Why Adobe's AI Agents Are Built for Busywork, Not Your Edits

4 hours ago 2

Adobe used its MAX keynote to show ChatGPT running inside Photoshop and Claude running inside Illustrator, and the reaction from people who edit their own pictures has been quick and skeptical. If an outside AI can operate the same apps you use every day, you have every right to ask where your own judgment fits and whether you'd ever want to hand it the controls.

Since MAX, two comments keep turning up in photography communities, on Instagram, and on YouTube: "AI doesn't know how I edit," and "I like being creative, I like to do it myself." Glyn Dewis agrees with both, and he still thinks most people are misreading what Adobe actually showed. His argument rests on the difference between "an AI carrying out steps inside an app and an AI deciding what your picture should look like." He compares it to a new intern on day one. You wouldn't hand that person your portfolio and ask for edits in your style, since they'd be guessing, but you would ask them to resize 40 files for a website, name them a certain way, and drop them into a specific folder. That second kind of job has a clear finish line, and Dewis sees it as the kind this technology is built for.

To see what happens with the vague version of the request, Dewis reopened the unedited raw file of a Miami Beach sunrise and gave Claude a single instruction: edit this for me. He wasn't using Adobe's integration, which hasn't shipped yet. Claude was operating his computer directly, moving the mouse and clicking through Lightroom the way a person would. It raised the exposure, pulled the highlights way down, pushed the shadows all the way to +60, warmed the white balance, and added some vibrance, texture, clarity, dehaze, and a light vignette. The lifeguard tower went from near silhouette to readable, and the sky kept its color while the sun stayed intact. Every change came from Lightroom's own sliders, so nothing was generated and all of it stayed adjustable. It left the profile, crop, and color mix alone and used no masks. When it was done, Claude noted that raising the shadows so much turns a moody sunrise into an evenly lit scene and suggested pulling them back to keep the silhouette feel. It had to assume what sort of picture he was after, and it said so. Dewis' own finished version has birds in the sky that he caught in separate frames and added in Photoshop, something no AI could have known about.

Dewis also separates two kinds of AI that tend to get lumped together. Claude worked with the same sliders you'd use, while some of Photoshop's newest tools generate new pixels, and that changes how closely you need to check their work. In a DPReview demo with Adobe's head of Photoshop product management, a Generative Restyle pass meant for the background of a wedding portrait also changed the woman's face and the forest behind her, since it regenerates the whole image. Relight works at 2K and upscales to your file size, which can cost some detail, and Adobe has been upfront about that. Both draw on generative credits, with Restyle costing 80 and Relight 30, though Relight previews are free. The MAX demos themselves were tightly defined jobs: a poster resized for several platforms, with its text translated, in Photoshop, a design split into separate colors for screen printing in Illustrator, and a long podcast cut into social clips in Premiere. None of it was instant, either. Building nine editable artboards for that poster took ChatGPT just over nine minutes.

Weddings and events are where Dewis sees the strongest case. Come home with thousands of frames and the usual move is to edit one, then sync or apply a preset to the rest, but a preset puts identical numbers on a frame shot in bright sun and one from a dim church, which leaves you correcting them individually. Tools like Imagen, Aftershoot, and Evoto already handle this differently. You feed them your own finished edits, somewhere between 200 and a few thousand images depending on the tool, and they adjust each new frame instead of stamping the same settings across a whole shoot. Aftershoot will also take one of your Lightroom presets and convert it into a style that responds to changing light. Where Dewis sees Claude being genuinely useful is that same job done inside the apps you already use: you edit one picture, show it the look, and it carries that look across the set while adapting to each frame, lifting a darker one more and pulling a warm one back.

None of this is built into Photoshop yet. Adobe's version, which it calls local MCP integrations, goes into a private beta in early 2027. MCP is a standard way for an outside AI such as Claude or ChatGPT to operate an app, here working on the files stored on your own computer, and Adobe's written list names Photoshop, Illustrator, After Effects, and Premiere while leaving Lightroom off. Adobe also hasn't specified what each model will be allowed to do, so what MAX showed is closer to an outline than a finished product, and it's something you choose to connect. Adobe CTO Ely Greenfield has said agents can't yet reliably tell whether work suits a brand campaign, which leaves people in charge of the creative direction.

On the creativity point, Dewis notes that plenty of photographers already use looks someone else built, from presets to plugins like Boris FX Optics and Topaz. Asking Claude or ChatGPT for a teal and orange grade made from Photoshop's own adjustment layers, which you can still change afterward, is a starting point you chose in the same way. He does side with the skeptics on one thing. He'd welcome a version of Photoshop with no AI in it, the same program people know without what some call the bloat, bought once with occasional updates, though Adobe hasn't announced anything like that.

Dewis ends the video by asking which repetitive task you'd hand to an AI agent first.

Read Entire Article