AI Reflects the Information We Control. That Should Worry Us.

A look at how government edits, platform restrictions, and school-level censorship quietly influence what AI can learn and what it can never say.

Gemini AI Created Image

This essay continues the conversation that began with my piece on .gov sites and the collapse of automatic authority. Many of you commented on social media, saying the quiet part out loud. You felt uneasy long before the federal shutdown banners and timeline edits confirmed what you already sensed. Institutions once considered neutral are no longer behaving that way.

What comes next is the part we have not talked about enough. It affects librarians, educators, researchers, and any community that depends on accurate information. When official platforms are edited, reframed, restricted, or distorted, those changes do not stay isolated. They filter into the digital world, and from there into the systems we are increasingly told to trust.

Artificial intelligence learns from the records it is given. When the record is incomplete or shaped by political motives, the model inherits those distortions. This is one of the most significant information literacy challenges of our time.

AI Does Not Learn From Truth

It learns from whatever has been published, scraped, archived, copied, reposted, and left visible long enough to be absorbed. It cannot separate credible information from information that survived because someone in power allowed it to exist.

AI models learn from:

• Public websites

• Government statements

• News coverage

• Social media

• Digitized books

• Online archives

• Educational materials

• Searchable documents

• And everything that was never deleted

If a government edits its own timeline, those edits become part of the dataset.

If a state removes certain books from public access, those books disappear from local digital collections.

If political actors flood the internet with misleading narratives, the model absorbs the volume and the tone.

If a country restricts foreign news, the model only sees what the government permits.

AI does not know what is missing.
It cannot recognize what was censored.
It cannot retrieve material that never made it into the corpus.
It simply reflects the environment in which it was trained.

The Hidden Chain of Information Control

Most discussions about censorship focus on a single layer: a banned book, a blocked website, or a political message posted on an official page. We rarely talk about the chain of influence that follows.

Here is what the chain looks like in practice:

• A government edits or shapes its own information.

• Platforms amplify or suppress certain narratives.

• Schools restrict what students can read.

• Groups flood the internet with misleading language.

• Search engines promote what is most visible.

• AI models learn from what remains.

By the time an AI model generates an answer, the original truth has already passed through multiple filters. Each filter removes something. Each filter adds something else. What we are left with feels authoritative, but it is often the end product of a shaped record.

This is not science fiction. It is happening quietly in real time.

Examples We Can Already See

When Russia suppressed independent media and restricted access to outside reporting after the invasion of Ukraine, Russian-language datasets were left with state messaging as the primary record.
• Reporters Without Borders: Russia’s media environment
https://rsf.org/en/country/russia
• Amnesty International: Kremlin crackdown on journalism and anti-war speech
https://www.amnesty.org/en/latest/news/2022/03/russia-kremlins-ruthless-crackdown-stifles-independent-journalism-and-anti-war-movement/
• Human Rights Watch: State censorship and increasing isolation
https://www.hrw.org/report/2025/07/30/disrupted-throttled-and-blocked/state-censorship-control-and-increasing-isolation

When China censors books, historical topics, and political discussion online, Chinese-language AI models learn a version of the record that omits entire chapters.
• Asia Society: A history of banned books in China
https://asiasociety.org/new-york/very-brief-history-banned-books-china
• PEN America: State restrictions on children’s books
https://pen.org/press-release/china-new-order-restrict-access-childrens-books-troubling-step-toward-ideological-conformity/
• ICSIN: Internet censorship and the effort to erase political content
https://icsin.org/blogs/2023/10/10/internet-censorship-in-china-the-struggle-to-swat-flies-away-2/

When school districts in the United States remove LGBTQ or race-related books, local digital collections become less diverse. Any school-based model trained on district materials will inherit those gaps.
• Teachers College, Columbia University: Overview of U.S. book banning
https://www.tc.columbia.edu/articles/2023/september/what-you-need-to-know-about-the-book-bans-sweeping-the-us/
• The Guardian: Surge in book bans documented across the U.S.
https://www.theguardian.com/us-news/2025/feb/07/book-bans-pen-america-censorship
• Center for American Progress: Curriculum restrictions and censorship in schools
https://www.americanprogress.org/article/book-banning-curriculum-restrictions-and-the-politicization-of-u-s-schools/

When political actors create large-scale misinformation campaigns, the language and volume of those posts become part of the environment AI learns from.
• Stanford Internet Observatory: Misinformation flows during the invasion of Ukraine
https://arxiv.org/abs/2212.00419
• ArXiv: Multilingual information flows and coordinated messaging across Russian media
https://arxiv.org/abs/2301.10856

These systems reflect the world they were trained in, not the world as it actually is.

Why This Matters for AI in Schools

Schools are adopting AI tools for tutoring, writing support, research help, and intervention planning. Every one of these tools depends on the integrity of the information that trained it. If the dataset reflects censorship, political framing, or missing voices, students inherit those distortions without realizing anything is absent. This affects how they learn, what they believe, and how they evaluate the world.

Why This Matters to Librarianship

Librarians have always taught students how to evaluate sources. What we did not anticipate was that the sources themselves would shift so quickly and so quietly. We assumed that official records, public institutions, and government websites would remain stable points in an unstable information landscape.

That assumption is no longer safe.

Students now live in a world where:

• Government information can be edited without public notice.

• Political messaging can appear inside official documents

• Whole communities can be erased from the record through book bans.

• Misinformation can outweigh fact by sheer repetition.

• AI systems can repeat these distortions without knowing they are distortions at all.

Our role has changed. We are no longer guiding students to trusted sources.
We are teaching them to question the entire pathway by which information reaches them. This essay is the second piece in a larger examination of information power and the systems that shape what students are allowed to know.

Librarians have always stepped in when the information landscape shifts. Our responsibility now is to teach how information changes long before it reaches the search box or the AI prompt

---

If you want to go further with this work, the portion below gives you the practical tools you can use tomorrow with students in grades eight through twelve. It includes a full, ready-to-use lesson on how information changes as it moves through government platforms, media coverage, and AI systems. If you want to teach evaluated trust with clarity and confidence, everything you need is below.