What AI should never choose: the decisions that stay yours

Every culling tool measures the same short list of things, and all of them will keep getting better at it. None of that is the reason a gallery sounds like you. This is the other half of selection, the half that stays yours no matter how good the software gets: the meanings you were told in advance, the promises you made, the consent you owe, and the taste that makes the work recognizable.

Two kinds of decision arrive in one pile

A shoot comes home as one undifferentiated pile of files, which is why the two kinds of decision inside it get confused. They happen in the same hour, at the same desk, with the same keyboard shortcuts, so they feel like one job. They are not.

The first kind has an answer. Is the subject in focus. Are the eyes open. Are these nine frames the same second of the same moment. Write the question down and two reasonable photographers looking at the file would agree on it. Adobe's Assisted Culling scores exactly three questions of this type, subject focus, eye focus and eyes open, and we went through all of them in our review of the feature. Every other tool in the category measures a near identical list, and that is not a criticism of any of them. Measurement at one in the morning is the worst work in photography, and handing it to something that does not get tired is a good trade.

The second kind has no answer of that sort. Which of two true frames belongs to this family. Whether the picture of the toast is the man speaking or the face of the woman listening to him. Whether the guest laughing in the background agreed to appear in a gallery that stays public for a decade. These are not harder measurement problems waiting for a better model. They are a different species of question, and a machine's answer feels hollow not because the model is weak but because the information was never in the file.

So this article is not a list of things AI gets wrong. We keep that list, honestly and in detail, in what AI culling misses. This is the other list: the decisions that would still be yours if every tool in the category became perfect tomorrow morning.

Write one sentence about the job before you open the folder

Here is the cheapest habit in this article, and the one most likely to change a gallery. Before the first pass, write a single sentence describing the job in human terms.

"This was a loud, affectionate city wedding that went completely quiet during the father's speech." "This brand launch was precise, optimistic, and full of people making things with their hands." "This family does not pose, and everything good happened after the four year old stopped cooperating." Write it in the car on the way home, while you still remember the room.

Then use the sentence as a filter over the shortlist. You are no longer asking whether a frame is a good photograph, a question with a thousand defensible answers. You are asking whether it is that day, which has far fewer. A star rating compares a frame to every photograph you have ever taken. The sentence compares it to one event, and one event is the thing your clients are actually buying.

It changes small decisions too. How many calm frames the set needs between the peaks. Whether black and white belongs here or is a habit imported from a different wedding.

The meanings you were told about in advance

Somewhere before the shoot there was a conversation, and in it you were handed information that exists nowhere in the pixels. The grandmother who is not well. The brother who flew eleven thousand miles and leaves on Tuesday. The dress that belonged to someone who died. The dog. The friend who has been through the worst year of her life and is finally laughing in the third frame of a burst.

A model looking at that third frame sees a face, in focus, eyes open, one of a group of forty similar frames from the same ten minutes. Everything it can see about the picture is ordinary. The one fact that makes it the photograph of the evening for that family is a fact it was never given, and no amount of training on other people's weddings will supply it.

The practical move is small. Keep a short list of names and meanings in the same note as the timeline, and read it before the final pass rather than after you have sent the gallery. Six names, one word each about why. That note is the difference between remembering the grandmother at selection time and remembering her when the client emails to ask whether there are any more of her.

The same is true of ritual. A moment that carries enormous weight in the room can look, to a general model, like a group of people standing fairly still. Nothing is wrong with the model. It was not in the room.

Consent is not a score

Some frames are technically excellent and must never be delivered, or must be delivered privately and never published. The guest who asked to stay off social media. The child whose parents said no. A medical detail somebody keeps to themselves. A person crying in a way that was not meant for an audience.

No selection score has access to any of that. Consent lives in a conversation, often a single sentence said in a doorway while you were changing a lens, and there is no field in a RAW file where it can be written. Selection is where that sentence gets honored or quietly forgotten, which makes it a professional responsibility rather than an optimization problem.

Two habits cover most of it. Write restrictions down the moment you hear them, because you will not remember at midnight. And keep the line between delivered and published deliberate: the family can have the frame, your feed does not need it. The client facing version of this conversation is in how to explain the cull to clients.

The unglamorous frames you promised to make

Then there is the half of the gallery nobody will ever repost. The family formals in the exact groupings the aunt wrote out. The grandparent arriving. The accessible entrance that made the day possible for someone. The sponsor banner. The empty venue. The cake the baker will want for her own portfolio.

None of these win anything. All of them are the job. And a selection process built entirely around what looks strong will quietly starve this half of the gallery, because every one of these frames loses a fair fight against a portrait taken in good light.

So keep a must-cover list that sits outside any rating or score, and check it as a separate pass before delivery: requested groupings, key people, the rituals, the venue, the vendor commitments, anything written into the brief. Where the line falls between what you promised and what you would have chosen is worth settling once rather than per job, which is what we mean by having an editorial standard.

The opposite failure is just as real, and more common in careful photographers. A checklist can crowd out the strange, slightly wrong frame that is the only reason anyone hires you rather than someone else. Cover the promises, then put back the two frames that no brief asked for.

What a score answers, and the question underneath it

Every automated criterion in this category has a human question sitting directly underneath it. The machine's version is answerable. The one underneath it usually is not, at least not from the file.

Read the right hand column and notice what it has in common. Every question there needs information that was never in the file: something you were told, something you promised, something you saw with your own eyes from three feet away. This is why the goal is not a more accurate tool. The goal is a clean division of labor, with the left column handed over completely and the right column protected.

Taste is the part that does not transfer

Here is the quiet risk in all of this. If every photographer in a city accepts the same ranked set from the same three or four tools, the galleries in that city converge, at the exact moment couples are choosing between photographers on how different their work feels.

Your visual language is not one grand decision. It is a set of small preferences repeated a few hundred times per shoot: negative space or density, documentary imperfection or controlled portraiture, movement or stillness, one decisive frame or a run of three that builds. A model trained on broad preferences across many photographers is pulled toward the middle of all of them, and that middle is competent and nobody's signature.

Use these tools to take away drudgery, not difference. The technical screen is drudgery, so give it away with both hands. The final call on the frames that carry the signature is the difference, so make it yourself, every time, awake. Some tools promise to learn your preferences from your past edits, and it is worth being clear eyed about what that can pick up, which we look at in training AI on your own style.

Sequence belongs in the same category. The order of images changes what they mean, and no scoring system has an opinion about pacing, contrast or breath. That is a separate craft and we give it its own article in building a gallery with emotional range.

A process that protects the human half

None of this means going back to the slow way. It means spending the recovered time in the four places that deserve it, in this order.

One. Let the software run the technical screen and group the near identical frames. Do not supervise it, that is the entire point of it. Two. Open the human-required minutes in capture order and look at every frame in them regardless of what the software thought: vows, speeches, the reactions during the ceremony, the first two minutes after a first look. Three. Compare the high value near duplicates yourself, at size. Four. Shape the set, then check it against the three lists you wrote earlier.

The check at the end is six questions:

• What was the emotional center of this job, in one sentence?
• Which people, objects and rituals were named to me in advance?
• What has to be in there because I promised it?
• Is there anything here that somebody asked me not to share?
• Does the set have pace, or is it eight hundred good frames in a row?
• Which slightly imperfect frame is more truthful than its cleaner neighbor?

That is ten or fifteen minutes on a wedding, and it is the ten or fifteen minutes clients respond to. The argument for automation was never that selection does not matter. It is that selection is a real skill and it deserves the awake version of you, not the version still going at one in the morning because the technical screen ate the evening.

That split is what Kepla is being built around: an AI first pass on iPhone, iPad and Mac that hands you picks to approve, with no file ever deleted, moved or renamed. The picking app is still in development and we will keep saying so until the day it ships. The booking page is live and free today, if you want the other end of the job handled while you wait.

FAQ