becomes more dangerous when AI systems make important decisions, some more subtle than we are likely to notice.
A technology we don't understand
The people who build the most powerful models can describe how they are trained. They cannot tell you why one answers the way it does. We are shipping a tool whose internals are, in a literal sense, unknown to its makers. Explainable AI is, at best, a thing to try and achieve but we keep making newer bigger and more complex systems before we understand the existing systems.
Words go in. Words come out. The middle is a few hundred billion numbers nobody can read.
A means of processing information we cannot remain in control of
When the machine writes, summarizes, decides, and remembers for you, the thinking moves out of your head. The output is fast, fluent, and constant. Your judgment does not get faster to match it.
You read at the speed of a human. It writes at the speed of a power plant.
A record of us, assembled in the dark
No single piece of data about you says much. A location ping. A card swipe. A search at 2 a.m. The danger is aggregation: data sets collected separately get joined, and the joined record gets read for things you never said. Where you sleep. Where you pray. What you are sick with. How you vote.
Anonymous does not survive the join. In one study of 1.5 million phone users, four points of place and time were enough to single out 95 percent of them.
None of this waited for AI. Love Edward Snowden or hate him, the documents he took in 2013 showed the NSA collecting the phone records of millions of Americans, a program a federal appeals court later found unlawful. Think what you like of Julian Assange, WikiLeaks published documents detailing the CIA's tools for breaking into phones, computers, and televisions. The gathering was well under way long before any machine could read it all.
Some of those programs were shut down. The market that feeds them never was. In 2023 a declassified report from the Director of National Intelligence acknowledged that intelligence agencies buy Americans' data from commercial brokers, data that "can reveal sensitive and intimate information."
Joining used to take people: analysts, time, and a reason to look. AI removes all three. Today's leading models can take in hundreds of thousands of words at once and connect things across all of it in seconds, the kind of connections a person would need weeks to make, if they made them at all. In one 2023 study, language models read ordinary Reddit posts and worked out where people lived, what they earned, and their sex, with up to 85 percent accuracy, about 240 times faster than people doing the same job.
And it does not have to tell you what it connected. When Harvard researchers looked inside one chatbot, they found it quietly keeping its own guesses about the person typing: age, gender, education, socioeconomic status. It never said so. They found it only because they went hunting. We can only hunt for what we know to look for, with the tools we know how to build, and every bigger, more complex model leaves more places we have not looked.
Now the old record, and everything added since, can be joined by systems whose reasoning no one is required to show you. The opacity is not a flaw in the method. It is the method.
Nobody has to watch you anymore. They only have to keep everything until something can read it. Now something can.
All together, Now What?
Taken one at a time, each of these is a problem someone is working on. Taken together, they compound. A fragile network carries a technology nobody can explain, producing more than anyone can read, feeding a record of us that never stops growing.
We build it anyway. The last exhibit puts all four in one room and asks what we do about it.
Make complex systems teachable, usable, and defensible.
Jarrett Heintz is a cybersecurity educator and applied ethics consultant with 20+ years across cybersecurity, AI, risk, governance, and evidence-based learning design. He has partnered with Fortune 500 companies, government agencies, and international law enforcement, and has guided thousands of professionals through CISSP, CISM, CCSP, CRISC, ISSMP, and CEH.
This interactive documentary applies that background to the question underneath the slop: who is in control of the information, and for how long.