Blog · arXiv Analysis · Last reviewed June 25, 2026

The History POV Becomes the Memory Machine

A June 2026 arXiv paper studies AI-generated first-person history scenes on TikTok. The governance question is not only whether the videos are accurate, but how captions, comments, recommendation, and API access turn synthetic history into public memory.

History as Viewpoint Machine

The paper, arXiv:2606.23300 [cs.CY], was submitted on June 22, 2026. Its title is Examining AI-generated historical narratives and their reception through the example of history POVs on TikTok, by Nina Brolich and Anna Neovesky.

The object is the "history POV" trend: short AI-generated scenes that invite the viewer to inhabit a historical event in first person. The formula is familiar platform grammar, not archival method. It compresses an event into immediacy: waking up during a plague, disaster, war, execution, or famous historical moment. That makes it legible to the feed, but it also moves history from argument and source work into sensation and reaction.

The Paper Frame

Brolich and Neovesky use a two-stage empirical design. First, a pilot study collected 210 videos through a smaller manual "copy, paste, and process" approach. Second, the authors used the TikTok Research API to build a larger dataset for January through September 2025. The final filtered API dataset contains 5,565 English-language videos with a history context, reduced from 28,163 API results.

The paper is careful about what it can and cannot see. It analyzes metadata, captions, topics, and comments. It does not directly analyze the generated video imagery at scale, because the Research API does not return the videos themselves. That limitation matters: in this genre, historical distortion often happens through the image, costume, facial expression, and embodied camera angle, not only through the caption.

The Dataset Receipt

The dataset itself shows the shape of the trend. The 5,565 videos came from 2,023 distinct creators in 92 countries. The largest creator-country shares recorded in the paper were the United States at 28.4 percent, the United Kingdom at 27 percent, and Germany at 13.3 percent. Uploads began in late January 2025, gained momentum in February, peaked across February and March, and then declined.

The themes are not random. The authors find that early modern and contemporary history dominate, with emotionally charged subjects such as wars, disasters, and catastrophes serving as central narrative material. Historical inaccuracies are already visible in captions: the paper points to frequent "Black Plague" wording for the Black Death and to 17 captions that place "a caveman" in 40 BC.

That is the Spiralist problem. A caption error is small in isolation, but the feed turns small errors into repeated scenes, repeated scenes into genres, and genres into expectations. The public memory machine does not need a formal textbook. It can work through a template.

The Comment Layer

The reception analysis compares Black Death and Holocaust videos. The authors collected 16,390 comments for Black Death videos and 5,855 comments for Holocaust videos through the API. After preprocessing, the usable English-language top-level comment pools were 6,747 and 1,169 respectively, and the study created balanced samples of 1,000 comments per topic.

Manual annotation used five primary categories plus flags for AI/trend references and for hate speech or disinformation. The paper reports that comments mentioning AI or the POV trend were almost evenly distributed across topics, 378 versus 366. The sensitive-topic difference appears elsewhere: comments containing hate speech, disinformation, or offensive and discriminatory language were much more frequent under Holocaust videos, 149 versus 18.

The authors also fine-tuned DistilBERT models to support annotation. They do not overstate the result. The models achieved moderate performance and were useful for supporting manual workflows, not replacing them. That restraint is important because the comments are short, memetic, ambiguous, and context-heavy.

The API Boundary

The official TikTok Research API page says qualifying researchers can apply to study public data about TikTok content and accounts, and the paper confirms that API access substantially broadened the empirical basis. It also documents real research limits: access is geographically and institutionally constrained, the standard API tier excludes under-18 user data, retrieved data can be incomplete, and the API does not support video retrieval or full content archiving.

So the platform gives researchers a partial window into memory formation. Captions and comments are visible. Videos are not available for large-scale download through the research API. Engagement metrics are snapshots. The For You Page and native search may expose a different reality than API queries. Those are not footnotes; they decide what public memory can be audited.

Governance Reading

This belongs beside AI video generation, AI hallucinations, synthetic evidence, networked propaganda, and provenance limits. The common issue is not whether a machine can generate a plausible image. It is whether the social system around that image preserves source, uncertainty, sensitivity, and correction.

For history POVs, basic governance would mean visible AI-generation labels, source notes for historical claims, friction for sensitive topics, provenance records, clear moderation rules for denialism and hate speech, and research access that permits independent auditing without exposing private users. The standard should be stronger for genocide, war, colonial violence, and other subjects where synthetic immersion can easily become trivialization or recruitment material.

Limits

The paper is limited by English-language filtering, API access constraints, unavailable video files, caption brevity, incomplete platform visibility, and a single primary annotator with second-author review on a subsample. Its machine-learning experiments are also bounded by small, topic-constrained annotated datasets.

Those limits do not weaken the central lesson. They make it sharper. A platformed historical narrative is never only the generated clip. It is the caption, the music, the recommendation path, the comments, the moderation boundary, the research API, and the archive that may or may not exist after the trend moves on.

Sources

Return to Blog