Documentary in the Age of AI

Katerina Cizek and shirin anlen

The documentary project rides on the relationship between seeing and believing.

For over a century, the camera lens has mediated that relationship; it produced an indexicality, a traceable lineage from the image back to a reality, to a fact and to a deeper truth. That lineage was never perfect, but it provided a foundation upon which documentary, journalism, and human rights documentation could operate. Over the decades, visual cultural practices evolved to make sense of what we see by using convenient and established shorthands: journalistic standards and verification, fact-checking, sourcing, archiving, forensic methods, scholarship, as well as scientific, artistic and community-based forms of authentication and corroboration.

In the last decade, synthetic media has entered the proverbial picture. Generative artificial intelligence models can now produce photorealistic images, audio and video without a human using a camera to record the material world. Multimodal models generate moving images from text-prompt. Voice clones can replicate individuals with uncanny accuracy. Entire characters, environments, and historical scenes can be synthesized. That direct connection between image and reality, however threadbare at times, is breaking. Add to the mix social media infrastructures that by design de-contextualize these (and any) images and toss them into unregulated real-time circulation. New interpretations are ascribed at the whim of whoever’s hands they pass through. A toxic storm. A new epistemic condition emerges: How do people know what they are seeing to decide what to believe?

What happens to the documentary project now?

25 years ago, one of the authors of this foreword made a documentary film called Seeing is Believing: Handicams, Human Rights and the News. It traced the then emergent use of handicams to document human rights abuses around the world, and it centred the work of a human rights media organization called Witness, which is where the other author of this foreword now works. In the film, we examined how cheap, ubiquitous and easy-to-use handicams in the hands of citizen journalists and activists could produce media as potent political and legal interventions. They could challenge official narratives and expose abuses. But even then, the thread between seeing and believing was tenuous. In the film, we examined how handicam video images depicting the 1991 brutal police beating of Rodney King, were manipulated in the first trial of the police officers involved. The defence attorneys slowed the video down, and presented it in court in frame by frame stills to successfully support their argument that correct police procedure had been followed. After the not guilty verdict was announced, Los Angeles erupted in riots. It took a second court trial to restitch the indexical thread of the footage to a reality in which the officers could be found guilty.

In the quarter century that has passed since we made our film, camera lenses have moved far beyond the now antiquated handicam and have proliferated everywhere. Camera lenses now exist in smartphones, dashcams, surveillance cameras, police bodycams, drones, webcams and countless embedded sensors. Billions of images and videos are produced daily, creating an unprecedented visual archive of human life. This phenomenon has radically changed the psychological and political perceptions of indexicality between image and reality, and in turn, the role of documentary.

Generative AI accelerates this shift further with AI slop. Synthetic images, videos, and voices are now produced at a massive scale, made entirely without camera lenses (or prey upon those images). It is a new era: Gen AI is killing the video star. Seeing is not believing.

This central consideration for documentary and human rights we have been exploring since 2018 through our ongoing research collaboration between Witness and MIT Open Documentary Lab’s Co-Creation Studio.

Beyond the issue of bypassing the camera lens, what happens when documentary editing is automated with AI tools, integrated directly into the media production pipeline? When characters can be constructed synthetically as lifelike avatars simulating believable human figures, faces and identities? When these avatars can exhibit realistic conversational capability generated by machines? When images can be enhanced, extended and 2D flat images can be transformed into 3D models using apps, and models that plug in to post-production systems? When images and audio can be “cleaned-up and enhanced”?

AI enhancement tools are now widely embedded across the media ecosystem: in dedicated software such as Topaz Video AI, in consumer apps, in cloud services, and increasingly, inside mainstream editing platforms like Adobe Photoshop and Premiere Pro. Often marketed as “up-rezzing,” “super-resolution,” or “AI restoration,” these tools promise to upscale resolution and “fill in” missing detail from poor source material. These AI tools are often categorized as assistants. The implication is that assistance is neutral: organizational, technical, supportive. But every pixel an AI system generates is not recovered; it is predicted. It is a reconstruction based on probability, and questionable association through algorithmic pattern recognition.

Documentary has always grappled with reconstruction. The difference is that conventional reconstruction techniques, for example, reenactments (such as The Act of Killing, 2012) or animation (such as Waltz With Bashir, 2008), foregrounded authorship; for the most part, they were clearly presented as the filmmaker’s’ version of events (even when revealed at the end of the work, such as in Stories We Tell, 2012).

In AI-generated reconstruction, the made-up material is obscure, random and comes out of nowhere.

Throughout history, in its best form, documentary reconstruction involved deliberate, documented choices. What to clean. What to crop. What to annotate. Those decisions were part of the film’s argument and ethics. They could be debated, but the viewer understood that what they were seeing was a creative reconstruction. AI automation blurs the line between recovery and generation, between damaged truth and plausible fiction. When the reconstruction process becomes invisible, so does accountability.

Not all pixels are equal. Enhancing background foliage may be inconsequential. Reconstructing a face is far more consequential. So is altering a gesture, or synthesizing speech which can fundamentally change the meaning of a video. We must ask: When does reconstruction become fabrication? When does improvement turn into reinterpretation?

At the same time, AI systems open genuinely new narrative capabilities: personalization, imaginative leaps, speculative and localized futures, and forms historically constrained by budgets, access, safety, format, or technical infrastructure. But for these opportunities to matter, they must be community-grounded, ethically governed, creatively effective, and operationally real, especially for documentary makers, human rights defenders, and frontline storytellers working under pressure.

There are clear and shining examples of how these tools have been creatively and ethically harnessed to bring audiences closer to truths that might otherwise remain unseen. Welcome to Chechnya (2020). Another Body (2023). Eno (2024). But for now, they remain edge cases, and the labour required to ensure no harm is done, is herculean.

Most urgently, the entanglement of AI and the documentary project goes far deeper than the crisis of trust, the ethics of an ever-expanding suite of tools, beyond questions about data training, data sovereignty and how data is collected and stolen. The documentary genre, in all its myriad manifestations, has the opportunity and the responsibility to confront the material conditions of the AI Industrial complex in the worlds it seeks to document, intervene in and consider. Documentarians need to engage with this frame as the very subject of our inquiry, not just behind the scenes or as background, but as the main action on centre stage: the military agendas, the mass surveillance, the environmental impacts, the insidious push for more compute (as rationale for more electricity, energy, pipelines), the exploitation of data workers along with human intellectual and creative labour. To confront the colonial empire of AI, as Karen Hao has starkly exposed.

The documentary project must pull out to expand the frame. The story of AI is not only about algorithms. It is about power, labor, environment, and governance. The mythology of AI inevitability, promoted by a concentrated group of corporate giants, demands critical scrutiny. At this historical juncture, the documentary project faces profound crossroads, At its worst, documentary could soon take its last gasp as another fallen video star. But at its best, the documentary project can shapeshift into new forms, new movements and new missions to lock arms with local and global coalitions. Through collective action, documentary stands a chance to uncover the myths and lies being told and sold about Artificial Intelligence. It can also participate and dive in developing new practices of verification, transparency, and alternative AI systems that already exist and build new systems of visual truth. This is the long game. Documentary has a critical role to rethread the needle of meaning-making. It involves rebuilding trust, rethinking authorship, developing new ethical frameworks, and resisting the systems that shape our visual world. Once again, documentary must rethread the connection between seeing and believing.

This volume is an invitation to build the power and possibility of the documentary in the era of AI.

Pietari Kääpä and Dafydd Sills-Jones

In September 2024, the Archival Producers Alliance (APA), consisting of over 300 documentary producers and researchers, published a set of guidelines to manage the use of Generative AI in documentary film. The guidelines were a response to long-term concern over the impacts of deep fakes and algorithmic manipulation tools that had generated concern over the possibility of authenticity and truth in media representation. In the field of documentary, controversies have ranged from the replication of the voice of Anthony Bourdain for the film Roadrunner (2021) to more recent completely AI-generated examples like About a Hero (2024) which all use various degrees of human-like performance. Simultaneously, AI tools had by 2024 become increasingly standardized in documentary production, including in the generation of basic content elements like CGI and backgrounds and had been, for example, used to hide the identities of informants or enhance archival material. Thus, for APA (Horton 2024), the challenge would not so much be about marginalizing or outlawing AI tools, but to reconsider the documentary form and its claim to truth and reality and, effectively, ‘to reaffirm the journalistic values that the documentary community has long held’.