I Opened “Transparency Mode” on Six Platforms at the Same Time — What Algorithmic Recommendation Has Become in 2026

By Lucien Viremont | Updated on January, 2026 | 🕓 12 min read

Key Highlights

- Why do algorithmic systems increasingly predict user behavior instead of simply responding to it?

- Why are hidden commercial incentives embedded inside recommendation systems?

- How has advertising evolved to become less recognizable as advertising?

- Do “transparency modes” and non-personalized feeds actually reduce algorithmic bias?

- Why can non-personalized recommendations sometimes feel more trustworthy?

- What practical steps can ordinary users take to audit algorithmic influence?

This is not an explainer about algorithms. It is a clumsy but honest record of the past eight months I spent, as an ordinary user, tracing recommendation systems across six platforms, three browsers, and regulatory filings spanning two continents.

1. From “You Might Like This” to Predicting Needs You Haven’t Even Thought About Yet

One evening in late 2025, I ran a pointless little experiment. I opened TikTok on two devices at the same time: one logged into my account, the other in a completely fresh incognito session. I searched the exact same phrase on both:

“How to repair old jeans.”

The results stopped me cold.

On the logged-in account, the first three videos were:

a Dutch designer demonstrating visible mending techniques,

an advertisement from a fast-fashion brand promoting its “used clothing recycling program,”

and an AI-generated virtual influencer promoting a sewing machine.

In the incognito session, the first three results were completely different:

an Indian tailor showing traditional repair methods,

an American DIY creator posting a compilation of failed attempts,

and a straightforward tutorial video with no shopping links at all.

This was not simply the difference between “personalized” and “non-personalized.”

It felt like two entirely different realities.

By 2026, recommendation systems had crossed some invisible threshold. TikTok’s algorithm no longer relied primarily on the labels creators assigned to their own videos. Instead, computer vision and speech recognition systems now directly “watch” the content themselves. The platform interprets what appears in the video instead of trusting creators to describe it accurately.

YouTube’s recommendation architecture had also evolved beyond a single unified algorithm. Home, Suggested, Search, and Shorts now operate as four partially independent ranking “rooms.” The exact same video might dominate in one room while disappearing completely in another.

Amazon was even more unsettling.

Its recommendation engine processes more than 5.6 billion user behavior events every day, pulling information from a 24-petabyte data lake and coordinating over forty specialized microservices to generate personalized recommendations in an average of 142 milliseconds. One service handles behavioral analysis. Another handles collaborative filtering. Another processes real-time events.

Together, they function like an extraordinarily precise machine.

The problem is that the machine’s objective has never been merely “helping users.”

Later, I came across research showing that Amazon’s personalized recommendation systems account for approximately 35% of retail revenue and contribute to a 29% increase in average order value. At the same time, the company reportedly runs more than 7,000 experiments per year, with core optimization metrics focused on click-through rates, conversion rates, and platform profitability.

When I started noticing that high-margin products systematically appeared ahead of high-value products in search rankings, the meaning of those experiments became clearer.

The algorithm is not answering the question “What is best for you?” It is answering the question “What is most profitable for the platform while still making you feel like this was your own decision?”

2. When Advertising Learns How to Disappear

In January 2026, I overheard a conversation in a café.

Two young people were discussing a skincare product. One of them said:

“I saw so many people using it on TikTok. It must actually be good.”

What she did not realize was that a significant portion of those “people using it” videos were undisclosed sponsored promotions.

This is the most dangerous evolution of advertising:

it no longer looks like advertising, so people stop defending themselves against it as advertising.

At the end of 2025, TikTok updated its policies to require all commercial content to use built-in disclosure tools, otherwise risking removal from the For You feed entirely. The FTC also intensified enforcement around influencer marketing disclosures.

But recommendation-driven commercial content had already evolved into forms that were far harder to identify.

The first layer: “organicized” commercial content

Creators receive free products and publish “honest reviews” without using disclosure tools. Whether platforms actually distribute transparently disclosed content with the same algorithmic reach remains difficult to independently audit.

The second layer: AI-generated endorsements

In 2026, TikTok formally prohibited the use of AI-generated fake personalities to endorse products. But enforcement still relies heavily on automated detection systems — and marketers are rapidly learning how to evade those systems.

The third layer — and the hardest to notice: content farms optimized for AI crawlers

As users increasingly stop searching through Google and instead ask ChatGPT, Perplexity, or Meta AI questions like “What is the best product?”, brands have started reverse-engineering the information ecosystem feeding those AI systems.

They mass-produce structured “Top 10” lists, comparison tables, pseudo-neutral Reddit comments, and Q&A forum posts. To human readers, much of this material feels repetitive and shallow. But to AI crawlers, it is ideal training material.

When AI assistants later “summarize the internet” to recommend products, the supposed “community consensus” may actually be the byproduct of a carefully orchestrated SEO campaign.

During this tracking project, I noticed an unsettling pattern.

Amazon’s “Frequently Bought Together,” TikTok’s “You May Like,” and Instagram’s “Explore” all appeared to operate on some variation of the same underlying equation:

Click probability × conversion likelihood × platform profitability

That formula is never written anywhere users can see it.

Yet it quietly determines what billions of people encounter every day.

3. The Rise of Transparency Tools — and the Darkness They Still Cannot Reach

In August 2025, the European Union’s Digital Services Act (DSA) fully activated its transparency requirements for recommendation systems.

Under Article 27, very large platforms must explain the “main parameters” behind their recommendation systems. Article 38 further requires them to provide at least one recommendation option that does not rely on user profiling.

It sounded like a victory for users.

But when I actually tested these transparency features, the reality felt far more complicated.

TikTok did introduce a non-personalized For You feed option back in August 2023, and added keyword filtering and topic management tools in June 2025. Instagram allows users to adjust recommended Reel topics. Facebook restored chronological feed options in parts of Europe.

Still, three problems stood out to me — problems largely absent from official narratives.

First: “non-personalized” does not mean “unbiased”

When I switched TikTok into non-personalized mode, recommendations were no longer based on my viewing history. But the feed was still shaped by broader commercial partnerships, regional weighting systems, and moderation policies.

It felt like a restaurant telling you:

“We no longer remember your preferences.”

while still offering only the dishes they most want to sell.

Second: the scale of transparency reporting is overwhelming

According to the DSA Transparency Database operated by the European Commission, platforms reported more than nine billion content moderation decisions during the first half of 2025 alone. Approximately 99% of those actions were initiated proactively under the platforms’ own internal policies rather than in response to user reports of illegal content.

In other words, the “natural” feed people experience is already the result of billions of invisible interventions.

Third — and perhaps most ironic — regulatory fines themselves reveal how serious the problem is

In December 2025, the European Commission fined X €120 million over misleading verification mechanisms tied to “blue checkmarks” and failures in advertisement transparency systems.

That same month, TikTok committed to updating its advertising repository within 24 hours and including complete advertisement content records.

These enforcement actions were significant breakthroughs.

But they also exposed something uncomfortable:

before regulators intervened, these platforms had never voluntarily provided meaningful transparency.

I experimented with several third-party auditing approaches — simultaneously running personalized and non-personalized sessions while recording ranking differences for identical searches.

Very quickly, though, I hit technical walls.

Platform APIs remain inaccessible to ordinary researchers. Although Article 40 of the DSA theoretically grants researchers data access rights, application procedures are complicated and platform cooperation remains inconsistent.

An ordinary user — or even an independent journalist — still has almost no realistic way to verify whether an algorithmic system is treating them fairly.

4. A Counterintuitive Discovery: Sometimes Non-Personalized Feeds Feel More Honest

Among all my messy cross-platform experiments, one finding stayed with me more than any other.

I compared personalized and non-personalized recommendations for news and health-related content.

Under personalized systems, platforms consistently appeared to prioritize emotionally activating material:

controversy,

anxiety,

outrage,

polarization.

Under non-personalized feeds, the emotional spectrum felt wider, more varied, and frankly more boring.

That observation echoed a 2025 study from the University of Washington. Researchers found that when human reviewers collaborated with AI hiring systems, people tended to unconsciously mirror the system’s hidden biases unless those biases were extremely obvious.

We are remarkably vulnerable to anything wrapped in the appearance of “data-driven objectivity.”

Another randomized field experiment involving StumbleUpon produced even more direct evidence. Researchers found that although personalized recommendation emails achieved higher open rates among new users, non-personalized emails ultimately generated significantly stronger click-through engagement and platform participation.

The researchers explained this through expectancy-disconfirmation theory:

people approach personalized content with higher expectations and therefore experience disappointment more easily.

With non-personalized content, expectations begin lower, making users more open to exploration.

That finding fundamentally changed how I think about algorithmic recommendation systems.

The problem with personalization is not that it is insufficiently intelligent. The problem is that it is too intelligent. It optimizes for immediate reaction rather than long-term satisfaction.

5. What Ordinary People Can Still Do: Four Practical Actions From My Own Experiments

I am not a technologist. I am not a policymaker.

I am simply someone unwilling to surrender all decision-making autonomy to invisible systems.

Based on these eight months of clumsy experimentation, here are four practical habits I now follow — none of which require technical expertise.

1. Conduct a “dual-window audit” regularly

Once a month, search the same question simultaneously in incognito mode and on your personal account.

It could be:

“What are the best running shoes?”

or

“How should I invest in index funds?”

Compare the first three results.

The difference itself is the “attention tax” the algorithm is charging you.

2. Use non-personalized modes intentionally — but skeptically

Switch TikTok, Instagram, or YouTube into non-personalized or chronological modes occasionally.

Not because you can fully escape algorithms, but because doing so recalibrates your perception of what “normal” looks like.

At the same time, remember:

this is still only another version of filtering that the platform has chosen to allow you to see.

3. Treat AI-generated “summaries” with double skepticism

When you ask ChatGPT or Gemini “What is the best X?”, remember that the answer may partly originate from content farms optimized for AI crawlers.

Cross-check information.

Visit independent forums.

Browse professional communities.

Read obscure blogs maintained by people with no SEO team behind them.

Real quality often hides beyond the first page of search results.

4. Introduce deliberate “algorithmic friction” into high-stakes decisions

For health, finance, or major purchasing decisions, I now force myself to follow one rule:

before making a final decision, I must either discuss it with a real human being or consult at least two non-algorithmically mediated sources — such as physical magazines, library databases, or reports published by professional associations.

This is not anti-technology.

It is an attempt to reclaim physical boundaries around decision-making.

Conclusion: Algorithms Are Not the Enemy — Invisibility Is

The recommendation systems of 2026 are not conspiracies.

They are extraordinarily efficient commercial tools designed to maximize platform profitability while minimizing the user’s cognitive resistance.

Their danger does not come from their power alone.

It comes from the fact that this power is packaged as neutrality, convenience, and care.

The European Union’s DSA, the AI Act, and broader global regulatory efforts are slowly opening cracks in that façade.

But regulation will always lag behind technology.

And technology will almost always serve capital first.

Ordinary users may never fully escape these systems.

But we can refuse to become passive data points inside them.

The moment you recognize that “recommendation” is a commercial behavior rather than a public service, you reclaim the first inch of autonomy.

Algorithms can decide what appears in front of you.

But they cannot decide how you interpret it —

unless you allow them to.

About the Sources and Limitations of This Article

The cross-platform experiments described in this article were based on my own informal observations and do not constitute academically rigorous evidence.

Platform policies, technical descriptions, and regulatory details referenced here were drawn from publicly available disclosures, regulatory filings, and academic literature, with references listed below.

If you notice factual inaccuracies or outdated information, I welcome corrections.

In an era of rapidly evolving algorithms, any description of the “current state” is likely to be temporary.

FAQs

1. Are personalized recommendations always harmful?

No. Personalized systems can improve convenience, discovery, accessibility, and relevance. The concern discussed in this article is not personalization itself, but the lack of transparency around the commercial and behavioral incentives embedded within recommendation systems.

2. Why do algorithms prioritize emotionally charged content?

Many recommendation systems optimize for engagement metrics such as watch time, clicks, comments, and shares. Content that triggers strong emotional reactions often performs better under those metrics, which can unintentionally incentivize polarization, outrage, or anxiety-driven material.

3. Can users fully escape algorithmic influence?

Realistically, no. Most large digital platforms rely on some form of ranking or filtering system. Users can, however, reduce dependence on personalization by using chronological feeds, non-personalized modes, independent sources, and deliberate cross-checking habits.

4. What is the Digital Services Act (DSA)?

The Digital Services Act is a European Union regulatory framework designed to increase platform accountability, transparency, and user protections online. It includes rules requiring large platforms to explain recommendation systems and provide alternatives to profile-based recommendations.

5. Why are AI assistants vulnerable to manipulated information?

AI assistants often summarize publicly available online content. As brands and marketers increasingly produce material optimized for AI crawlers rather than human readers, recommendation quality may be distorted by SEO-driven content ecosystems rather than genuine expertise or consensus.

6. What are “AI content farms”?

AI content farms are large-scale networks of articles, reviews, forum posts, and comparison pages designed primarily to influence search engines and AI systems. These pages often imitate authentic reviews or discussions while being strategically optimized for visibility.

7. Is using incognito mode enough to avoid personalization?

Incognito mode can reduce some account-based personalization signals, but platforms may still use device information, regional data, browser fingerprints, or broader popularity trends to shape recommendations.

References

1. The Filter Bubble — Pariser, E. (2011). The Filter Bubble: What the Internet Is Hiding from You. Penguin Press.

2. The Age of Surveillance Capitalism — Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.

3. Zhao, X., Chen, C., & Wang, Y. (2022). Commercial bias in recommender systems: A survey. ACM Computing Surveys, 55(7), Article 142.

4. Digital Services Act — European Commission. (2025–2026). Digital Services Act Implementation Reports and Enforcement Actions.

5. Fedorov, L. (2026). Explainable Recommendations in Large-Scale Content Feeds. Universal Library of Innovative Research and Studies, 3(1).

6. Wilson, K., et al. (2025). People mirror AI systems' hiring biases. AIES '25.

7. KAIST/StumbleUpon Field Experiment (2025). Personalized vs. non-personalized engagement for at-risk new users.

About the Author

Lucien Viremont, PhD – AI Decision Systems & Cognitive Risk Researcher

Lucien Viremont, PhD is a researcher and writer focusing on AI decision-making systems, cognitive bias in algorithmic environments, and the psychological impact of automation on human judgment. He holds a PhD in Cognitive Science from Stanford University, where his research explored how humans calibrate trust in machine-generated recommendations. He has worked with AI ethics labs and decision intelligence startups in the US and Europe. His writing focuses on how AI systems shape — and sometimes distort — human reasoning, autonomy, and responsibility in complex decision environments.

Editorial Transparency Statement

This article combines personal experimentation, publicly available regulatory documents, academic research, platform policy disclosures, and independent analysis.

The cross-platform comparisons described in the article were informal observational experiments conducted by the author and should not be interpreted as controlled scientific studies. Some interpretations reflect the author’s analysis of broader industry trends and public reporting rather than direct access to proprietary platform systems.

Every effort has been made to ensure factual accuracy at the time of publication. Because recommendation systems, AI policies, advertising standards, and platform regulations evolve rapidly, some technical or regulatory details may change after publication.

The article was written independently and was not sponsored by any technology company, platform, advertising network, or regulatory body.

Disclaimer

The information provided in this article is intended for educational, informational, and editorial purposes only. It does not constitute legal advice, financial advice, investment advice, regulatory guidance, or professional cybersecurity consultation.

References to specific companies, platforms, technologies, or regulatory actions are included for commentary and analysis purposes. Platform policies, enforcement practices, recommendation systems, and regulatory obligations may vary by region and may change over time.

Readers are encouraged to consult official platform documentation, regulatory sources, academic publications, and qualified professionals when making important decisions related to privacy, technology use, finance, health, or digital security.