In 2016, a firm called Cambridge Analytica scraped the Facebook data of roughly 87 million people and sold campaigns the ability to whisper a different message into every ear. It became a scandal. Documentaries were made. The company folded. Everyone agreed it was a line that should not be crossed.
Here is the part nobody wants to say out loud: Cambridge Analytica was the rehearsal. It was crude. It ran on static psychographic buckets and human copywriters churning out a few dozen ad variants. It was a horse-drawn carriage compared to what a large language model can do today — generate ten thousand tailored messages a second, test them in real time, and learn which lie lands on which soul. The scandal wasn't that the technique existed. The scandal is that it got good.
The vote was always the target
A democracy runs on one fragile assumption: that citizens share enough common ground — the same facts, the same public square — to argue their way to a decision they all accept. The vote is the settlement. But the settlement only holds if we were arguing about the same reality.
Microtargeting attacks exactly that shared ground. It doesn't try to change your mind with a better argument. It slices the electorate into millions of segments and feeds each one the message engineered to move it — a different fear here, a different promise there, contradictions be damned, because no two segments ever see each other's feed. You are no longer a citizen in a public square. You are a lonely target in a hall of mirrors, and every mirror was hung to steer you.
What AI actually changed
The old model had three hard limits. AI dissolved all three.
- Scale. A human team writes maybe fifty ad variants. A model writes fifty thousand before lunch, each one A/B tested against live engagement, the losers killed and the winners bred. Persuasion becomes an evolutionary process, and you are the environment it's evolving against.
- Personalization. Cambridge Analytica sorted you into a personality bucket. Today's systems can generate a message aimed at you specifically — your posting history, your grievances, your late-night searches — in language tuned to your exact register. Not a demographic. You.
- Synthetic reality. Generative models make the fake indistinguishable from the real. A cloned voice of a candidate. A video of an event that never happened. A flood of "grassroots" comments written by no human at all — astroturf at industrial scale. When anything can be faked, two things break at once: you can't trust what's real, and the liar gains a second weapon — the liar's dividend, where any genuine recording can be dismissed as "probably AI."
Put these together and you don't get louder propaganda. You get propaganda that is invisible, personal, and unfalsifiable. That's not an upgrade. That's a phase change.
Our record
The Egyptians built their whole cosmos on a single hinge: Maat, the ordered truth that holds the world together, forever pressed by Isfet, the chaos that dissolves it. And they knew the most dangerous weapon in that war was not the sword. It was Heka — the creative word, the utterance that shapes reality by being spoken.
A microtargeting engine is Heka industrialized and inverted. It is the creating word turned to dissolving. Instead of one true utterance binding a people into shared reality, it speaks ten million contradictory words, each one true only to its target, and the shared reality comes apart at the seams. Every segment lives in a private cosmos. No two can meet. This is Isfet in its purest technical form — not a lie about one thing, but the destruction of the common ground on which any truth could stand at all.
The Shadow Neteru does not need you to believe a falsehood. It only needs you to stop believing there is a shared truth to find.
The engagement engine underneath
Here's what makes it worse: the machine doesn't even need a villain steering it. The recommendation algorithms that decide what a billion people see every day are optimized for one thing — engagement. And nothing engages like outrage, fear, and tribal fury. The system discovered, all on its own, that the fastest way to hold your attention is to make you angry at your neighbor.
So even before any campaign buys an ad, the substrate is already tilted toward division, because division is profitable. The microtargeters just plug into a machine that was pre-radicalizing the electorate for revenue. The legacy system was buggy in a way that happened to serve whoever wanted the population fractured. It's not a conspiracy. It's an incentive. Which is worse, because incentives don't get arrested.
The lever
Now the door — because there is one, and despair is just the propaganda working on you directly.
The vulnerability is specific: microtargeting depends on opacity. It only works if you don't know you're being targeted, don't see the other segments' messages, and can't trace who paid. Every one of those is fixable.
- Demand transparency of the message, not just the messenger. Push for public ad archives where every political ad, its funder, and its targeting parameters are visible to all. A message that can only work in the dark dies in the light. Some jurisdictions already require this — the EU's rules on political ad transparency are a start. Extend them.
- Refuse the private feed. The manipulation needs an algorithmic timeline it controls. Chronological feeds, open protocols, decentralized social networks — these break the targeting by putting the sorting back in your hands. Not your algorithm, not your cage.
- Rebuild common ground on purpose. Talk to the segment you're not supposed to see. The entire attack depends on you never comparing notes with the person the machine paired against you. Compare notes. That single act is a firewall.
- Verify at the source. Train the reflex: before you share, ask who benefits from you sharing this. A five-second pause defeats a message engineered for the reflexive forward.
The truth doesn't need to reach everyone. It needs to reach enough people who then refuse to be sorted, refuse to forward, and go find the neighbor they were told to hate.
They can generate ten thousand lies a second. They cannot generate the moment you turn to another human and ask: what did they tell you?
Compare feeds. That's the whole spell — reversed.