The future of AI in photography: four grounded bets
Most writing about the future of AI in photography is a press release with the dates filed off. This is an attempt at something more useful: four things I think happen next, each one anchored to a fact you can go and check today, each one labeled clearly as my opinion, and each one with the specific thing that would prove me wrong.
The rules I am writing under
Three rules, stated up front so you can hold me to them.
Anything about the present is linked. If I say a tool does something today, there is a link to the vendor's own page saying so. Where the industry has not measured something, I say that instead of producing a number.
Anything about the future is labeled opinion. Not "experts predict", not "the industry is moving toward". Mine, with my name on it, and a note about where my interest lies: I am building a photo picking app, which makes me an interested party in at least two of these bets. Read them with that in mind.
Every bet comes with a falsifier. A prediction you cannot lose is entertainment. Each section below ends with the thing that would tell me I was wrong, so this page can be checked in three years instead of quietly rewritten.
If you want the present-tense version of all this, with no forecasting at all, that is a separate piece: what AI has and has not changed about the job in 2026.
The first pass moves off the desk
What is true today. Every serious selection tool is a computer program. Aftershoot's download page says it is available on desktops only. Narrative's help center says flatly that Narrative is not available on iPad. Imagen runs in a browser, but your RAW files have to finish uploading before anything happens. In August 2026, the first pass over a wedding happens at a desk, after you get home.
What changed underneath. The hardware argument is mostly over. Apple's A19 Pro puts Neural Accelerators into every GPU core alongside a 16-core Neural Engine, and in June 2026 Apple shipped Core AI, an architecture for deploying full-scale models locally on Apple silicon. A phone in a coat pocket now has more machine learning capacity than the laptop most culling software was designed for.
My call, and it is opinion. Within a few years the first pass will happen wherever you are, on the device already in your hand, and the desktop becomes the place you go for the handful of comparisons a phone screen genuinely cannot settle. Not because phones win, but because the expensive part of the current workflow is not compute, it is the two hours between the last dance and the moment you sit down. Compare the two positions honestly in phone versus computer culling before you take my word for it.
What would prove me wrong. If in three years the fastest way to review a wedding is still a desktop-only feature, and the mobile versions exist but nobody uses them past the first week, then the bottleneck was never location and I misread it.
Style learning gets very good at sliders and stays bad at taste
What is true today. Style learning already ships, and the vendors are honest about its shape if you read the documentation instead of the headline. Imagen's Personal AI Profile needs at least 2,000 photos you have already edited in one style with the same camera profile. It learns global adjustments: tone, color, texture, the tone curve. It explicitly ignores masking, cropping, object removal and retouching, and it does not adapt to subject matter or lighting. Narrative sells trained color and black and white presets on its Premium and Ultra tiers, and Aftershoot's 2026 roadmap lists profiles you can train on your own images and profiles that keep learning from your adjustments.
My call, and it is opinion. This gets genuinely excellent over the next few years, and it gets excellent at exactly one thing: making the tenth photo look like the first. Consistency is a slider problem and slider problems fall to machine learning reliably. What I do not expect is the jump from consistency to judgment. Choosing between two technically identical frames because in one of them the bride's father has just realized what is happening is not a style transfer problem. It is not obvious that more training data touches it at all.
What would prove me wrong. A tool that, given an unlabeled folder from a shoot it has never seen, picks the frame where the moment turns, and does it consistently enough across photographers that it is not luck. If that arrives, my mental model of what these systems are doing is broken, and I would rather find out from a test than from a demo. Here is how to run that test on your own shoots.
Choosing photos stops being a product and becomes a feature
What is true today. Adobe now ships Assisted Culling inside Lightroom Classic as an early access feature. It scores subject focus and eye focus, checks whether eyes are open, flags misfires and exposure problems, and then applies flags, ratings and labels or builds a collection from the picks. That is a real chunk of what standalone tools charge for, and their published pricing ran from $10 to $60 a month when I checked in August 2026. It now arrives inside software most photographers already pay for.
My call, and it is opinion. Bundling wins the middle of the market. The photographer who wants a decent automatic first pass and no new software will take the one already inside their editor, and standalone tools will stop competing on "we can find blur" because that stops being worth money. What they compete on instead is the stuff a bundled feature is structurally bad at: where the work can happen, how the handoff behaves, whether your decisions survive a move between devices, and how quickly you can overrule the machine. That is also why the interesting comparison in two years will not be accuracy at all, a point we make in the current culling software comparison.
What would prove me wrong. If Adobe's feature stays in early access, stays inconsistent, and photographers keep paying separately for a better one, then selection was a harder product than I am giving it credit for and the specialists keep the category.
Three things I do not think will happen
Predictions are cheap in the direction of more. Here are three in the other direction.
Nobody delivers a gallery they did not look at. Not because the software will not offer it, but because the failure mode is unsurvivable. One gallery missing the grandmother who died in March costs more than a decade of saved evenings. The photographers I know who use these tools most heavily are also the ones who look at every frame the machine set aside, and I expect that to stay true no matter how good the scores get. What happens to those frames matters, which is why we wrote about the photos you do not pick.
Nobody generates the wedding. The commercial and stock damage is real and documented, but it lands on photographs that exist to illustrate an idea. A wedding photograph exists to prove a day happened, in a room, to specific people. That is not a prompt.
The industry does not standardize on erasing files. Every mainstream tool today marks: flags, ratings, labels, collections. My read is that this is not conservatism, it is the only design that survives contact with a photographer who changes their mind at 11pm. Any tool that irreversibly discards frames to look decisive will lose trust faster than it gains users, and it should.
What still needs a person in five years
The last column is opinion. The middle one is not: it describes what shipping software does in August 2026.
| The task | Where software is today | My call for 2031 (opinion) |
|---|---|---|
| Finding blur, missed focus, closed eyes | Solved well enough that doing it by hand is hard to justify | Fully automatic, invisible, nobody discusses it |
| Grouping near-identical frames | Standard across every tool | Automatic, and better at knowing when a burst is one moment |
| Matching your color | Works from thousands of your own edits, global adjustments only | Very good, still global, still needs your hand on the odd frame |
| Choosing between two good frames | Proposed, then you approve | Still yours, on the frames that carry the story |
| Deciding what a gallery is about | Not attempted by anyone | Still entirely yours |
| Telling a client the truth about their day | Not attempted by anyone | Still entirely yours, and increasingly the product |
Read the bottom half of that table as the actual answer to "will AI take my job". The parts of the work that are legible, repeatable and boring are going. The parts that require a person to have been present, to have an opinion, and to be accountable for it are not going anywhere, and the more automated the top half becomes the more the bottom half is what you are actually selling.
How to prepare without betting on any of this
The useful thing about these four bets is that the preparation is the same whether I am right or wrong.
Write your criteria down. If you cannot say why you keep one frame and not the near-identical one beside it, you cannot evaluate a tool that claims to do it for you, and you cannot train anyone else either. Start with your own selection criteria.
Trial on finished work, not live work. Run a free trial over a shoot you have already delivered and compare against what you chose. Aftershoot gives 30 days with no card, Narrative offers a free trial of Ultra, Imagen gives two free culling projects. The disagreements are the whole result.
Avoid long lock-ins while the category is moving. Annual billing is usually a third cheaper, and if bet three is right, the price of this capability falls. Pay monthly until the shape of the market settles.
Keep your originals, always. Every prediction here assumes you still have the frames. Any workflow that depends on being sure the first time is a bad workflow regardless of who is doing the choosing.
Where Kepla actually is
So that nothing above reads as a roadmap: our photo picking app for iPhone, iPad and Mac is still being built. When it ships it picks your best shots and marks them, and it never deletes, moves or renames a file. What works today is the free booking page, which holds a date and takes a deposit. That is the honest status, and I would rather state it than imply more.
If you would rather read about what is already true than what might be, the companion survey linked at the top has no forecasting in it at all. And for phone-side reality as it stands today, there is the state of picking photos on an iPhone or iPad.
FAQ
Will AI photo culling ever run on a phone?
The hardware case is already settled. Apple's A19 Pro adds Neural Accelerators to every GPU core, and its Core AI framework lets developers run full models locally on Apple silicon. The software has not followed yet: Aftershoot is desktop only and Narrative states it is not available on iPad. My opinion is that this changes within a few years, and I am building in that direction, so weigh it accordingly.
Will AI ever learn my personal style?
Partly, and it already does. Imagen's Personal AI Profile learns your global adjustments from at least 2,000 of your own edited photos, and Narrative sells trained presets on its higher tiers. What that captures is consistency of color and tone. It is not the same as taste, since none of these systems adapt to subject or lighting, and choosing which moment matters remains a human call.
Will AI replace wedding photographers?
In my opinion, no, and the reason is structural rather than sentimental. Generative models compete with photographs that illustrate an idea, which is why documented losses sit in commercial and stock work. A wedding photograph exists to prove a specific day happened to specific people in a specific room. Someone has to have been there, and the client is buying that as much as the pictures.
Is standalone culling software going to disappear?
Not disappear, but I expect it to change what it sells. Adobe now ships Assisted Culling inside Lightroom Classic as an early access feature, which puts a decent automatic first pass in software many photographers already pay for. Specialist tools will likely compete on where the work can happen and how cleanly decisions move between devices, rather than on finding blur.
How do I test any of these tools properly?
Use a shoot you have already delivered, so you know the right answer. Run the tool over the same folder, then compare its selection against yours and look only at the disagreements. Count how many of its misses were frames you would have been embarrassed to lose. Two or three real jobs will tell you more than any published accuracy figure.
Should I wait for better tools before changing my workflow?
Waiting costs you the evenings you would have saved in the meantime, and the habits transfer anyway. The preparation that pays off regardless is writing down your selection criteria, keeping every original, avoiding long contracts while prices are still moving, and testing on finished work. None of that depends on which prediction turns out to be right.
Never sort photos at 1 AM again.
Kepla for Mac clears the obvious misses from a card, names the reason on every frame it sets aside, and leaves the choosing to you. Nothing is ever deleted, moved or renamed. Free through the private preview · the first hundred photographers keep it at $99 a year.