For digital archivists, The End Is Nigh is no longer an impending risk, it is with us today. I’m calling on all digital archivists, and honestly this needs to carry over to all curators of any source of content, to mark 2022 as the first year that all sourcing must be considered suspect for our collections (frankly we should probably say 2020 if we’re being honest).
What do I mean when I say suspect? Let’s step back 24 hours to me before penning this essay and set some reference for this discussion, as yesterday’s me had a wholly naïve perspective on the extent of the problem before us. For the last number of years I’ve been talking to my friends, family and peers about the need to for a system to establish digital chain of custody from a reliable source as a potential method to specifically counter the grown threats of deep fake technologies. While this concern seems to be taken seriously by cyber security professionals and a hand full of journalist, the implications need a wider, serious public discourse.
The concern as only heighted in the last couple of years as the proliferation of digital touch up technologies has flourished and become accessible outside of a professional setting. We now see an army of casual computer users applying these technologies to personal and public data (from restoring 100 year old crinkled photos of great, great grandma to colorizing newsreels from the 1920’s, etc…).
Yesterday my more general concern was the pollution of contemporary history with colorized, digitally enhanced content blending in with the source material and being taking as original. Now I see that the problem is far greater than I could have ever imagined.
Yesterday my wife (who has made a career developing AI/AGI technologies behind the scenes) was playing with the latest and greatest of publicly accessible ML systems for content generating and with relatively little effort composed a documentary video of 100% synthetically generated content that on the surface might be perceived as a factual presentation. By this morning my stomach was clenching as my world view of the risk of digital archive contamination now became 100% tangible and worse. The reality is that this goes beyond a worry about a bad actor into the space of casual contamination by well-intentioned yet uniformed third parties.
For this reason, I implore that anyone doing digital archiving needs to add an extra step of assuring chain of custody and sourcing for literally everything that comes into your possession going forward. The ability to make a convincing period video on any subject is now in the hands of the world, and for example with even the most rudimentary skills someone can render a fully synthetic video to a used VHS tape and send it into a collection for storage and if not caught can risk becoming source content for someone later.
This is going to be a problem for physical collections as well but physical forensics still gives us a way to verify age on materials for now. The immediate risk is digital where information moves so freely between archivists.
ALL of the material you receive going forward should be considered suspect and I ask that all archivists make a heightened effort to notate the date and time of ANY item comes into your collection and if possible any verifiable chain of custody documentation that can be collected as well. YOU as the digital archivist are going to be the last line in protecting the integrity of our past, present and future. Without your help nothing will be real, everything will be possible, history will be lost and fantasy will inform the future.
The Comments
My replies have been lost to the retraction. What survives is the community’s side of a conversation responding to an argument they can no longer read. I’m preserving the comments as they were.
MrSansMan23:
I would imagine that the verification methods used to tell if a painting or old photograph is fake will have to apply to modern ones to help with solving this problem. Eg a lie is hard to keep up cause it has to conflict with all the other real evidence. Eg different camera angles of the same event.
[my reply, deleted]
Yes you could contaminate it with fake footage but every single frame would have to align exactly with what you’re trying to fake. eg shadows, reflections, how the light reflects off someone’s face, the colour of the day needs to all match, eg someone on a green screen in a Studio couldn’t match the brightness of the sun on a clear blue sky day at noon.
[my reply, deleted]
Even if people aren’t watching for it people who know what to look for like I mentioned above can’t tell other people exactly why it’s inconsistent. Eg “at 0:23 in both videos shows the person falling but in the first video the shadows textures on the face don’t match see here side by side for each image” sure you can say the debunk is also faked but the debunkers should have the viewer verify themselves “go find the two videos where ever you want and see if they have the flaw there still”.
[my reply, deleted]
True however I believe that archive as many primary sources is best method to prevent such mistakes as you mentioned in the first paragraph.
A good analogy would be using the 9/11 impact footage of the south tower. All the real footage of it happening match the same smoke pattern just at different angles. You can find the same event being filmed from many angles, different lighting conditions, etc.
nikPitter:
What you consider to be reality is also potentially subjective to the eyes/ears of others. i.e photo and videographic evidence has been highly subject to manipulation and alternative interpretation since it has existed. A sequence of shots edited in different ways can tell completely different stories. Crucial context can be cropped out of photos. There is no shortage of misleading content on all platforms. Yes, AI will accelerate the problem but this I think might improve matters.
My problem has not been so much that untruths are perpetrated by misuse of technology, but that it seems to be of little concern. The gulf of Tonkin incident as a historical example was staged with only words - no photographic evidence. The retraction of this staged incident took place several days later but by that time the stage had been set for the invasion to begin. The consent had been manufactured. The general public of the USA in that case accepted printed word in newspapers as truth without further verification.
The point is that the issue is not technological but human. That some people are credulous enough to believe that photo or video is irrefutable evidence in and of itself.
My hope is actually that AI is going to make this even more of a concern, due to the attention it garners and the ethical concerns it raises. More thought and effort will hopefully be put into education, regulation and critical thinking, and through law. Its optimistic I know. Really I hope it will accelerate what should have been happening since the camera was invented in terms of public knowledge.
I guess this doesn’t negate your points entirely, but who is to say (or how do we prove to the public) that our archived documented version of truth is the real one? This is something for governments, and courts of law to establish precedents around. Assuming there is rule of law, and fair access to it. An interesting age to get into law of many types. Sure if you possess what you believe to be documentary evidence of an event, archive it. At that point its a matter who has the resources to convince interested parties that their ‘reality’ is the base reality.
By the way Ive been keeping an eye on AI graphics also, particularly animation and video, as I work in a CGI related area and am yet to see anything that can produce temporally cohesive video. The closest of concern may be papers demonstrating manipulation of people in videos to change their actions/movement. Pretty interesting. Of course its only a matter of time (like weeks or months) until we do have fully coherent video, but Im curious what your wife has been using for this.
Three and a Half Years Later
I wrote this ten days after ChatGPT launched. The night before, my wife had demonstrated a synthetic documentary video using publicly accessible tools, and what I saw made my stomach drop. Not because the technology was impressive. Because the contamination pathway was obvious and nobody was talking about it.
The video is called “Among Us.” It’s a fully synthetic documentary about self-organizing systems. The presenters don’t exist. The academics don’t exist. The credentials are fabricated. The script was generated. And it’s still on YouTube, where it has sat for three and a half years (it was briefly removed and restored so the original publication date is incorrect) with sixty-eight views and two comments as of the time of this post.
Here’s the tool chain used to build the video:
- Faces: Nvidia StyleGAN2
- Facial animation: MyHeritage Deep Nostalgia
- Photography and graphics: Stability.AI Stable Diffusion, OpenAI DALL-E 2
- Facial correction: Tencent ARC Labs GFPGAN, Snapchat Face Swap
- Video editing: Corel FastFlick with Mixkit elements
- Music: Generative.FM, Wolfram Tones
- Voice over: Murf.AI text-to-speech
- Script: OpenAI GPT-3 via Copy.ai
Every one of those tools has been superseded by something an order of magnitude more capable. She built this with duct tape and determination across a dozen different services. Today you could do the same thing with two or three tools and less effort. The Norbert Wiener quote she put in the video description said it: “it’s the dose that makes the poison.” Each individual tool seemed like a curiosity. The combination was the revelation and an emergence of a new type of content production.
Reading these surviving comments to my post now is an education in how seriously people took the problem in December 2022.
MrSansMan23 built their entire defense on forensic verification. Shadows, reflections, frame-by-frame analysis, cross-referencing multiple camera angles. Their gold standard was the 9/11 south tower footage: real events filmed from dozens of independent angles, where the smoke patterns all match. That model assumes two things. First, that generation tools would continue to produce detectable artifacts. Second, that real events reliably have multiple independent sources to cross-reference against. The first assumption was wrong within months. The second was never true for the vast majority of archival material. One-camera footage, personal recordings, institutional documentation, local events. Most of what fills an archive has exactly one source. There is no second angle to compare against. The forensic defense was built for the exception and applied as if it were the rule.
nikPitter was more thoughtful and more interesting. Their argument was that the problem is human, not technological. People have always been credulous. Manipulation has always existed. The Gulf of Tonkin was manufactured with nothing but words. All true, and all beside the point. The problem I was describing wasn’t about bad actors deliberately manipulating archives. It was about well-intentioned people casually contaminating them. Colorizing a newsreel. Enhancing a photograph. Compositing footage with synthetic elements because it looks better. The contamination I was worried about doesn’t require malice. It requires accessibility. And accessibility is what changed in December 2022.
nikPitter also said something that has aged in a way I don’t think they intended: “My hope is actually that AI is going to make this even more of a concern, due to the attention it garners and the ethical concerns it raises. More thought and effort will hopefully be put into education, regulation and critical thinking.” It is now June 2026. The education didn’t arrive. The regulation is stalled in committee. The critical thinking is losing ground to convenience. The optimism was understandable. The outcome is not what they hoped for.
One more thing from nikPitter: “I work in a CGI related area and am yet to see anything that can produce temporally cohesive video. Of course its only a matter of time (like weeks or months) until we do have fully coherent video.” They were right about the timeline and still treated urgency as optional.
I retracted this post because the heat wasn’t worth the fight. That was the wrong decision. The Archivist crew on the Discord were the most reactive calling me a sensationalist fearmonger and that was where I knee jerked and pulled the post. I was just not in a place in my life right then to have a fullthroated fight with people I considered peers who were willfully in denial of the problem. Sometimes you pick your fights and you pick wrong. Not because the post would have changed anything. It wouldn’t have. But because the act of retracting it is its own small lesson in how warnings die. Not because they’re wrong, but because saying them costs more than practicing silence. That’s a problem I’m still thinking about and have covered elsewhere.
The archival contamination I described in 2022 is now so widely acknowledged that it barely qualifies as a hot take. Chain-of-custody verification is an active area of research. Provenance metadata standards are being developed. None of it is moving fast enough.
The last line of the original post still holds. I wouldn’t change a word.