The Algorithm Knows Who You Were Standing Next To

What TikTok taught me about a new kind of privacy—and why the same problem may be coming for artificial intelligence

I have developed a theory about TikTok.

I cannot prove it.

My theory is that physical proximity matters more to recommendation algorithms than we understand—that geolocation, network information or some combination of digital signals generated when two people spend time together can eventually influence what appears in their respective information environments.

I started wondering about this because of something stranger than receiving eerily accurate advertisements.

Sometimes, after spending significant time with someone, my TikTok feed seems different.

A new subject appears.

A particular kind of video begins circulating.

Something enters my algorithm that doesn't seem to belong to me.

And occasionally, looking at what has suddenly appeared allows me to infer something about the person I have just been around.

Not something they told me.

Something private.

That observation raises a question considerably more interesting than whether TikTok is somehow listening to our conversations:

Can an algorithm accidentally tell you someone else's secrets?

What TikTok Actually Knows

It is important to distinguish what can be documented from what I suspect.

TikTok's current U.S. privacy policy says the platform automatically collects substantial information about its users, including interactions with content and advertisements, duration and frequency of use, engagement with other users, IP addresses, mobile carriers, device information, network type and approximate location derived from device and network information.

With permission, it can collect contacts. Depending on a user's settings, it can also collect approximate or precise device location.

TikTok also receives information from advertisers and other partners about activity occurring outside its own app.

The company does not publicly say that it notices two people sharing a cell tower and then uses that fact to cross-pollinate their For You feeds. I have found no evidence establishing that mechanism.

But the distinction may ultimately make my original theory less interesting rather than more.

Because an algorithm may not need to know explicitly that two people were together to infer relationships between their worlds.

The Algorithm Doesn't Need the Secret

Imagine two people have lunch.

One of them is privately considering moving to Japan.

She hasn't told the other person.

But perhaps for weeks she has been watching Tokyo apartment tours. She has searched for Japanese-language videos. She watches airline content. She follows expatriate creators. She pauses on videos about Japanese neighborhoods.

Those behaviors create a pattern.

Now the two people spend several hours together.

Perhaps they share a location. Perhaps they have mutual contacts. Perhaps they photograph the same restaurant. Perhaps they search for something they discussed. Perhaps one sends the other a link. Perhaps their subsequent online behavior changes in tiny ways neither consciously notices.

There are potentially hundreds of signals.

Then something peculiar happens.

Japan begins appearing in the other person's feed.

Tokyo apartments.

Japanese lessons.

Airlines.

Expatriate life.

No notification appears saying:

Your friend is considering moving to Japan.

The algorithm hasn't necessarily disclosed anything at all.

Instead, it has altered someone's information environment.

And the human being does the rest.

She thinks:

That's strange.

Then:

Why am I suddenly seeing this?

And finally:

Is this about her?

The machine provides the clues.

The human performs the intelligence analysis.

Privacy After Recommendation

We tend to imagine digital privacy as a question of possession.

Who has my data?

Who can read my messages?

Does an application know my location?

Who bought my browsing history?

Those remain important questions.

But machine inference introduces another one:

What can other people learn about me from the information algorithms choose to show them?

Researchers studying recommendation systems have warned for years that personalization creates privacy problems precisely because recommendations can be constructed from sensitive social relationships and inferred attributes.

The problem becomes particularly interesting when the system never explicitly reveals the underlying information.

A recommendation can itself become evidence.

And evidence can be combined with human knowledge.

Your friend knows where you went last weekend.

Your colleague knows that you have been unhappy at work.

Your sister knows whom you are dating.

Your neighbor knows which car is usually outside your house.

The algorithm knows something entirely different: thousands of behavioral correlations derived from millions of users.

Neither possesses the whole picture.

But put the machine's recommendations in front of a human who possesses contextual knowledge and suddenly the combination may reveal something neither could have discovered independently.

That is a different privacy problem.

It is inferential privacy.

The World's Largest Accidental Intelligence Network

This makes recommendation algorithms begin to resemble intelligence systems.

Traditional intelligence collection seeks information.

Who met whom?

Where did they go?

What did they purchase?

What changed?

What are they preparing to do?

But increasingly, the valuable product isn't a single intercepted message or secret document.

It is inference.

Ten thousand individually insignificant signals can collectively suggest something consequential.

A satellite image changes.

Shipping patterns change.

Someone purchases unusual equipment.

A government official cancels a trip.

Aircraft move.

Online rhetoric changes.

Financial transactions behave strangely.

None proves anything.

Together they might.

The intelligence analyst's job is to determine whether the pattern represents signal or coincidence.

TikTok performs a vastly more mundane version of this operation every time it constructs a For You feed.

It observes signals.

It predicts relevance.

It ranks possibilities.

Then it decides what deserves your attention.

The Most Powerful Decision May Be What You Never See

This is where recommendation algorithms become philosophically interesting.

Suppose an algorithm considers one million possible pieces of content and selects 100 for you.

You choose what to watch.

Technically, you remain completely free.

But the machine made another decision first.

It determined the 999,900 things from which you would never choose.

The same problem becomes considerably more serious when artificial intelligence enters government, medicine, finance, policing or intelligence analysis.

Imagine an intelligence system examining one million signals overnight.

In the morning, it tells an analyst:

These seven deserve your attention.

The analyst examines all seven and makes the final judgment.

We might describe that system as maintaining a "human in the loop."

But where was the consequential decision actually made?

The human decided what the seven signals meant.

The machine decided which 999,993 signals the human never saw.

That is not merely automation.

It is control over attention.

When Machines Know Things They Cannot Explain

There is an even stranger possibility.

A sufficiently complicated machine-learning system can produce a useful prediction without giving its human user anything resembling a conventional chain of reasoning.

The system may effectively say:

Look here.

Why?

Because thousands of variables interacting across enormous datasets make this particular thing statistically unusual.

Sometimes it will be right.

Sometimes it will be wrong.

Sometimes it will discover relationships a human would never have noticed.

And sometimes a human will discover meaning in a coincidence because the machine placed the coincidence directly in front of her.

That creates an epistemological problem we are only beginning to confront:

How do you distinguish an extraordinary machine inference from an extraordinary machine coincidence?

The Algorithm and the Analyst

This question extends far beyond TikTok.

Artificial intelligence increasingly operates between human beings and overwhelming quantities of information.

The machine doesn't necessarily make the final decision.

It ranks.

Filters.

Prioritizes.

Recommends.

Flags.

Predicts.

Those verbs sound less dramatic than decides.

But they may describe something just as powerful.

Because whoever determines what receives human attention helps determine what humans ultimately believe.

This is particularly important in intelligence.

For most of history, intelligence organizations confronted scarcity. The challenge was obtaining information that adversaries wanted hidden.

AI creates the possibility of the opposite problem.

Information becomes abundant.

Human attention becomes scarce.

The central question therefore changes.

It is no longer merely:

What does the machine know?

It becomes:

What has the machine decided that I should know?

The Human Is Still Part of the Algorithm

There is one final twist.

My TikTok theory may be wrong.

Perhaps the people I spend time with have absolutely nothing to do with the strange changes I sometimes perceive in my feed.

Humans are ferocious pattern-recognition machines ourselves.

I may notice the uncanny recommendations and forget the thousands that mean nothing. Once I suspect a connection, confirmation bias makes additional examples easier to notice.

But that doesn't eliminate the phenomenon.

It completes it.

Because the modern recommendation system isn't simply software.

It is software interacting with a human mind.

The algorithm ranks information.

The human interprets it.

The algorithm produces another recommendation based partly upon that behavior.

The human interprets again.

Machine inference and human inference begin feeding one another.

And somewhere inside that loop, knowledge can emerge that nobody explicitly disclosed.

Perhaps that is the privacy question we should be asking as artificial intelligence becomes increasingly embedded in everyday life.

Not simply whether machines know too much about us.

But whether machines can arrange the world around us so effectively that we begin knowing too much about one another.

And whether, when that happens, we will even know where the knowledge came from.

Sources and verification note

TikTok's July 2026 U.S. privacy policy says it collects usage, device, network and approximate-location information; with appropriate permissions/settings it can also collect contacts and more precise location information. TikTok also says it receives certain off-platform activity information from advertising and measurement partners. TikTok's 2026 Local Feed explicitly uses location as one factor in recommending nearby content. These documented practices do not establish that TikTok uses shared cell towers or physical co-presence between two particular users to influence their respective For You feeds. That remains the author's hypothesis and personal observation.

Research into recommender systems has separately demonstrated longstanding privacy concerns involving social recommendations and the inference of private attributes from user data.

Author's Disclosure

This essay was developed with the assistance of AI as an editorial, organizational, research, and critical analysis tool. AI was used to test the author's original hypothesis against publicly available information, distinguish documented practices from speculation, and assist with drafting and refinement. The central observation, hypothesis, argument, and conclusions remain the responsibility of the author.