The Restoration
A tall tale, yes, but a cautionary one about generative "preservation."
The first time I heard someone say that generative AI could have the capacity to manipulate our history, I dismissed it as the sort of prediction people make whenever a new technology appears. We have a habit of assuming every advancement could either save civilization or end it, and reality usually lands somewhere closer to the centre. The printing press did not destroy memory. Photography did not eliminate lying. Photoshop did not make every image meaningless.
False histories are not always created through elaborate conspiracies. They appear slowly, through repetition. A mistake appears in an obituary, which is copied into a family tree, which is cited by a local historian, which eventually becomes the source used to correct someone who thought they remembered what actually happened. The right people have died, the newspaper has been poorly digitized, and surviving relatives doubt whether their own memories can be trusted.
I accepted the Mercer job without much thought.
An estate company had been hired to empty a farmhouse after the last surviving member of the family died. There were no children and no close relatives prepared to inherit the property. The furniture would be auctioned, the house sold. The local historical society had expressed interest in the family’s photographs, provided someone could determine whether they contained anything of significance.
There were, according to the estate manager, somewhere between three and four thousand of them. She said this apologetically, as though she had personally taken every picture.
“They’re mostly in boxes,” she told me over the phone. “Some albums too. We found slides, negatives, old discs. It’s a lot.”
“How much of it is labelled?”
She laughed, which answered the question.
The estate company needed someone willing to compare the photographs against public records, newspaper archives, cemetery databases, school yearbooks, property documents, and whatever scraps of family history had already found their way online. The historical society wanted names, dates, locations, and just enough supporting information.
Those are modest ambitions until you have tried to meet them.
The photographs arrived in twelve cardboard storage boxes on a Thursday morning. One of the boxes originally held liquor bottles, another was for an eletric mixer, and several bankers boxes. Someone had written MERCER PHOTOS across the top of each one in black marker, although several also contained tax receipts, Christmas cards, appliance manuals, and the sorts of paper that accumulates in houses where nothing is thrown away because everything could matter someday.
The prints ranged from formal studio portraits to badly exposed snapshots. There were tintypes, cabinet cards, Polaroids, school photographs, wedding pictures, military portraits, and hundreds of small square prints with white borders and dates stamped in the margins. Some were pristine, others had been bent, torn, water-damaged, or stuck together by decades of humidity. The negatives had been stored in envelopes with notes such as LAKE, CHRISTMAS, and 1964 (MAYBE).
There were also six compact discs in cracked plastic cases. Each had been labelled in blue marker: MERCER 1 through MERCER 6.
The estate manager mentioned that a volunteer digitized part of the collection years earlier. She believed the images were also available through an online archive maintained by the historical society.
“Might save you some time,” she had said.
The archive was better organized than I expected. Each image had been assigned an identification number, a short description, an estimated date range, and a list of known names. Many entries included comments from local residents who had identified buildings, vehicles, or distant relatives. The site was not sophisticated, but it had clearly been built by people who cared.
More importantly, the photographs had been restored.
Dust had been removed. Exposure had been corrected. Torn corners had been reconstructed. Faded faces emerged from shadows, and water stains disappeared from pale summer skies. The photographs still looked old; they had not been sharpened into some sort of waxy oversaturated imitation.
There is nothing inherently wrong with restoring a photograph. Preservation often requires intervention. Archivists repair tears, stabilize paper, adjust scans, and remove defects introduced by damaged equipment or poor storage. Digital restoration can reveal what was always there but had become difficult to see.
The important word is reveal. Restoration should help us perceive what the original contains. It should not decide what the original contained and place it there.
That boundary has always been difficult. A person repairing a damaged portrait may reconstruct the edge of an ear or continue the line of a dress across a tear. Before generative AI, every restoration involved a judgment call. A human restorer knows when they are making an inference, but a generative model has no relationship to truth at all. It identifies patterns and predicts what belongs next. It does not ask what was there, it simply asks what would look plausible.
One of the reasons people are so easy to mislead with photographs is that we spend very little time really looking at them. We see a grandmother standing outside a farmhouse and our brain concludes, Grandmother. Farmhouse. Summer. We do not inventory the porch columns, count the fence posts, or study the direction of every shadow. Recognition is efficient, and allows us to understand an image in a fraction of a second.
I began by matching the physical photographs to their online counterparts. This was slower than it sounds. The numbers written on the backs of the prints did not always correspond to the digital filenames, and the volunteer who created the archive had occasionally grouped several photographs under one entry. Some images had been scanned from albums and cropped so tightly that the page itself was no longer visible. Others existed online in multiple versions with slightly different descriptions.
None of that was unusual. Archives are made by people, and people are contradictory.
The first photograph that bothered me was numbered M-1968-044.
Three children stood beside a white wooden fence, squinting into bright sunlight. The oldest boy had his eyes closed. The younger two wore matching striped shirts, although one shirt had been tucked into shorts and the other had not. Behind them was a field, a line of trees, and the edge of a barn that appeared in dozens of other Mercer photographs.
The physical print had a tear through the lower-left corner. It removed part of the youngest child’s shoe and splintered one of the fence posts. The restored image repaired both. The child had two complete shoes, and the post continued neatly to the ground.
What I did not expect was the second horizontal rail.
In the physical photograph, the fence had two rails running behind the children. In the restored version, it had three.
It disappeared naturally behind the children’s bodies, followed the angle of the existing boards, and carried the same paint wear as the rest of the fence. Had I seen the restored version by itself, I would not have questioned it.
I enlarged both images and compared them side by side. The new rail was not just an addition to the torn area but had been added across the full width of the picture, including sections where the original was perfectly intact.
The restoration software must’ve made a mistake.
Older automated tools sometimes extended nearby textures when repairing damage, and newer generative systems are designed to create visually coherent results rather than historically accurate ones. The model saw a fence and predicted a more complete fence.
I recorded the discrepancy in my notes and moved on.
The second inconsistency involved a dog.
The photograph had been taken on a porch, probably in the late 1950s. Two women sat on a wooden swing, each holding a glass. A black Labrador stood partly behind them. In the physical print, only the dog’s head, shoulder, and one front paw were visible. The rest of its body was obscured by the swing.
In the archived version, the dog had been moved slightly forward. Its entire chest was visible, along with both front legs and the beginning of its back. The result was anatomically plausible and consistent with the lighting. It did not have extra paws, distorted fur, or any of the obvious defects people associate with AI-generated images.
I checked whether the archive might have used a different print or negative. Families often took several photographs in quick succession, and similar frames can be mistaken for one another. I searched the boxes, reviewed the discs, and examined every nearby image in the online collection.
There was only one version.
By Friday afternoon, I had documented twenty-three alterations. Most were small. A wrinkled tablecloth became smooth. A missing porch spindle reappeared. A child’s obscured hand acquired fingers that were probably correct but could not be verified. In one photograph, a faded sign on a building became legible enough to read MERCER GENERAL, although the original contained no discernible lettering.
None of these changes were dramatic which made them more troubling.
If a restoration reveals a face hidden in shadow, the detail should exist somewhere in the source data. If it reconstructs an area destroyed by a tear, the result may be convincing, but it is still an inference. If it creates a readable sign where no readable letters exist, it’s manufactured.
I contacted the historical society on Monday.
The woman who answered introduced herself as Elaine. She had volunteered there for more than twenty years and remembered the Mercer digitization project.
“We were very proud of that one,” she said. “People from the family kept finding us online. We got messages from all over.”
“Do you know what software was used to restore the photographs?”
“There were a few different programs, I think. The volunteer who did most of it passed away.”
“Were the original scans retained?”
“I would assume so.”
“Do you know where?”
There was a pause long enough that I heard her keyboard begin clicking.
“We changed website companies a few years ago,” she said. “The new one improved a lot of the images automatically.”
“What do you mean by improved?”
“Cleaned them up, ya know? Removed scratches, made the faces clearer.”
“Did anyone review the changes?”
“Well, there weren’t changes, really, there were enhancements.”
The distinction seemed important to her.
I asked whether I could receive copies of the original, unprocessed scans. She said she would contact the company that hosted the archive and look through the society’s drives.
Before we ended the call, she added, “The restored ones are much nicer.”
The photograph that propelled me forward was in a paper album near the bottom of the ninth box. Its pages had once been black but had faded to a soft charcoal grey. Each picture was held in place by triangular paper corners, and someone had identified the subjects in white pencil beneath them.
The image showed four women standing outside the Mercer General Store in 1947. Three were young, possibly sisters. The fourth was older and wore a dark dress buttoned to the throat. The photograph had been folded before it was placed in the album, leaving a crease through the older woman’s face.
The paper had separated slightly along the fold. A narrow section of the image was hidden beneath the raised fibres, including part of the woman’s left cheek and the corner of her mouth.
I found the matching archive entry and opened the restored version.
The crease was gone. The woman’s face was complete.
At first, I saw nothing that distinguished it from the other restorations. The missing area had been filled smoothly. Her mouth continued across the damaged section, and a small dimple appeared beside it. The result was subtle enough that most viewers would never have known the photograph had been damaged.
I removed the physical print from its paper corners and examined the fold beneath a magnifying lens.
The hidden section remained unclear. The emulsion had cracked and curled inward. There was no visible information from which the shape of the smile or the dimple could have been recovered.
Then I noticed another photograph on the same album page that showed the same woman in profile.
She had the same dimple in precisely the same place.
A generative system trained on a collection could use information from one photograph to complete another. A volunteer could have done the same consciously. The archive contained hundreds of Mercer faces. Perhaps the restoration software had identified the woman across multiple images and reconstructed her damaged features from an intact photograph.
That explanation was technically possible.
It was also far more sophisticated than anything the historical society believed it had purchased.
I searched the archive for every photograph tagged with the woman’s name. There were fourteen. In three, the dimple was visible. In the others, she was turned away, too distant, or unsmiling.
The restored 1947 photograph matched the earlier images exactly.
I sat with the physical print, the archived restoration, and the later portrait arranged across my desk. For several minutes, I tried to decide whether I was impressed or alarmed.
Then I turned the 1947 print over.
Someone had written a note on the back in pencil.
First photograph after Ruth lost the left side of her smile.
“Hmm,” I thought. Perhaps a stroke or some sort of paralysis that left her smile flat an undimpled on the side the photo had suggested
Beneath the note I’d made about Ruth’s photograph, I wrote:
Most likely explanation: Modern restoration software reconstructed missing portions of damaged photographs by referencing other images in the same archive.
It was a sophisticated capability for a volunteer historical society. But it wasn’t impossible. Consumer software had become startlingly good at identifying recurring faces across collections. If the model recognized Ruth in a dozen photographs, perhaps it had simply inferred the missing portion of her smile from another image.
I decided to leave Ruth alone for the afternoon.
The Mercers had lived in the county for well over a century, which meant census records, school yearbooks, church bulletins, property transfers, wedding announcements, obituaries. Local newspapers have a wonderful habit of documenting the lives of ordinary people in extraordinary detail. A county archive can tell you when someone won a pie contest, broke an ankle, bought a new tractor, or hosted twelve relatives for Thanksgiving. It is simultaneously invaluable and deeply nosy.
By the end of the first day I identified nearly thirty people with reasonable confidence.
The work was pleasantly repetitive. Compare a face to a yearbook. Confirm an address. Find a marriage announcement. Check whether the children listed in the obituary matched the people standing beside them in a Christmas photograph. Every confirmed identity made the next one easier.
The archive, for all my growing reservations about its restorations, was proving genuinely useful.
Names aligned.
Dates aligned.
Buildings matched insurance maps.
Vehicle registrations corresponded with the estimated years of the photographs.
If I had been evaluating the archive for historical usefulness, I would have given it an enthusiastic recommendation. But I was trying to determine whether it still represented the photographs it claimed to preserve.
The first interview I conducted was with a woman named Margaret Mercer, who was seventy-eight and still lived about twenty minutes from the farmhouse.
She apologized three separate times for not having made coffee before I arrived, despite the fact that she had no reason to expect I was coming at all. Small town Canadians of a certain generation possess an almost innate ability to feel guilty.
The walls of her living room were crowded with family photographs.
Most had been printed from digital scans.
I recognized several immediately.
“I’ve been looking through your family’s archive,” I told her.
“Oh, it’s wonderful, isn’t it?” she replied. “We never thought we’d have all those pictures back.”
I asked whether she still had many originals.
“A few.”
She disappeared into another room and returned with a small floral tin that once held Christmas cookies.
Inside were perhaps twenty photographs wrapped in wax paper.
“I couldn’t tell you why I kept these ones,” she said.
I selected a photograph almost at random. It showed six people gathered around a picnic table sometime in the early 1960s. I had seen its restored counterpart online the night before.
“Can I ask you something a little strange?”
She smiled.
“I’ve lived long enough, nothing seems all that strange.”
I opened the restored version on my laptop and placed it beside the physical print.
“Which one looks right?”
She didn’t hesitate.
“That one,” she pointed to the laptop.
“What makes you say that?”
“It just does.”
I pointed to one of the men.
“The restored version has him wearing a wedding ring, but the print doesn’t.”
Margaret leaned closer.
“I suppose...” she began, “...I never noticed.” After a moment she laughed softly “Funny, isn’t it? I remember him wearing a ring.”
“But not whether it was actually in the picture?”
She looked at me for a long time before answering.
“I remember him often wearing one. Maybe I’m remembering him instead of the photograph.”
I nodded.
“That’s very common.”
Every time we remember something, we’re rebuilding it from fragments; some visual, some emotional, some borrowed from later experiences. We don’t retrieve memories nearly as much as we rewrite them.
Margaret made us tea. While the kettle boiled she asked what exactly I was investigating.
“I think some of the restored photographs contain details that weren’t in the originals.”
She shrugged, “I suppose that’s what restoration means.”
“No,” I said, perhaps more sharply than I intended. “It isn’t.”
She looked surprised, and I apologized. Then I pulled the fence photograph from my folder and laid the restored image beside it.
“It built a different fence.”
Margaret adjusted her glasses. She looked once and then twice. Then, a third time.
Finally, she frowned.
“I think I remember that fence,” she said, but nothing else followed.
She could no longer distinguish between remembering the day the photograph was taken and remembering the version of the photograph she’d spent the last five years looking at online.
Margaret was not the only Mercer who struggled to separate the photographs from the memories attached to them.
Over the following week I spoke with nine more relatives. Some lived nearby. Others joined by video call. I stopped asking people whether they recognized the individuals in the photographs. Instead, I began asking them to describe the photographs before I showed them.
Tell me what you remember.
One cousin remembered a porch swing that seemed to appear only in a restored image. Another insisted her grandfather’s suspenders had been blue, despite the fact that the original colour slide, one of the few in the collection, showed they had always been brown. A woman in Alberta described the lettering on the Mercer General Store sign in remarkable detail.
The lettering did not exist in the physical photograph.
None of this was especially surprising. Memory is porous. If you show someone the same corrected image often enough, it simply becomes the photograph they remember.
The younger the relative, the more likely they were to describe the restored image.
The older the relative, the more likely they were to remember details from the physical print.
People who had grown up handling the albums remembered the albums. People who had grown up searching for family history online remembered the archive.
A few days later, the historical society called to tell me they had located the original scans.
“They’re coming from the hosting company,” Elaine said cheerfully. “Apparently they keep backups.”
An hour later, a download link arrived in my inbox. The folder contained just over twelve hundred TIFF files, uncompressed, high resolution.
Exactly what I had hoped for.
If you’ve never worked with digital evidence, you might take for granted that a photograph is the photograph. In reality, every save creates another object. A print is an object, a negative is an object, so is a scan. A cropped scan is another. A compressed JPEG sent over email is another again.
One may faithfully represent the other, but they are not the same thing.
Every digital file can be run through a mathematical function that produces a unique value called a cryptographic hash. Think of it as a fingerprint. Change a single pixel, and the fingerprint changes completely. The hash won’t tell you which file is authentic, but it will tell you whether two files are identical.
I generated hashes for the scans from the Mercer CDs. Then I generated hashes for the files the hosting company had described as their original backups.
None of them matched.
Perhaps the CDs had been rescanned years later. Perhaps the archive had recompressed the files. Perhaps someone adjusted the brightness before uploading them. Every explanation I considered was ordinary and accounted for some of the discrepancy. None accounted for all of it.
The dimensions matched, as did the filenames and the scratches in the paper. But when I overlaid the images, they drifted apart in dozens of tiny places.
I called the hosting company.
The support representative was patient.
“Our restoration pipeline has changed several times over the years,” he explained. “It automatically improves material.”
“Right, I get that.”
“The original scans don’t appear to have been modified.”
“But they aren’t identical.”
“They’re functionally identical.”
I wrote that phrase down. “Functionally identical.” It sounded like something a lawyer would say.
I spent the next several evenings tracing the photographs outward instead of backward. Genealogy websites, local history blogs. Family memorial pages and social media.
The same restored images appeared everywhere.
Because they looked better; they were brighter and cleaner. They looked the way people wished old photographs looked. So those were the versions they downloaded, which became the versions they shared with cousins. Those became the versions printed for funerals. Eventually, they became the versions scanned again.
The restorations had not replaced the originals through malice but through preference.
I finished my report three weeks later.
It was longer than anyone had asked for.
I recommended preserving every physical print, every negative, every scan, and every restored image as separate artifacts. Do not overwrite originals, document every enhancement. Treat restoration as interpretation, not preservation.
Record the software used, record the date and the settings.
The historical society accepted every recommendation.
“I suppose we never thought about future researchers,” Elaine admitted.
“You were thinking about the public and what they want to see.”
She smiled.
“We wanted people to enjoy them.”
“I know.”
This is not a true story, but a story nonetheless.