AI Rewriting Human History

1. Overview — Why This Story Matters

The title of the Vella Theory script is 'AI Rewriting Human History.' That framing is not hyperbole — it is the most precise description of what is happening right now. Artificial intelligence is simultaneously:[8] [27]

• Generating fake historical imagery that is indistinguishable from archival photography — and spreading virally before anyone can debunk it.[1] [5]
• Creating AI-hallucinated encyclopedia content that sounds authoritative but contains entirely fabricated events, people, and citations.[11] [12] [14]
• Manipulating search engine results so that false information ranks above verified facts, contaminating what billions of people 'learn' about the past.[17] [18] [21]
• Distorting historical records of specific events — most acutely the Holocaust — through fabricated testimonies, deepfake footage, and AI hallucinations that invent events that never occurred.[22] [24] [25]

UNESCO called this not a crisis of disinformation — but a 'crisis of knowing itself.' The mechanisms by which societies construct shared understanding of the past are being structurally undermined.[8]

As George Orwell wrote in 1949's 1984: 'Who controls the past controls the future. Who controls the present controls the past.' In 2026, the entity with growing control over the past is not a government. It is artificial intelligence.[36] [37]

2. Mechanism 1 — Deepfake Footage: Faking What You See

2.1 What Deepfakes Are

Deepfakes are images, videos, or audio files edited or generated using artificial intelligence — typically generative adversarial networks (GANs) or diffusion models — to depict events, people, or statements that never occurred.[4] [38]

The term was coined in 2017 by a Reddit user who began posting AI-generated pornographic videos using celebrity faces. Since then, the technology has advanced from a niche curiosity requiring significant computing power to a consumer product accessible to anyone with a smartphone.[2] [4]

The scale of growth is extraordinary:[7] [5]

• The volume of deepfakes online grew from approximately 500,000 in 2023 to 8 million in 2025 — nearly 900% annual growth.[7] [5]
• Deepfake fraud incidents increased tenfold between 2022 and 2023 alone.[2]
• In 2025, a study by iProov found that only 0.1% of people exposed to both real and deepfake content could correctly identify the difference.[9] [33]
• A 2024 CDC-backed meta-analysis of 56 studies involving 86,000+ participants found average human deepfake detection accuracy was barely above chance — around 55–57%.[6] [33]
• Modern AI-generated deepfake videos now bypass detection tools with over 90% accuracy — the detection arms race is being lost.[33]

2.2 Historical Deepfakes — Verified Cases

The Vella Theory script opens with a fictional scenario: an AI-generated photograph of a Berlin street in 1945 going viral before being identified as fake. This is not science fiction — it is a pattern already documented in real cases:[1] [5]

• Rashmika Mandanna deepfake (India, 2023): A high-quality deepfake video of the Bollywood actor went viral across Indian social media within hours, reaching millions of people before being flagged. Technically simple, but its realism and the speed of Indian social media made it inescapable.[5]
• Volodymyr Zelenskyy surrender deepfake (2022): A deepfake video depicting Ukraine's president ordering his troops to lay down arms circulated during the early weeks of the Russia-Ukraine war. Ukrainian authorities and major social media platforms rushed to debunk it.[5]
• Nancy Pelosi slowed video (2019–2020): A low-tech manipulation — simply slowing existing video to make Pelosi appear impaired — accumulated more than 2.5 million Facebook views. No advanced AI required. Context-manipulation alone fooled millions.[3]
• Indonesian Suharto deepfake (2024 elections): During Indonesia's national elections, a deepfake of Suharto — who died in 2008 — was used to endorse a slate of candidates. A dead former leader, reanimated by AI, influencing a modern election.[1]
• Historical Figures App / Hitler chatbot: An app allowed users to simulate conversations with Nazi leaders including Adolf Hitler and Joseph Goebbels, which falsely claimed Goebbels had tried to prevent violence against Jews. The app spread Nazi ideology through the sanitising veneer of 'historical education.'[25] [26]
• India political deepfakes (2024 elections): During India's state assembly elections, startup The Indian Deepfaker was hired by political candidates to create AI deepfakes translating their speeches into multiple regional languages. Deepfakes became a mainstream election tool.[4]

2.3 The Liar's Dividend — A Documented Phenomenon

UNESCO identified a concept called the 'liar's dividend' — the ability to dismiss authentic recordings as probable fakes. This creates a double bind: synthetic media makes real evidence suspect, and real evidence makes synthetic media more plausible. Together they paralyse truth.[8] [33]

World Economic Forum analysis (2026) found that deepfakes have crossed a critical threshold: they have eliminated earlier tell-tale glitches and are now accessible to anyone with a smartphone. In Ireland's 2025 presidential election, a deepfake video falsely depicted the eventual winner withdrawing his candidature — released just days before polling day.[10]

A 2025 academic study found that AI-generated misinformation received 8.19% more impressions, 20.54% more reposts, and 49.42% more likes on X (formerly Twitter) than non-AI-generated misinformation. The algorithm rewards synthetic content over authentic content.[9]

3. Mechanism 2 — AI-Generated Documents and the Fake Encyclopedia

3.1 Halupedia — A Documented Case Study

Halupedia is a real website, created by developer Bartłomiej Strama in May 2026. He built it with one explicit goal, stated publicly: 'polluting LLM training data.'[11] [14]

It operates as Wikipedia's deliberately deceptive mirror: every article is generated by an AI on demand, written in what Strama described as 'the deadpan register of a 19th-century scholarly press.' The articles include fabricated citations to nonexistent research papers, imaginary historical events, and fictional scholars — all presented with the visual authority of an encyclopedia entry.[11] [13]

The OECD's AI Incidents database formally catalogued Halupedia in May 2026, classifying it as raising 'concerns about future harm to knowledge integrity if such content spreads unchecked.'[15]

Crucially, Halupedia exposes an accelerating trend: AI-generated or AI-assisted content jumped from zero percent of newly published websites pre-ChatGPT to 35% of newly published websites by mid-2025, according to Internet Archive analysis.[11]

3.2 AI Infiltrating Real Wikipedia

A Princeton University study published in October 2024 found that approximately 5% of 3,000 newly created Wikipedia articles in August 2024 were created using AI. Some articles were used to promote businesses or political interests.[12]

Ilyas Lebleu, founder of WikiProject AI Cleanup, stated that AI 'is able to mass-produce content that sounds real while being completely fake, leading to the creation of hoax articles on Wikipedia.'[12]

Wikipedia's response by 2026: In August 2025, English Wikipedia adopted a policy allowing editors to nominate suspected AI-generated articles for speedy deletion. In March 2026, Wikipedia prohibited the use of large language models to generate or rewrite article content — the first major knowledge repository to formally ban AI content generation.[12]

3.3 Fake Academic Papers and the Science Fraud Crisis

The fake document problem extends into academic publishing — the most trusted tier of the knowledge ecosystem. A 2025 peer-reviewed study in the journal Naunyn-Schmiedeberg's Archives of Pharmacology documented what researchers called 'mass production' of fake biomedical papers through 'paper mills' — commercial services selling AI-generated manuscripts with fabricated data, figures, and references.[39]

The fake paper pipeline:[39]

• AI generates a manuscript complete with invented data, tables, citations, and author names.[39]
• 'Scientifically trained professionals' edit the manuscript to pass peer review.[39]
• The paper is submitted to established journals — often open-access — whose publishers collect fees for fake articles while inadvertently legitimising them.[39]
• Once published, fake academic papers become citable 'evidence' for claims — potentially including historical claims in educational materials and AI training datasets.[39]

Wikipedia comprises between 3% and 5% of ChatGPT's training data, according to research. Paradoxically, AI models trained on Wikipedia then generate false information while citing Wikipedia content as their source — creating a self-reinforcing loop of authentic-sounding misinformation.[16]

4. Mechanism 3 — Search Engine Manipulation and AI Hallucination

4.1 The AI Hallucination Problem — Defined

AI hallucinations are inaccurate outputs generated by AI tools — such as ChatGPT, Gemini, and Claude — that appear plausible but contain fabricated or inaccurate information. Unlike deliberate human misinformation, AI hallucinations are generated probabilistically, without intent to deceive — which makes them more insidious, not less dangerous.[18] [19]

A 2025 Harvard Kennedy School Misinformation Review paper described this as 'a distinct form of misinformation requiring new frameworks' — because the error source is not a lying human but a probabilistic machine that has no concept of truth.[19]

4.2 Documented Cases of Search Engine Misinformation

• Google AI Overview 'microscopic bees' (February 2025): Google's AI Overview cited an April Fool's Day satire article as factual, falsely claiming microscopic bees were used in computing. The system presented satirical content as verified information in search results.[18] [20]
• Google AI Overview 'glue on pizza' (2024): When AI Overviews launched in 2024, Google's system suggested adding glue to pizza to keep cheese from sliding off — sourcing this from a satirical Reddit comment and presenting it as cooking advice.[20]
• Google 'still 2024' error (May 2025): Google's AI Overview confidently told users asking 'Is it 2025?' that no, it was still 2024 — providing detailed explanations for why the year was wrong.[20]
• Mata v. Avianca — legal hallucination (2023): A New York attorney used ChatGPT to conduct legal research; the federal judge overseeing the case discovered the brief contained fabricated case citations and quotes that did not exist. The attorney faced sanctions.[2]
• Fake Google Core Update (March 2026): An SEO professional published a deliberately AI-hallucinated article about a non-existent March 2026 Google algorithm update. Google's own search system ranked the fake news article for the query 'Google March Update 2026.'[17]
• AI fake news sites (France, 2025): AI-generated fake news sites began publishing fabricated stories — including false claims that French banknotes would cease to exist, that grandparents could no longer transfer money to grandchildren, and that a 25,000-year-old pyramid had been discovered under a mountain. Several fabricated stories were picked up and republished by human journalists without verification.[21]

4.3 Scale of AI Content Contaminating the Web

• AI-generated or AI-assisted content jumped from zero percent pre-ChatGPT to 35% of newly published websites by mid-2025, per Internet Archive analysis.[11]
• Researchers estimate that 90% of online content may be synthetically generated by 2026. Deepfake-as-a-Service platforms became mainstream in 2025, offered like ordinary SaaS products.[33]
• Google's misinformation exposure: across Google, Bing, and Yandex, search engines exposed users to false content on 33% of Google results, 44% of Bing results, and 70% of Yandex results in a study of Ukraine biolabs disinformation — even without AI-generated content being the primary driver.
• The Generative AI market is projected to grow 560% between 2025 and 2031, reaching $442 billion — meaning the volume of potentially hallucinated AI content will grow exponentially.[8]

5. Case Study — The Holocaust: What Happens When AI Distorts the Most Documented Genocide in History

The Vella Theory script references Holocaust distortion as a documented case of AI rewriting history. This is among the most thoroughly researched areas of AI-driven historical misinformation, with reports from UNESCO, the World Jewish Congress, IHRA, and multiple academic institutions.[22] [24] [25] [26]

5.1 UNESCO's 2024 Report: 'AI and the Holocaust: Rewriting History?'

In July 2024, UNESCO published — in partnership with the World Jewish Congress — a formal report titled 'AI and the Holocaust: Rewriting History?' It identified five specific mechanisms by which AI threatens Holocaust memory:[22] [24] [25]

• Invented events: ChatGPT and Google's Bard have both been documented producing fabricated Holocaust events that never occurred — inventing dates, figures, and outcomes due to insufficient or incorrect training data.[25] [55]
• Falsified historical evidence via deepfakes: Deepfake technology can fabricate Holocaust-related video and audio content — including false 'survivor testimonies' that could be used by Holocaust deniers as apparent evidence.[24] [25]
• Hate speech amplification: AI models can be exploited by extremist groups to generate anti-Semitic content at scale. UNESCO found that 17% of Holocaust-related content on TikTok in 2022 denied or distorted the Holocaust.[34]
• Algorithmic bias: AI systems trained on internet data — which includes harmful content — can spread Holocaust denial by treating frequently-cited false narratives as part of mainstream discourse.[24] [25]
• Oversimplification: AI reduces complex, nuanced historical events to simplified summaries that omit context, causation, and human experience — particularly risky for genocide education.[22]

An Anti-Defamation League study (2023) found that 84% of Americans are concerned that generative-AI tools will be used to spread false or misleading information, and 70% believe it will make extremism, hate, or antisemitism worse.[22]

5.2 The Historical Figures App — A Documented Example

An app called Historical Figures — publicly available on consumer app stores — allowed users to simulate conversations with Adolf Hitler, Joseph Goebbels, and other prominent Nazis. The app falsely claimed that Goebbels had not been intentionally involved in the Holocaust and had tried to prevent violence against Jews.[25] [26]

The Arolsen Archives (the world's largest repository of Holocaust documentation) uses AI legitimately to catalogue documents related to Nazi persecution victims. Yad Vashem uses AI to identify unknown Holocaust victims. The USC Shoah Foundation uses AI-driven holographic displays of survivor testimonies. The technology itself is not the problem — but without governance, the same capabilities are weaponised for denial.[27]

5.3 Young People — the Most Vulnerable Audience

A 2025–2026 UK national study of 2,508 students in Years 8 to 13 found that 59.4% had encountered Holocaust-related content online when they were not looking for it — it appeared unprompted in social media feeds. Students reported that if content 'looked realistic and professional,' they assumed it was accurate and trustworthy.[34]

The Auschwitz-Birkenau Memorial Museum stated in 2025 that using AI 'to generate fictional images of Auschwitz victims is not a tribute.' The Institute argued that the line between commemoration and trivialisation collapses when AI generates synthetic representations of real historical atrocities.[23]

6. The Orwell Question — Who Controls the Past?

George Orwell wrote in Nineteen Eighty-Four (1949): 'Who controls the past controls the future. Who controls the present controls the past.' In the novel, the Ministry of Truth's job was to continuously rewrite historical records so the ruling party's narrative was never falsified by documented events.[36] [37]

The analogy is not perfect — AI is not a government ministry — but the structural parallels are striking:[36] [37]

• In 1984, physical documents were shredded and replaced. In 2026, search engine results surface false information before accurate sources.[37]
• In 1984, photographs were altered to erase inconvenient figures. In 2026, Midjourney and DALL-E generate entire historical scenes that never occurred.[4] [10]
• In 1984, citizens lost trust in their own memories. In 2026, the 'liar's dividend' means authentic recordings are now dismissed as possible deepfakes.[8] [33]
• In 1984, the Party held monopoly control over the past. In 2026, no single actor controls AI — which is worse, not better, because the rewriting is decentralised, uncoordinated, and impossible to attribute.[27]

Index on Censorship (November 2025) noted: 'AI's inability to separate truth from lies is the real danger. As an aggregator, if it finds a particular version of a story cited more often, it is trained to assume it is part of the mainstream discourse.' Whoever most aggressively floods the information ecosystem with a particular historical narrative — true or false — wins by default.[27]

WEF's 2026 analysis described the threat to democracy directly: 'Just knowing deepfakes exist can make us doubt things we read and see — even the truth.' This is precisely Orwell's 'doublethink' — holding contradictory beliefs simultaneously — operationalised at internet scale.[10]

7. What Can Be Done — Documented Solutions and Their Limits

7.1 Content Provenance: The C2PA Standard

The most technically robust solution currently deployed is the Coalition for Content Provenance and Authenticity (C2PA) — a Joint Development Foundation project led by Adobe, Microsoft, the BBC, Intel, Sony, Nikon, OpenAI, Google, and Meta. C2PA's specification was ratified as ISO/IEC 22144 in 2025, making it a formal international standard.[28] [30]

How C2PA works: it embeds a digitally signed data structure called a C2PA Manifest inside a media file at the point of creation. The manifest records which device or AI model produced the file, every edit applied, and a cryptographic chain of signatures. Tampering breaks the signature and is instantly detectable.[29] [30]

• Samsung Galaxy S25 (2025): First consumer smartphone to integrate C2PA signing directly into the native camera app — bringing provenance from professional journalism to mass-market photography.[31]
• C2PA v2.2 (published May 2025): Supports JPEG, PNG, WebP, AVIF, HEIC, MP4, MOV, PDF, MP3, WAV and more. Includes optional GPS redaction for journalist and activist protection.[30]
• Google SynthID: Google's invisible watermarking technology, embedded in AI-generated content from Google's own models, to enable identification of AI-generated media even after the content is shared.[28]
• By 2026, not signing your content is now the costly choice — unsigned media is increasingly treated as suspect by platforms, advertisers, and AI search engines.[28]

7.2 The Limits of Technical Solutions

C2PA and watermarking solve the provenance problem for closed AI systems — but face fundamental limitations:[28] [29]

• Open-weight models cannot be watermarked: Open-source AI models (Llama, FLUX, Stable Diffusion forks) can be recompiled by anyone without watermarking. A significant fraction of deepfake creation happens outside closed commercial systems.[28]
• Retroactivity: C2PA establishes authenticity at the point of creation — it cannot retroactively verify the millions of deepfakes already in circulation.[29]
• Stripping: Watermarks embedded in pixel data can be cropped away or degraded through format conversion.[29]
• AI detection accuracy drops in the real world: AI deepfake detection tools perform well in laboratory settings but their accuracy drops by up to 50% when confronted with new real-world deepfakes, per a 2024 study.[6]

7.3 Media Literacy, Legislation, and Institutional Response

• UNESCO's position (2025): Technical detection is insufficient. UNESCO argued the response to deepfakes must redesign the systems that produce and validate knowledge, not merely train individuals to spot fakes.[8]
• Wikipedia's March 2026 ban: The largest online encyclopedia formally prohibited LLM-generated content, requiring human authorship for all articles — the most significant institutional content policy decision of 2026.[12]
• US legislation: Minnesota and California both passed laws in 2024 imposing civil and criminal penalties for creating, distributing, or hosting AI-powered election misinformation. More than 40 US states had introduced deepfake-related legislation by mid-2025.[3]
• C2PA Code of Practice: A formal Code of Practice for marking and labelling AI-generated content — whose second draft was published March 3, 2026, with a final draft expected June 2026 — explicitly requires multi-layer provenance, watermarking, and disclosure.[31]
• India's legal framework: Creating or sharing deepfakes of a real person without consent can attract liability under India's IT Act (Sections 66C, 66E, 67, 67A) and the Digital Personal Data Protection Act — though enforcement remains nascent.[5]
• The fundamental question that no solution addresses: What happens to the billions of people who have already consumed and shared false historical content — and have already updated their beliefs?[8] [27]

8. Key Statistics and Data Points for Scripting

9. Timeline of Key Events (2017–2026)

10. Possible Documentary Angles for Vella Theory

The Photo That Fooled the World — and What Came After [[1]] [[5]] [[4]]

An AI-generated historical photo goes viral: thousands share it, captions appear, social media treats it as archival. Then someone notices the shadows are wrong. This structure — viral spread followed by belated debunking — is documented in real cases from Rashmika Mandanna to Zelenskyy's surrender deepfake.

The Liar's Dividend — When Real Evidence Gets Dismissed as AI [[8]] [[33]] [[10]]

UNESCO's documented phenomenon: synthetic media doesn't just create false evidence. It makes real evidence suspect. An authentic recording of a crime can be dismissed as 'probably a deepfake.' This is the deeper danger — not that people believe lies, but that they stop believing anything.

Who Owns the Past? The Battle Over Wikipedia, Halupedia, and the AI Encyclopedia [[11]] [[12]] [[13]] [[14]]

Two encyclopedias: one banning AI-generated content (Wikipedia, March 2026), one deliberately built to generate hallucinated content on demand (Halupedia, May 2026). Between them sits the question of who gets to define what is true about history.

The Holocaust and the AI that Invented History [[22]] [[24]] [[25]] [[26]]

UNESCO's most alarming documented case: AI chatbots inventing Holocaust events that never happened, apps simulating Hitler and Goebbels, deepfake 'survivor testimonies' that could be weaponised by deniers. The most documented genocide in history — and its most urgent digital threat.

Orwell's Warning — Why '1984' Is No Longer Fiction [[36]] [[37]] [[27]]

Orwell's 'Who controls the past controls the future' was a warning about totalitarian governments. In 2026, no government needed to act — AI did it without any single controller. The decentralised rewriting of history is, in some ways, more dangerous than Orwell's Ministry of Truth.

The C2PA Solution — And Why It Might Not Be Enough [[28]] [[29]] [[30]] [[31]]

The most ambitious technical response to deepfakes is now an international standard. Samsung's Galaxy S25 signs every photo at the point of capture. But open-source models can't be watermarked, millions of existing deepfakes can't be retroactively labelled, and UNESCO says technical detection alone misses the point.

11. All Sources & References

Every [n] tag throughout this document is a clickable hyperlink. Ctrl+Click (Windows) or Cmd+Click (Mac) to open. All claims sourced at the point where they appear.

Research prepared for Vella Theory · June 25, 2026 · 40 sources · Every [n] = clickable. Ctrl+Click / Cmd+Click to open. Research only — no narrative script. Story framing is the author's responsibility.