> ## Content Index
> Fetch the complete content index at: https://journal.shiroand.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Before You Choose, Something Has Already Chosen What You Can See
- URL: https://journal.shiroand.io/before-you-choose-something-has-already-chosen-what-you-can-see/
- Published: 2026-10-22T19:29:33.000Z
- Updated: 2026-10-22T19:29:33.000Z
- Author: Kosuke Shirako
- Tags: Observations, Essays

## A simpler way to think about algorithmic accountability

When we talk about algorithms, we often start with a difficult question:

**Why did the system make this decision?**

Why was this job recommended?

Why did this product appear first?

Why did one piece of news reach me while another did not?

These are important questions.

But they are also difficult to answer.

Modern recommendation systems are complicated, constantly changing, and often inaccessible from the outside. Even when we can inspect parts of a system, reconstructing exactly why one result appeared at one particular moment may be impossible.

So perhaps there is another place to begin.

Not with what the algorithm was thinking.

But with what the person was actually allowed to see.

## Choice begins before the click

Imagine two people opening the same job platform.

They type similar searches.

They live in the same city.

They may even have similar qualifications.

But the jobs shown to them are not necessarily the same.

One person may see AI engineering positions.

Another may see cybersecurity roles.

A third person may see something completely different.

Each person feels as though they are choosing freely.

And they are choosing.

But they are choosing from different visible worlds.

That distinction matters.

Because before we choose between A, B and C, some system has already decided that A, B and C are the options worth putting in front of us.

The invisible part is everything we never saw.

## The problem with looking only inside the algorithm

Much of algorithmic accountability has focused on the black box.

Can we inspect the source code?

Can we understand the model?

Can we identify bias in its internal logic?

These approaches are valuable.

But in practice, they can run into serious limits.

The software may be proprietary.

The model may change continuously.

Its behavior may depend on thousands of variables.

And the exact state that produced a particular recommendation may disappear before anyone investigates it.

This led us to a different question:

**What if we audit the visible world instead?**

## Treat the screen as evidence

This is the basic idea behind the Visible World Audit.

Instead of claiming to know why an algorithm produced something, we begin by recording what actually happened.

What appeared?

In what order?

What changed?

What happened immediately before the change?

Did two controlled users see different things?

And can we preserve that evidence so someone else can examine it later?

The distinction sounds small.

It is not.

It changes the audit from:

**“We think the algorithm did this because…”**

to:

**“This is what appeared, at this time, under these conditions.”**

That second statement is much easier to defend.

## Observation is not explanation

One of the most important disciplines in this approach is knowing what we do not know.

Suppose a recommendation feed changes shortly after someone opens an email.

We can observe:

- the email was delivered
- the email was opened
- the feed later changed

But we should not automatically say:

**The email caused the feed to change.**

That may be true.

It may also be false.

There may be another variable we cannot see.

So the Visible World Audit separates evidence into categories such as:

**Observed** — what was directly recorded.

**Stated** — what the platform itself explicitly disclosed.

**Inference** — what we think the evidence may mean.

**Unknown** — what cannot be established from the available evidence.

That last category is important.

A good audit should not eliminate uncertainty by pretending it does not exist.

It should preserve uncertainty where uncertainty is real.

## Why disappearing interfaces matter

Digital environments are unusually difficult to investigate because they do not stay still.

A personalized feed today may be different tomorrow.

A recommendation may move.

A model may be updated.

A cache may refresh.

A page may quietly change after another interaction.

Weeks later, someone may ask:

**What exactly did this person see?**

By then, the answer may no longer exist.

So another part of the Visible World Audit is preservation.

The purpose is to create a frozen record of a temporary digital environment: what appeared, where it appeared, when it appeared, and what sequence of interactions surrounded it.

In other words, a feed that normally disappears becomes something that can be revisited.

## A small experiment

During internal validation, we tested the approach using controlled synthetic personas in a synthetic recommendation environment.

One persona interacted with AI-related jobs.

Another interacted with cybersecurity-related jobs.

A neutral persona did nothing.

Over time, the two active personas developed different recommendation environments, while the neutral control remained unchanged.

The experiment does not prove that a particular real-world platform discriminates.

It was not designed to.

What it demonstrates is something narrower:

**different patterns of interaction can be measured as different visible environments.**

And those environments can be compared without claiming to know the hidden logic that produced them.

## Why this matters beyond AI

This question is not limited to recommendation engines.

It appears anywhere systems determine what becomes visible.

Recruitment.

E-commerce.

Advertising.

News.

Search.

Entertainment.

Finance.

Education.

Even public information.

In all of these environments, there is a difference between:

**what exists**

and

**what becomes visible to a particular person.**

As personalization becomes more powerful, that gap may become increasingly important.

## The missing question

For years, we have asked:

**Is the algorithm fair?**

That question will remain important.

But it may be too large to answer all at once.

A more practical question might come first:

**What did the system actually show?**

And then:

**Did it show something different to someone else?**

Before we judge the entire intelligence of a system, we can preserve the evidence of the world it produced.

That is the shift I find most interesting.

Algorithmic accountability may not always begin by opening the black box.

Sometimes it may begin by carefully recording what appeared on the screen.

## One question to keep

The next time a platform presents you with a list — jobs, products, videos, people, articles, opportunities — it may be worth remembering:

**You are choosing from what became visible.**

And visibility itself is increasingly a decision made by systems.

That is why the architecture of what appears deserves to be observed.