Fred is a media executive, producer, and co-founder of Machine Cinema, based in Los Angeles. His career spans film and television, emerging technologies, and the creator economy. As Head of Non-Fiction Programming at Anonymous Content, he oversaw projects including Undercurrent: The Disappearance of Kim Wall (HBO), At the Ready (Sundance 2021), and Murders Before the Marathon (Hulu). His independent credits include Oscar short-listed Hidden Letters and Lowndes County and the Road to Black Power (Tribeca 2022). Earlier in his career he served as VP of Development at Viceland alongside Creative Director Spike Jonze, and as a Senior Director of Programming at A&E.

At Starling Lab, Fred has focused on how spatial intelligence and generative AI can be applied responsibly to non-fiction storytelling — and what verification frameworks are needed as captured and synthetic media become increasingly difficult to distinguish. His work includes developing a tranparent media “ingredient list” guidelines for AI creative workflows, research into 3D reconstruction of archival and historical materials, and contributions to Stanford’s EE292J, Designing for Authenticity.

Voice as Archive

One of the earliest fault lines emerged around synthetic voice.

The controversy surrounding Morgan Neville’s biographical documentary about the late Anthony Bourdain, Roadrunner, centered not on the existence of AI-generated audio, but on its undisclosed use. Neville took a small number of sentences from a text exchange Bourdain had written and brought them to life with a voice clone. Once discovered, the backlash was swift — not because the words were fabricated, but because audiences felt the social contract had been surreptitiously breached. The lack of transparency left the filmmaker’s storytelling intentions exposed to concerns about posthumous consent.

That reaction stands in contrast to Andrew Rossi’s The Andy Warhol Diaries on Netflix, which also used generative voice technology, but did so transparently, in collaboration with Warhol’s estate, and with extensive public explanation. The same act — using AI to allow a person to posthumously speak words never actually recorded — produced a very different reception.

A third case, American Murder: Gabby Petito, complicates the picture further. Even with disclosure and family involvement, some viewers expressed discomfort with hearing private diary entries voiced aloud. Transparency is necessary, but it isn’t always sufficient.

Historical Re-Creations Where Archives Don’t Exist

Another rapidly expanding use case involves historical recreation.

In Free Leonard Peltier, director David France and collaborator Jesse Short Bull used AI-assisted recreations to visualize moments for which no archival footage exists, working in consultation with archival professionals and disclosing these choices clearly to audiences.

Similarly, the Sky History series Killer Kings represents one of the first broadcast television productions to use AI-generated reenactments at scale. The motivation was not novelty, but feasibility: historical series of this scope had become financially prohibitive using traditional methods.

In both cases, AI is not replacing archival evidence. It is extending what can be responsibly represented, allowing history to be visualized where silence once stood.

Identity Protection: Deepfake as Ethics

Some of the most ethically compelling uses of AI in documentary involve protecting human subjects rather than simulating events.

In Welcome to Chechnya, face-replacement technology was used to obscure the identities of LGBTQ+ refugees while preserving their emotional presence. Rather than blurring faces or hiding subjects in shadow, the film composited donated faces from activists onto participants, maintaining eye contact and expression.

A similar strategy appears in Another Body, which profiles victims of non-consensual deepfake pornography while using synthetic techniques to protect them from further harm.

In both cases, synthetic media arguably increases truthfulness by enabling stories that could not otherwise be told without erasing the people at their center.

Generative Form: What Is a Documentary Now?

Beyond content, AI is reshaping documentary form.

In ENO, the biographical documentary about music producer Brian Eno, the filmmakers designed an experience where a different version of the film is shown at every screening. Eno and his collaborator Brendan Dawes built a bespoke hardware unit called Brain One, which uses generative technology to reinterpret hundreds of hours of archive and interviews into a new experience — for the audience and even for the director himself. There is no definitive cut. Documentary becomes a performance, a living system rather than a fixed artifact.

A complementary counterpoint appears in Railbound, a short film by creative director Alex Naghavi, built from the documentary photography of Mike Brodie. Brodie’s still images, long recognized for their raw intimacy, document people living as hobos on trains. Railbound animates these photographs into motion using AI tools, while remaining explicit about its hybridity from the opening credits onward.

Developed in close collaboration with Brodie and grounded in respect for the original archive, the project uses AI to create continuity and movement while leaving every narrative and emotional decision — pacing, tone, sound, performance, color — in human hands.

Placed side by side, ENO and Railbound suggest two futures for documentary form: one oriented toward variability and system-level authorship, the other grounded in archival care and transparent interpretation. Together, they make clear that the question is no longer whether documentaries will become hybrid, but how that hybridity is disclosed, governed, and understood.

Creators and the Contemporary Documentary Ecosystem

These developments are not confined to traditional documentary spaces.

Across the contemporary creator economy, creators are increasingly acting as documentary participants. Often working solo or in small teams, they combine research, narrative, and visuals in ways that bypass institutional commissioning altogether.

AI functions here as a force multiplier, accelerating research, lowering production costs, enabling “impossible” visuals, and expanding who gets to attempt non-fiction storytelling.

This is the optimistic side of decentralization. Stories that would never clear traditional gates now find audiences. At the same time, the risks outlined in my previous post remain, and are amplified by scale.

Platform Anxiety: Authenticity After Abundance

One of the clearest signals that something fundamental has shifted comes not from critics or academics, but from platforms themselves.

In recent public commentary, Adam Mosseri, head of Instagram, has articulated a concern that goes beyond moderation or labeling. His argument is that authenticity itself is becoming infinitely reproducible. As AI-generated media grows indistinguishable from captured reality, the qualities that once made creators matter — being real, being present, being unfakeable — are no longer scarce by default.

What makes Mosseri’s perspective especially relevant is that he doesn’t frame this as a problem solvable through disclosure alone. Labels may help, he suggests, but they will inevitably fail as synthetic media improves. Instead, he gestures toward a deeper requirement: systems that can fingerprint real media at capture, establishing a chain of custody that allows platforms and audiences to distinguish documented reality from simulation.

In other words, Mosseri is pointing toward provenance, not just transparency. This aligns closely with the direction of research efforts like Starling Lab, which focus on authenticating origin rather than chasing increasingly sophisticated fabrications after the fact.

Mosseri also anticipates the human consequence of this transition. As synthetic content floods feeds, audiences will shift from default belief to default skepticism, paying more attention to who is sharing something and why. Surface-level disclosure won’t be enough. What’s required is context that can scale — signals about capture, authorship, and continuity that don’t rely on trust in any single institution.

When platform leaders begin publicly advocating for cryptographic provenance and capture-level verification, it signals that the challenge has moved from cultural concern to infrastructural necessity.

Acceleration in the Wild: Venezuela and Creative Proximity

If the earlier examples illustrate careful, intention-driven experimentation, recent events around Venezuela show how quickly creative practice can collide with live reality.

Following reports of political upheaval involving Nicolás Maduro, social platforms were inundated with misleading visuals: AI-generated images, recycled footage, and out-of-context photographs presented as breaking news. This dynamic was documented in reporting that showed how outdated visuals and synthetic imagery fueled widespread confusion during the unfolding events.

What made this moment especially revealing was not only the volume of misinformation, but the breadth of participation. The influx of synthetic imagery did not come exclusively from coordinated disinformation campaigns. Much of it emerged organically from within the AI creator community itself.

Well-known AI artists — many of whom focus on education, tutorials, and speculative storytelling — used the same tools they employ for short films and creative experiments to respond to the unfolding events. These works were not framed as documentary claims or journalistic interventions. They were acts of expression, produced in real time, using tools designed for cinematic realism.

One such example involved a popular AI creator demonstrating how Nano Banana Pro could generate a hyper-realistic, multi-panel “photo-documentary” imagining a fictional operation surrounding Maduro’s removal from power. The exercise was presented as a technical and creative demonstration, not reportage. Yet the imagery adopted the full visual grammar of breaking news.

None of this content was intended to mislead. But it underscores a real shift: creative tooling now sits so close to reality that expressive experimentation and perceived documentation can collapse into the same visual space, especially during moments of crisis.

Designing Authenticity Into the Experience

This collision brings us to the design challenge ahead.

If documentary forms are fragmenting — from feature films to feeds, clips, and explorable virtual environments — then authenticity can no longer remain implicit. It cannot live only in credits, press materials, or after-the-fact explanations.

The shift must be from “trust me” to “here’s how this was made.”

Consider the nutrition label, and what a phenomenon it’s been for the way people interact with what they consume and their health. It’s no coincidence that we talk about a media diet — media is something we consume too, and the relationship between maker and audience carries a similar obligation. What the nutrition label did for food — creating a shared, legible contract between manufacturer and consumer — is exactly the kind of user experience worth designing toward here.

Standards like C2PA are doing the foundational work, establishing the ingredients list: a verifiable record of what a piece of media is made of and how it got there. That’s the necessary starting point. But the infrastructure only matters if it reaches people in a form they can actually use.

The nutrition label works because it operates at two levels at once. Most people glance and move on — the presence of the label is itself a signal of accountability. The person who wants the full breakdown can find it. One standard, two levels of engagement, billions of transactions, all of it embedded in the act of buying food rather than added on top of it.

That layered dynamic is exactly what’s missing from how we currently present media provenance. We’ve been designing for the researcher, not the viewer. Movie ratings and game ratings point toward the right instinct: trust signals need to be ambient and low-friction before they can be meaningful. The original blue checkmark worked the same way — a simple icon implying a verified relationship between an identity and an institution.

Imagine extending that logic to YouTube, where most non-fiction content is now consumed. A small, color-coded icon — visible at a glance — tells you immediately where a piece of content sits on the spectrum from fully documented reality to openly synthetic work. A click, or even a QR scan, takes you deeper: into the metadata, the sourcing, the full ingredients list for anyone who wants it. Something that intuitive, baked into the YouTube experience, could just as easily travel to streaming platforms, broadcast, and social video. The standard expectation becomes: before you consume, you already know what you’re getting into, and you know where to go if you want to know more.

(UX Design suggestion: adding a simple icon inside the existing interface, color-coded to correspond to simple criteria akin to Movie or Game ratings, on a spectrum reflecting fact, fiction, and the space in between measured on a transparent system. Clicking the icon reveals verified metadata, + keyword tags for a range of formats including “verite”, “op-ed”, “investigative”)

(UX Design suggestion: After clicking eye icon, detailed information display including sources used to make the documentary, disclosure on AI usage, provenance info such as C2PA, and modification history info.)

That also means making meaningful distinctions that currently get flattened. An op-ed is not investigative journalism. A produced reality show is not a verified documentary. These don’t need to be condemned for what they are — but they shouldn’t share the same unlabeled space either. The label doesn’t judge the content. It just tells you what’s in it.

The technical foundation exists. What’s been missing is the will to treat media provenance as a consumer experience — something designed into the transaction, not bolted on afterward.

Experimentation Is the Point

There are no final answers. Documentary has always evolved through practice — through artists testing boundaries, audiences reacting, and norms slowly forming. AI accelerates this process, making both the risks and the possibilities more visible.

What matters now is not certainty, but intention. Transparency. Infrastructure that supports interpretation rather than replacing it.

Documentary remains one of the few spaces where society can rehearse how to live with ambiguity. In that sense, it is not being displaced by AI — it is being asked to lead.

Authenticity, Semiotics, and the Collapse of the Myth of Objectivity

Fred Grinstein is a documentary producer, media executive, and educator working at the intersection of non-fiction storytelling and emerging technology. His work spans legacy television, independent documentary, and research collaborations focused on media authenticity and synthetic media.

Through the Looking Glass (Again)

We’ve been somewhere like this before.

In 1896, early film audiences reportedly recoiled when a projected image of a train appeared to rush toward them. Whether or not the story is literally true, it captures something enduring about how new media can briefly destabilize perception. Early cinema collapsed the boundary between representation and reality, and over time audiences adapted. They developed a form of media literacy: an ability to read the frame, recognize artifice, and distinguish cinematic depiction from the world itself.

That arc — shock, adaptation, literacy — has become the story we tell ourselves about every new medium that emerges. It reassures us that confusion is temporary, and that our understanding will catch up.

In the summer of 2025, a short AI-generated video depicting bunnies bouncing on a trampoline circulated widely online. Many viewers initially assumed it was real. The reaction wasn’t panic or awe, but something quieter: casual plausibility. The video did not feel extraordinary. It felt believable. And unlike early cinema, there was no stable grammar for audiences to learn from — no consistent cues to decode, no shared visual language that reliably signaled illusion.

This is where the Lumière analogy begins to break down. We are no longer dealing with a medium that can simply be “read” into literacy. We are through the looking glass, in a media environment where the signals that once helped us distinguish reality from artifice are themselves synthetic, automated, and endlessly recombinable.

The discomfort many of us are experiencing now is not simply about fake videos. It reflects a growing awareness that the infrastructure we assumed was holding non-fiction media together may never have fully existed. What sustained trust instead was a fragile system of consensus built from institutions, conventions, and shared habits of interpretation. It went unexamined because it worked well enough.

That system has been under strain for some time. Artificial intelligence did not introduce the pressure, but it is accelerating it toward a more complete and undeniable state. Media literacy remains important. But unlike earlier media shifts, the pace, adaptability, and scale of synthetic media now evolve faster than shared norms can stabilize, making the emergence of a universally reliable literacy far less certain.

Part I: A Moment of Duality

AI is producing both inspiring and troubling outcomes at the same time. It is opening up creative possibilities, lowering barriers to entry, enabling stories that were previously too expensive, dangerous, or fragmented to tell, and expanding who gets to participate in cultural production. At the same time, it is amplifying disinformation, collapsing visual trust, and normalizing distortion at unprecedented speed.

It’s overly simplistic to frame this as a debate about whether AI is “good” or “bad.” A better question is: why was our media ecosystem so unprepared to absorb its effects?

One possibility is that we overestimated the strength of the trust systems we thought were already in place.

This strain did not emerge entirely by accident. Alongside scale, habit, and technological change, shifts in incentives and distribution in television, streaming, and social media gradually loosened the connection between credibility signals and shared accountability. What AI does is make that loosening impossible to ignore.

The Semiotics Problem

Non-fiction media has always relied on semiotics — visual and narrative signals that communicate seriousness, authority, and credibility. Camera framing, lower thirds, interview posture, typography, tone. These cues did not guarantee truth, but they created a workable proxy for trust.

Reality television played a significant role in blurring these signals. Journalistic tropes were reused as entertainment grammar across the genre. Sitdown interviews with competition reality contestants gave an air of “newsworthiness” to their presentation. Documentary filmmaking techniques once reserved for meaty topics like war zone coverage were deployed to bring dramatic weight to unusual workplaces, or to the personal lives of fabulous housewives. And without a newsroom and its journalistic standards, these “verite” style productions openly used scripts and direction to deliver for the viewer. This was not deception by design; it was optimization for scale, legibility, and production efficiency. At the time, it felt harmless.

I remember once describing a show I worked on as “basically like wrestling,” referring to the fact that viewers watch indifferently to the authenticity of the action before them. The comment stayed with me — not because it felt wrong in the moment, but because, in retrospect, it captured something quietly shifting. We were teaching audiences that the language of credibility could be separated from the obligation to truth.

This dynamic was not confined to reality television. Social media platforms soon abstracted the same signals even further. Fake newspapers succeeded not because their claims were persuasive, but because their fonts, layouts, mastheads, and column structures looked right. Credibility work was being done before content was even processed. Semiotics detached from institutional accountability.

You might even argue that the fake documentary style of TV sitcoms like The Office or Modern Family, or the success of found-footage horror films like The Blair Witch Project and Paranormal Activity, also diluted trust signals along the way, each one innocently exploring new entertainment territory.

AI completes this trajectory. Today it automates credibility aesthetics at scale. Semiotics made credibility legible, platforms made it viral, and AI makes it effortless.

The Collapse of Top-Down Narratives

For much of the twentieth century, non-fiction media operated within top-down structures. Broadcast news, major publications, and editorial gatekeeping created shared reference points. Some reflect with nostalgia on an era when newsmen like Walter Cronkite offered a broadly accepted standard for the daily record. In retrospect, it’s also generally acknowledged that these same individuals and the newsrooms they fronted lacked the representation we’ve come to expect — whether in terms of gender, race, or lifestyle. These systems were imperfect and exclusionary, but they provided consensus.

That architecture has fractured. Information now flows through preference-driven feeds, citizen journalism, influencers, and self-published media. This decentralization has expanded representation — a real and necessary benefit — but it has also dissolved shared baselines.

The warts of this system are uncomfortable. In a fully decentralized media environment, young creators with large audiences and no editorial oversight can publicly speculate about the Holocaust, not through evidence, but through the performance of doubt. This is not an argument against free speech. It is an acknowledgment of what unfettered access to broadcast power includes.

The problem is not that more people can speak publicly; it’s that we never designed shared norms, literacy, or infrastructure for a world in which everyone can.

Avatars are the logical endpoint of this shift. As authority fragments, trust migrates to interface, to the affect of a trustworthy performance, and to always-on availability. Synthetic personalities scale certainty and intimacy far more efficiently than institutions ever could.

AI did not cause this collapse. It accelerated it.

The Myth of Objectivity

Documentary has never been objective. That idea was always more aspiration than reality.

Filmmakers from Errol Morris to Werner Herzog have long argued that documentary media itself is not a transparent window onto the world, but a construction shaped by choices, context, and interpretation.

We see this tension across the documentary canon itself. Michael Moore’s films openly foreground his perspective, collapsing the distance between journalist and participant. Films like Exit Through the Gift Shop or Author: The JT Leroy Story openly embrace sleight of hand as techniques to deliver what their filmmakers see as a methods to unpack deeper more compelling truths about their dynamic subjects. Even highly polished “authorized” documentaries — celebrity-driven biographical series like Beckham where the subject participates as en EP in shaping the narrative — carry the aesthetic weight of documentary while transparently advancing a point of view.

Audiences have long understood this. We don’t watch these works expecting neutral arbitration; we watch them expecting interpretation. Documentary has always contained this elasticity between reportage and argument, between record and persuasion. Truth, in this tradition, emerges through investigation and transparent process — not through the camera alone.

AI makes this impossible to ignore. When images, voices, and scenes can be generated or altered, the fantasy of mechanical objectivity collapses. What remains is accountability.

Part II: From Authenticity to Infrastructure

If debates about documentary truth are to move forward, they must move past an obsession with detecting fakery after the fact.

This is where the work of Starling Lab becomes instructive. Rather than attempting to identify synthetic media once it has already circulated, Starling focuses on infrastructure for authenticity: methods to capture, store, and verify media across its lifecycle.

This provenance-first approach emphasizes establishing a verifiable chain of custody from the moment media is created, anchoring metadata such as time, location, and source at capture, and preserving it through secure storage and verification. The goal is not to define truth, but to reliably distinguish documented reality from simulation.

In the Starling Framework, a Root of Trust is the foundational source within a system where cryptographic trust is first established, ensuring the authenticity and integrity of digital records. Establishing a clear “root” is essential for creating a verifiable audit trail that tracks the creation and entire lifecycle of content, regardless of whether it is AI-created synthetic media or traditional content filmed with a camera. The Root of Trust serves as the bedrock for a “trustless” system, where any third party can independently confirm a digital asset’s origin and history without relying solely on the reputation of a centralized authority.

Capture

The Starling Framework prioritizes establishing this Root of Trust at the Capture phase — or, in the case of generative AI work, the Creation phase — as close to the original event as possible. Ideally, this is anchored in the hardware or firmware of a device, such as a camera’s sensor, a smartphone’s secure enclave, or a computer’s GPU, to bind “birth certificates” of authenticity to an asset before it can be manipulated at the user level. By systematically hashing and signing digital media at this earliest stage, the framework creates an immutable provenance trail that remains with the data throughout its entire lifecycle.

In the Starling Framework, both “real” content created with an actual camera and AI-generated synthetic content are treated equally. Both get “birth certificates” of authenticity. The framework provides the transparency necessary for audiences to evaluate what they are seeing from an informed position.

Store

Digital assets naturally decay due to hardware failure, media obsolescence, or malicious attacks, making resilient preservation essential. Documentary filmmaking, journalism, and other disciplines also require edits and adjustments to content as part of standard workflow: interviews are edited for brevity and clarity, photos are cropped and toned, and historical images are restored and repaired.

Starling addresses this by leveraging tamper-evident data structures and constant audits to demonstrate integrity over time and immediately expose changes. Integrity proofs from the capture step are registered on consensus-based ledgers (blockchain), creating an immutable record of the content and its time of creation that remains accessible even if original hardware fails.

Verify

Starling Labs embraces both human and technological approaches to verification, creating digital audit trails akin to a notary’s physical ledger.

Starling workflows surface cryptographic integrity evidence and immutable registrations, allowing audiences to evaluate the evidence for themselves rather than relying on blind trust. These methods automate complex parts of the verification cycle, removing opportunities for human error and providing real-time authentication.

Ultimately, this enables current and future inspectors to definitively understand the provenance and integrity of digital media throughout its entire history.

The Starling Lab Framework’s provenance and verification do not create trust, nor do they resolve meaning or ethics. But they help stabilize the preconditions for trust by establishing a shared and verifiable record of what happened, when, and where.

Without baseline agreement on those fundamentals, discourse collapses into endless skepticism. With them, we can argue productively about interpretation, responsibility, and impact.

In other words, authenticity infrastructure allows us to move the conversation up the stack.

The Opportunity

AI did not break the documentary. It revealed the fragility of relying on habit, authority, and convention to convey authenticity rather than on designed systems.

That realization is unsettling, and understandably produces fear and pessimism. But it is also clarifying.

The opportunity now is not to restore a mythic past of objectivity, and not to surrender to relativism, but to build infrastructure that acknowledges how non-fiction actually works: as a negotiated, interpretive, consensus-driven practice.

Today’s audiences expect credits at the end of a documentary that list the director, producers, music credits, and all the people who participated in creating the work. What we are suggesting would be an extension of those credits, functioning similarly to an ingredients list on food packaging. But instead of words printed on a cardboard box, it would be addresses on a blockchain ledger pointing to the primary sources used to create the documentary. Further entries could list the modifications made to those sources in the making of the film. In this way, audiences would know not only who made the documentary, but what raw ingredients went into it, and how they were processed to arrive at the final result.

Documentary sits at the center of this moment not because it is uniquely threatened, but because it has always operated at the boundary between fact and meaning. AI simply makes that boundary visible.

What we choose to build there — soberly, optimistically, and intentionally — will determine whether non-fiction remains a space for shared understanding, or dissolves into noise.

Limits of the Argument

This essay does not argue that artificial intelligence is the sole or primary cause of declining trust in non-fiction media, nor does it attempt to resolve the political, economic, or psychological forces shaping this moment. It also does not claim that new forms of media literacy will not emerge, or that decentralization is inherently harmful. Rather, it focuses on a narrower observation: long-standing weaknesses in how non-fiction media relied on consensus, semiotics, and institutional habit are now being accelerated by synthetic media at a scale that challenges individual cognition. The infrastructure questions raised here are not presented as sufficient solutions, but as necessary conditions for meaningful discourse to continue.

I’ve been experimenting with 3D reconstruction from archival photographs for the past year at Spatial Lab, and when we started in early 2025, I wasn’t sure high-quality single-image-to-3D was even possible. Gradually I developed processes with diffusion models that created convincing 3D scenes from 2D images, but they were painstaking manual processes that took hours to complete. Now, with Apple’s SHARP, there is a fast, opensource, image-to-3D process freely available.

SHARP converted a single 1930s Dorothea Lange photograph I fed it into a three dimensional object in seconds. A flat image from the Library of Congress became something I could manipulate and view from different angles.

From a single photograph to 3D. No multi-angle capture or orbiting video required. For someone like me who is used to doing 3D capture with multi-camera rigs, capturing hundreds and sometimes tens of thousands of images to reconstruct scenes, this felt astonishing.

In December, Apple released SHARP, which they describe as “photorealistic view synthesis from a single image.” A more practical description: it’s a 2D-to-3D depth extraction process that turns any photograph you feed it into a 3D scene. It works on a mobile phone photo shot today as well as it does on a hundred-year-old archival image.

The code is open, it runs on Mac, Linux, and Windows, and you can install it yourself or use the prepackaged version in Pinokio. If you’ve ever struggled to get a breakthrough research paper’s code running on your own machine, only to discover it requires an 8-GPU cluster and a terabyte of RAM, you’ll appreciate what Apple has done here. SHARP is genuinely accessible, and that matters because accessibility is what turns a research curiosity into an infrastructure problem. When anyone can do this, everyone will.

And everyone did. Shortly after its release, my LinkedIn feed and 3D Gaussian splat hosting sites like SuperSplat were flooded with SHARP outputs of photos of all kinds turned into 3D.

To be honest, at first I had mixed feelings about SHARP. While it does feel magical in the way it adds depth to 2D photos, the results are fairly limited. The resultant 3D is only really viewable a few degrees off center. After that, the limits of monocular depth extraction become obvious. Faces and objects are curiously squashed when viewed from the side, and the more you move away from the original viewpoint of the photo the more the reconstruction deteriorates. But within the limitations the results are truly sharp, retaining the photorealistic and crisp character of source images where diffusion model processes often turn mushy.

Part of me was also frustrated that the painstaking processes were now obsolete, and a unique capability I’d once enjoyed was now available to everyone and their dog at no effort. But that’s just life in a fast-moving discipline, I suppose.

SHARP doesn’t generate worlds. It doesn’t hallucinate scenery beyond the frame or fill in what the camera didn’t see. What it does is take the flat image you give it and infer depth: Deciding what’s in the foreground, what’s in the background, and the general shape of objects in the scene. The result is a 3D relief of the original photograph, not an environment you can walk through.

That sounds modest. But it sits at one end of a spectrum that extends to tools like World Labs’ Marble, Google’s Project Genie, and Tencent’s HunyuanWorld, which produce full 360° 3D environments from text and image prompts. At that end of the spectrum, the provenance question is blunt: how much of what I’m seeing exists in any source material at all? The answer, often, is very little.

SHARP’s provenance question is subtler, and in some ways harder. Every pixel you see in a SHARP output comes from the original photograph. Nothing has been added. But the spatial relationships between those pixels — what’s in front of what, how far apart objects are, the curvature of surfaces — are entirely inferred by the model. The image is real. The depth is a guess. And the guess is informed not by the photograph itself, but by the model’s understanding of how the world generally looks.

For cultural memory and education, this distinction might not matter much. A 3D rendering of a historical photograph that adds a sense of depth without fabricating content is a powerful tool for engagement. But for evidence and documentation, the distinction matters enormously. If a legal team presents a 3D reconstruction of a crime scene photograph, the jury needs to know that the spatial relationships they’re perceiving, such as which object was closer to which, whether a doorway was within reach, are inferences, not measurements. The image says “this is what was there.” The depth says “this is roughly where we think it was.” Those are different claims with different evidentiary weight.

What we observed. In our experiments converting archival photographs with SHARP, a few things stood out.

The depth estimation is more convincing than you’d expect, and that’s potentially a problem. When we converted FSA photographs from the 1930s, the results felt spatially plausible. Objects appeared to sit at reasonable distances from each other. Rooms had a sense of volume. But nothing in the output distinguishes estimated depth from measured depth. A viewer has no way of knowing whether the spatial arrangement they’re seeing reflects the actual scene or the model’s best guess about scenes that generally look like this.

Right now, SHARP’s own limitations provide a kind of accidental honesty when you move a few degrees off center and the illusion breaks, signaling the boundary between what the model knows and what it’s guessing. But it won’t always be this way. As monocular depth estimation improves, the artifacts will shrink, and the line between measured and inferred will become invisible. The time to build trust infrastructure is now, while the seams are still showing.

And regardless of how good the depth estimation gets, metadata doesn’t survive the conversion. A photograph with C2PA content credentials enters the SHARP pipeline; what exits is a 3D object with no connection to the original provenance chain.

The spectrum from monocular depth estimation to full world generation is filling in fast, and the provenance challenges are different at each point. For SHARP, the question is: are the spatial relationships documented or inferred? For world models, the question is: how much of this environment exists in any source material at all? Both need answers, and both need different labels.

At Spatial Lab we’re working on a provenance layer that persists through reconstruction, for any given 3D creation spelling out what the inputs were and what processes and transformations occurred. We’ve been prototyping what we call a “nutrition label” for synthetic media, and the need for it just became a lot more urgent.

If you’re working with archival photographs, spatial documentation, or digital evidence and grappling with these questions, we’d like to hear from you — especially if you’ve tried to maintain provenance through a 3D reconstruction pipeline. Reach out at info@starlinglab.org.

Spatial Lab is a publication of Starling Lab, a joint initiative of Stanford University and USC focused on data integrity. We cover spatial intelligence technologies for journalism, law, and historical documentation.