Mid-to-Late Stage Summary of a Small - Human Observation Project

I. “Small” — The Birth of a Human Observation Project

1.1 Project Background

Recently, the company has been promoting Generative AI. However, compared with traditional AI (such as speech recognition, computer vision, and predictive algorithms), these technologies currently lack clear real-world application scenarios.

More often, their promotion is driven by a mindset of “others have it, so we must have it too—even if it’s useless, we can’t afford not to have it.” That is: technology comes first, and use cases are searched for afterward. Frankly speaking, this is no different from earlier waves such as Big Data, the Metaverse, or Blockchain.

As a result, I have been more consciously (and sometimes unconsciously) trying to use various consumer-facing AI tools (or AI tools without specific scenarios).

One day, when I tried to transform my own appearance into that of a woman using AI, I noticed that the result was very different from the typical “internet influencer face” commonly seen online.

This made me wonder: could I create a virtual social media account to increase brand exposure for the company (and enhance users’ impression of us)?

1.2 Choosing Xiaohongshu (RED)

When selecting a platform, I considered Weibo, Xiaohongshu (RED), Douyin, and others—platforms with suitable content formats and user scale.

In the end, I chose Xiaohongshu.

The main reasons were:

1. Xiaohongshu’s user demographics are highly aligned with our company’s customer base.
2. Coincidentally, my original Xiaohongshu account had been banned because I posted a photo of my son without clothes (all historical posts became inaccessible).
3. AI is currently unable to generate videos with high consistency in human appearance.
4. Weibo lacks strong interactivity and functions more like an official announcement board.

1.3 Vertical Content Selection

Because AI-generated human images struggle to accurately present products or product details, it is not feasible to produce direct product recommendation or store exploration content.

Therefore, the content had to focus on the simplest form: daily life sharing.

1.4 Data and Sample Bias

The data is based only on users who have already followed me or interacted with me (and Xiaohongshu does not expose detailed analytics).

Additionally, Xiaohongshu has its own interest-based recommendation algorithm, so I cannot observe how all users react to my content.

I can only analyze the reactions of users who are already interested in the content I produce.

II. Observing the Role of AI

2.1 AI Capabilities and Suitable Work

2.1.1 Definition of AI in This Study

There are many types of AI: game NPC interaction, industrial robotic arms, video recognition, audio recognition, etc.

Here, I refer only to Generative AI, namely text-generation and image-generation models.

2.1.2 Capabilities of AI

From my personal perspective, AI’s capabilities can be summarized as:

It cannot satisfy requirements that demand precise output, but it performs well for needs with no strict accuracy requirements.

Many “one-click” website, app, or game generators online illustrate this well: the results are often crude or highly homogenized.

Even for simple image or text generation, AI cannot create content it has never seen or does not know.

Therefore, unless the user’s description is extremely precise or the requirement is very simple, AI is often incapable of truly meeting practical work needs.

2.1.3 Work AI Cannot Handle

For example, tasks like finance or auditing, which require extremely high numerical accuracy, cannot be handled by AI at all—it can only serve as a reference.

For roles such as programmers or engineers, where inputs and outputs are clearly defined, AI can handle part of the workload.

However, the biggest problem is this: code is generated in seconds, bugs are fixed for years.

While it appears to accelerate output, the resulting quality is often disastrous.

2.1.4 Work AI Can Handle

Non-accuracy-critical content production.

Such as low-level music creation, writing, or illustration—fields where there is no absolute notion of “right” or “wrong”.

AI is well suited here.

Not because AI produces outstanding content, but because it can generate content quickly and at scale, after which humans can filter and select usable results.

2.2 Application of AI in This Project

AI was used in only two aspects:

1. Image generation
2. Copywriting generation

2.2.1 Image Generation

One major goal of this project was to explore how well people can identify AI-generated images.

I observed a very interesting phenomenon:

The human brain tends to accept coherent narratives as real—even if individual elements appear fake.

Many images were obviously AI-generated at first glance, yet people still accepted them as “real” because I posted a mix of real and fake content over time.

People reasoned: since some content is real, then everything must be real.

Even more strangely, one person initially pointed out that the images were AI-generated—but after seeing more posts, they kept asking me how I trained my body.

2.2.2 Copywriting Generation

AI-generated copy is either too stiff or too obviously AI-generated.

I was also unwilling to spend time crafting prompts or examples.

As a result, I stopped using AI for copywriting altogether.

III. Xiaohongshu Recommendation Mechanism

I originally wanted to elaborate more on this section.

However, Xiaohongshu seems to have many operating entities, and I do not have sufficient data to support strong conclusions.

So I will only describe some observable characteristics.

3.1 LBS (Location-Based Services)

Xiaohongshu assigns significant weight to LBS.

Most followers are from the same city; very few are from outside.

If no location is tagged, exposure drops significantly.

Interestingly, when tagging universally shared locations such as “subway” or “train station”, exposure spikes sharply.

Personally, I suspect this might be a bug.

3.2 Recommendation Recall Mechanism

During the early cold-start phase, recommendation weight seems particularly critical.

If initial viewers do not like the post, exposure drops sharply.

3.3 Approved on the Surface, Rejected in Practice

Xiaohongshu reviews posts before publishing.

However, some posts marked as “approved” do not appear in self-recommendation lists and receive almost no traffic.

3.4 Riding Trends

When a topic becomes popular, adding relevant tags or keywords does help to some extent.

But in practice, this is difficult.

For example, how many people actively browse posts about Sushiro?

Or topics like “Xiao Luoxi”, which tend to get banned very quickly.

However, tags like marathons, celebrity names, or scenic spots can capture some search and recommendation traffic.

IV. Observations of People

Based on my content niche, I mainly encountered three types of people:

1. People looking for a partner
2. People looking for one-night stands
3. People who just want to see attractive women

As mentioned earlier, my sample is biased. The following summarizes only those I personally interacted with and does not represent everyone—only certain social phenomena.

4.1 People Looking for a Partner

Many directly send personal information such as age, height, income, and residence.

When asked why they look for partners online, they usually say their social circles are small and they cannot find partners in real life.

This is indeed true.

4.1.1 Work Environments That Limit Social Interaction

For jobs involving little external communication and long working hours (often high-paying roles), coworkers are mostly of the same gender—such as finance, programmers, kindergarten teachers, or school teachers.

When the workplace is dominated by the same gender, free time is limited, and the environment is fixed, it becomes extremely difficult to meet the opposite sex.

4.1.2 Lack of Personal Hobbies

Many young people today were raised with goal-oriented education—study hard, get a good job, make money.

They had neither time nor interest to develop personal hobbies.

4.1.3 Lower-Income Groups Face Less Pressure

Lower-income groups tend to have higher social mobility and meet more people across different work environments.

Many of these relationships lack grand expectations for the future.

As a result, the pressure is lower—people simply want companionship, mutual support, or to enjoy the present together.

4.2 People Looking for One-Night Stands

This part was genuinely disturbing.

While I understand that people use platforms for different purposes, some individuals directly sent messages stating their genital size.

That was extremely disgusting.

There are even rumors that on other platforms, people directly send explicit photos.

Even those who were not explicit at first often revealed their intentions quickly, repeatedly pushing to meet offline—their motives were obvious.

4.3 People Who Just Want to See Attractive Women

These individuals were relatively harmless.

They rarely sent disturbing private messages, though some comments were inappropriate.

Personally, I consider that acceptable.

4.4 Summary of the Xiaohongshu Platform

Overall, the platform is relatively decent.

At least it provides a reasonable level of user protection.

Explicit content is restricted or banned.

Given this, it is difficult to imagine how uncontrolled some overseas platforms might be.

V. Public Opinion Bias and Data Manipulation

This experiment reminded me of a quote:

Today, a squirrel dying in front of your house may concern you more than someone dying in Africa.

Based on this experiment, it became very clear how easily public opinion bias and data bias can be created.

5.1 Why Doesn’t the U.S. Want Foreign Control of Local Media?

When I post negative content about a political figure from a certain party on Facebook—complete with solid evidence—the post is not recommended in public feeds.

It can only be seen by visiting my profile.

Meanwhile, positive portrayals of that same figure are continuously recommended.

As a member of the public, would you ever see my message?

This is what public opinion manipulation looks like.

In modern society, I do not need to arrest you.

I only need to reduce your social media exposure to zero.

You speak—but it is as if you never did.

5.2 Data Bias and Information Cocoons

Regarding recommendation algorithms, consider this:

If you only know that someone has bought Chinese sausage, what should you recommend next?

Chewing gum? Coriander?

No.

The optimal choice is to recommend Chinese sausage again.

Why? Because recommending something else carries a high risk of failure, while sausage has already succeeded once.

This is the lowest-risk strategy for operators and those obsessed with “golden cases.”

But is this a good recommendation for the customer?

In reality, it deprives users of potential possibilities—they lose exposure to better or alternative choices.

Eventually, users may not even realize that other options exist.

They continue buying sausage, reinforcing the cycle and further narrowing their exposure.

VI. Final Thoughts

At this point, I feel somewhat fatigued by it all.

I may stop soon.

Hopefully, by the end, there will be deeper insights—or at least a more refined understanding of what I’ve described above.

Goodbye.

The text translation was completed by a large language model.