There’s something deeply human about watching a tech company stumble in public. Sony launched its AI Camera Assistant on the Xperia 1 XIII with what seemed like a straightforward promise: let machine learning help you take better photos. The idea makes sense in theory. Artificial intelligence has transformed image generation and editing. Why not photography suggestions?
Then the examples started rolling in, and it became clear that something went very wrong.
The backlash was swift enough that Sony felt compelled to clarify what the feature actually does. The company insists it doesn’t edit your photos, merely offers suggestions based on lighting, depth, and subject matter. Point your camera at something and the AI will propose four different adjustments: exposure, color, and background blur variations. Sounds reasonable. The product video even mentions it suggests “the most photogenic angle,” though what they actually showed was just a zoom recommendation, which is… not the same thing at all.
I found the disconnect telling.
The Examples That Broke Trust
When I looked at Sony’s initial examples on May 14th, I understood why people reacted poorly. The suggestions were uniformly worse than the originals. A sandwich looked washed out. A portrait in a meadow was so over-exposed it looked damaged. These weren’t minor tweaks that happened to miss the mark. They were objectively degraded versions of already decent photos.
Sony tried again with a new set of examples, likely hoping to rebuild confidence. I’m not sure it worked.
The first suggestion cranked saturation to cartoonish levels. The second flattened everything into a processed haze that removed any sense of dimension. The third made food look like it had been pasted into the scene with Photoshop. The fourth pushed contrast so aggressively that you wondered if the phone’s display settings were miscalibrated. Each one had serious problems. Each one looked worse than the original.
This matters because it reveals something fundamental about how we’re currently deploying computer vision in consumer products. We’ve gotten obsessed with what the models can do in isolation, but we haven’t always thought deeply about what users actually need.
What Went Wrong Under the Hood
I think the issue starts with training data and loss functions. Sony’s AI was likely trained on thousands of photos labeled as “improved” by human raters. But how do you define “photogenic”? What’s considered a better photo varies wildly based on context, personal taste, and artistic intent.
The suggestions we’re seeing look like the model learned to maximize for certain measurable features: saturation, contrast, exposure evenness. These are easy to quantify. They’re easy to train on. But they’re not what makes a photo actually good. A photograph is good when it captures intention, mood, and authenticity. Those are fuzzy concepts that don’t compress well into loss functions.
There’s also the question of what baseline the AI is being compared against. If it’s being trained to move photos toward some statistical average of “good photos on Instagram,” you’re probably going to get aggressively processed results. That works for some genres. For everyday photos? It’s a recipe for over-processing everything.
I suspect the engineers at Sony knew this might be an issue but pushed the feature to production anyway. Maybe it was a deadline. Maybe it was pressure to ship something with AI in the marketing materials. The cost of failure felt abstract enough at the time.
The Practical Reality for Users
If you own an Xperia 1 XIII, the honest recommendation is simple: ignore these suggestions and take your own photos. The AI isn’t ready for this task, at least not in its current form.
But here’s what’s interesting to me as someone watching the broader tech landscape: this failure is actually instructive. It shows us that slapping a model onto a product and calling it intelligent doesn’t automatically make it useful. There’s still an enormous gap between “the AI can make suggestions” and “the suggestions are actually worth following.”
We’re seeing similar patterns across other AI-assisted features in consumer devices. Recommendation systems that recommend the wrong thing. Auto-complete that suggests words you’d never use. AI features that technically work but fundamentally misunderstand what the user is trying to accomplish.
The problem isn’t that the technology is immature. It’s that we’re not spending enough time on the human factors. What do photographers actually want? What frustrations are they experiencing? What would genuinely make their workflow better? Those questions need to drive development, not the other way around.
Sony had an opportunity here to ship something genuinely helpful. Instead, it shipped something that made photos worse. That’s not a tragedy. It’s a learning moment. The question is whether the industry will actually learn from it or just move on to the next AI feature to hastily integrate into consumer products.