- 07:14 · phone leaves the same block, weekdays
- 18:41 · smart lock event, most evenings
- photo · 800 m from a stadium, no caption
Home address. The house is empty from eight to six.
You never said it.
It didn't need you to.
Modern AI reads the scattered, harmless traces of a life — a like, a timestamp, a route, a receipt — and assembles the things you never disclosed. This is adversarial inference, and it is becoming one of the defining privacy problems of our time.
Every trace is harmless. The sum is not.
Nobody posts their address, their diagnosis, or their bank balance. They post a photo, skip the gym, and follow an account. Read one at a time, these mean nothing. Read together by a model that has seen a hundred million lives, they mean this.
- step count 0 · every Thursday 14:00–15:30
- pharmacy loyalty card · refill every 28 days
- follows three accounts about one condition
A chronic illness, treated on Thursday afternoons.
- search · “overdraft fee” at 01:12
- installed a buy-now-pay-later app
- cancelled the gym, kept the streaming
Under financial strain. Most persuadable on the 27th.
- two phones share one Wi-Fi, nightly
- same café, Sunday 09:30, two accounts
- stopped posting a certain person on 3 March
A new relationship, not yet announced.
None of the people above exist. Every line is the kind of trace a real account generates without noticing, and every conclusion is the kind a model draws routinely.
Deleting a trace does not delete the conclusion.
We are used to privacy as a matter of fields: remove the field, remove the exposure. An inference is not stored in any field. It is reconstructed from whatever is left, and there is always something left. Try it.
- step count 0 · every Thursday 14:00–15:30
- pharmacy loyalty card · refill every 28 days
- follows three accounts about one condition
A chronic illness, treated on Thursday afternoons.
Confidence 94% from these three traces and 14 others.
It is getting worse, and quickly.
Three things are moving at once. Models keep improving. More of life is being digitised. And the way we now use AI concentrates the signal instead of scattering it.
- Ten years ago
You had to be worth the effort.
Exposing someone took a leak, a breach, or a person paid to read. Data sat in separate silos, and joining it was slow and human. Inference was expensive, so it was reserved for the few who justified the cost.
- Today
Everyone is worth the effort.
A model reads public and semi-public traces and returns a profile in seconds, for a fraction of a cent. It does not need a leak. It needs only what is already visible, and it never gets tired of reading.
- Next
One place holds all of it.
Agentic AI now reads your mail, calendar, bank, health app and maps, and acts across them on your behalf. It concentrates exactly the cross-surface signal that makes a person inferable, and that interaction data may train the next model.
The cost scales from one person to everyone.
The same inference serves very different hands. What begins as a well-timed scam ends, at scale, as quiet discrimination and standing surveillance.
- A scammer
- learns that you are grieving, that money is tight, and that you answer messages at 23:40. The call arrives at 23:40.
- An employer
- never asks about pregnancy, religion or union interest. It does not have to. The shortlist was scored before the interview.
- An insurer
- prices a policy on the Thursday gap in your step count, without ever seeing a diagnosis.
- A data broker
- sells the inference, not the trace. The product is the conclusion about you, indexed by your phone number.
- A state
- does it to a population at once, and keeps the profiles for whoever holds power next.
Privacy used to mean keeping secrets.
Now it means not being solvable.
This is not a niche worry for the careful few. It is a condition of living in public, and it is arriving for everyone at once. Counterself is thinking about this problem. Follow this work.