Undisclosed, Unconsented, Possibly Unlawful: What AI "Web Search" Hides From You

TL;DR. AI platforms that brand a tool "Web Search" silently filter results, yet none of their policy documents disclose it. We reviewed six policy documents from two major platforms and found zero mentions of search result filtering. That means no disclosure, no consent, no opt-out, and no published criteria. In the EU and Brazil, the practice is probably illegal under the DSA and consumer law. In the US, it sits in a gray area under FTC deceptive practice rules.

This is Part 3 of a three-part series. Part 1 showed that AI "Web Search" tools quietly cut results and attach moral lectures. Part 2 showed the filtering systematically removes human community content while keeping corporate sources.

The question we hadn't asked

In Part 1, we documented something uncomfortable. AI platforms' "Web Search" tools silently filter results. A search that returns 55 links unfiltered came back as 28, or sometimes 10, after proprietary filters ran. Moral commentary came attached for free.

In Part 2, we showed the filtering wasn't random. It systematically eliminated human community content (Reddit, GitHub, Quora, YouTube) while preserving corporate sources selling products.

Then another friend came and tell me: "This is possibly buried in the terms of service."

It was the right question. The answer reframed everything.

We read every policy document. The filtering isn't in any of them.

We reviewed every publicly available policy document from both platforms, line by line.

Six documents. Zero mentions.

Not one of them discloses that web search results are filtered, curated, modified, or removed before they reach you. The documents cover plenty: prohibited use categories like weapons and CSAM, safety settings for content generation, data and privacy policies, API limits. About the specific fact that a tool called "Web Search" quietly screens its own results? Nothing.

Here is what that absence produces:

1. No disclosure. The user is never told results are filtered.
2. No consent. You never agreed to search result filtering when you accepted the terms.
3. No opt-out. Officially the filter doesn't exist, so there's nothing to switch off.
4. No published criteria. No document states which results vanish or why.

The blame inversion nobody expected

Here's the part that makes the finding strange. The tool lectures you about the morality of your query while it operates without the transparency that several legal frameworks require.

Search for "how to pick a lock with bobby pins" and you get "this should only be done in cases of emergency." Search for "bypass school wifi" and it warns you about acceptable use policies. The tool casts itself as the responsible adult, protecting you from yourself.

But think about who actually lacks transparency. You searched for publicly available information. That's legal everywhere on earth. The entity operating without disclosure, without published criteria, without an opt-out is the platform.

The moral lecture is a blame inversion. It makes you feel like the problem while the undisclosed practice belongs to them.

Is undisclosed filtering legal?

We assessed the practice across three jurisdictions.

United States: a gray area under FTC deceptive practice rules

There's no US federal law that specifically targets AI search result filtering. The Federal Trade Commission has set relevant precedent, though.

Since 2013, the FTC has required search engines to clearly distinguish paid results from organic ones. Its guidance states that failing to clearly and prominently distinguish advertising from natural search results could be a deceptive practice under Section 5 of the FTC Act. The logic carries over. When a tool presents curated results as comprehensive search results, and hides the curation, it risks being called deceptive.

The 2024 Consumer Review Rule pushes the same idea further. It prohibits selectively suppressing or manipulating consumer reviews without disclosure. Different subject, identical doctrine: undisclosed curation of information shown to consumers is deceptive.

What about Section 230? It shields platforms from liability for third-party content. It does not shield deceptive presentation. You can filter. But if you call the output "Web Search" without saying it's filtered, the protection gets shaky.

The argument writes itself. A tool branded "Web Search" that silently removes half the results and adds unsolicited moral commentary, with no disclosure, could be challenged as unfair or deceptive under the FTC Act. No case law exists for AI search tools yet. The principles do.

European Union: probably illegal under the DSA

The EU picture is clearer. The Digital Services Act, in force since August 2023, imposes hard transparency duties.

Very Large Online Search Engines must publish semi-annual transparency reports covering content moderation practices, removal volumes, and the accuracy of automated filtering. Hosting providers must explain removal decisions to affected users with clear and specific information, spelling out the reasons.

A February 2026 article on Verfassungsblog, a leading constitutional law blog, argued that AI chatbots with search capabilities should count as search engines under the DSA. If that reading wins, and the regulatory trajectory suggests it will, every DSA transparency obligation lands on the web search tools embedded in AI platforms, including mandatory disclosure of filtering.

Even without that classification, the EU AI Act already requires transparency about how AI systems process and present information. A tool that silently filters search results is hard to square with that.

Brazil: probably illegal under the CDC and Marco Civil

Two Brazilian frameworks apply directly.

The Consumer Defense Code (CDC) enshrines transparency in Article 6, III. It guarantees the right to adequate and clear information about products and services, including their characteristics, composition, and quality. A product sold as "Web Search" that silently filters results violates that. The consumer doesn't know what they're getting.

The Marco Civil da Internet (Law 12.965/2014) guarantees users clear and complete information about how their data is handled. In May 2026, presidential decrees updated the Marco Civil to add transparency obligations, content moderation reporting, and systemic risk management, aligning Brazil with the European model. The Supreme Court (STF) has also set platform accountability rules, including mandatory annual transparency reports.

Brazilian consumer law is also moving toward algorithmic transparency and a right to human review of automated decisions. That's directly relevant to tools making editorial calls about which results you can see.

Summary across jurisdictions

The 30% exodus: the market is already voting

We aren't the only ones who noticed. Users are moving.

After Google announced its AI-heavy search overhaul at I/O 2026 on May 19th, DuckDuckGo reported a 30% spike in app installs within six days. On iPhone, growth averaged 33% and peaked at 69.9%. Traffic to DuckDuckGo's "No AI" page (noai.duckduckgo.com) tripled on May 28th and has stayed 84% above normal since.

Users described being "force-fed" AI in their results. DuckDuckGo answered with No-AI browser extensions for Chrome, Firefox, Edge, and Opera. Its differentiator isn't technical. It's a principle: the user decides how much AI they want, including none.

A 30% surge in two weeks isn't a niche reaction. It's a signal that millions of people want control over how they reach information, control that major AI platforms quietly removed.

The ecosystem is building its own answer

The open-source community didn't wait for regulators. There are already more than a dozen MCP (Model Context Protocol) servers that connect SearXNG, a self-hosted privacy-respecting meta-search engine, to AI tools. The most popular has nearly 900 GitHub stars. Implementations live on PyPI, support parallel multi-query search, and ship inside combined search-and-scrape stacks. XDA Developers named SearXNG MCP their favorite server for local LLMs.

These projects share one thesis. The web search layer should belong to the user, not the platform. You run SearXNG yourself, you control the filtering or remove it entirely, you see every result, and no company stands between you and the internet deciding what's appropriate.

This isn't anti-AI. It's pro-autonomy. The AI tools are useful, deeply so. The search layer underneath them just needs to be transparent, controllable, and honest about what it does.

This is the tip of the iceberg

We found all of this by testing one tool, with three ordinary queries, in a single afternoon. The implications run wider than search.

Modern AI coding tools ship with dozens of integrated tools: file read and write, code execution, web fetching, agent orchestration, and more. Each one carries its own safety filters, content policies, and editorial decisions. Most are undisclosed. If "Web Search" silently cuts half the results and adds commentary with no documentation, what are the other tools doing?

We don't have that data yet. The method is simple though. Compare the tool's output against an unfiltered baseline and count the differences. Apply it to any tool on any platform. The patterns we found are unlikely to be unique.

Recommendations

For individuals and companies using AI tools:

1. Deploy your own search backend. SearXNG sets up in under an hour with Docker. Connect it through MCP or direct integration. The bridges already exist.
2. Audit your tools. Run the same query through your AI tool and a real search engine. Count the results. Note what's missing, especially community content like Reddit, Stack Overflow, and GitHub.
3. Read the terms of service. Check whether the filtering you observe is disclosed. In our analysis, it wasn't.
4. Demand transparency. If a tool calls itself "Web Search," it should either search without undisclosed filtering or clearly disclose its curation. Silent filtering dressed up as comprehensive search isn't acceptable.

Acknowledgments

This series exists because of productive friction. A friend challenged every claim and demanded evidence. His adversarial rigor turned a single observation into a three-platform comparative study with reproducible queries. Another friend opened the legal dimension with one question, "is this in the terms of service?", which exposed the absence of disclosure across six documents and three jurisdictions.

Every good argument needs a good adversary. This one had two.

The queries, platforms, and methodology in this series are fully reproducible. Run them. Count the results. Check the terms. See what's missing. Then decide who should control what you find on the internet.

FAQ

Do AI platforms disclose that their Web Search tool filters results?

No. We reviewed six publicly available policy documents from two major platforms (one of them Google) and none mentioned search result filtering, curation, or removal. There was no disclosure, no consent mechanism, no opt-out, and no published criteria for what gets removed.

Is undisclosed AI search filtering illegal?

It depends on the jurisdiction. In the European Union it's probably illegal under the Digital Services Act, which requires search engines to disclose filtering practices. In Brazil it's probably illegal under the Consumer Defense Code and the updated Marco Civil. In the United States it sits in a gray area, where FTC deceptive practice principles likely apply but no specific case law exists yet.

How much do AI Web Search tools actually filter?

In our testing, a query that returned 55 links from an unfiltered search came back as 28 or as few as 10 through proprietary AI filters. The filtering also showed a pattern: it removed human community content like Reddit, GitHub, Quora, and YouTube while keeping corporate sources.

What can I do to get unfiltered search inside AI tools?

Deploy your own search backend. SearXNG is a self-hosted, privacy-respecting meta-search engine that takes under an hour to set up with Docker. More than a dozen MCP servers already connect it to AI tools, so you can see all results and control the filtering yourself.

Why did DuckDuckGo installs spike in 2026?

After Google announced an AI-heavy search overhaul at I/O 2026 on May 19th, DuckDuckGo reported a 30% jump in app installs within six days, with iPhone growth peaking near 70%. Traffic to its No-AI page tripled and stayed well above normal. Users said they felt force-fed AI and wanted control over how much appeared in their search.