I Don't Know Anything About Decorating. ChatGPT Doesn't Either. We Helped Anyway.

A friend who doesn't use AI tools asked me for decorating advice. They're furnishing a lake house living room and wanted to know what accent chair colors would work with what they already had.

I don't have an interior design background. But I do have ChatGPT. So I told them I'd see what the tools could come up with.

This is actually one of my favorite ways to test AI. Not on abstract prompts I've crafted for maximum performance, but on real questions from real people who have no idea what a large language model is and just want a useful answer.

So I uploaded a photo of the room and asked: "If I were to add two accent chairs to complement these couches, what color options should I consider?"

Before I get to the response, let's pause on that sentence for a second.

I uploaded a photo of the room.

My friend didn't upload it. I did, on their behalf. Before I did, I asked for their permission and scanned the image for anything personally identifiable. Nothing in the frame gave away a name, an address, or a face. I made a deliberate call that the photo was safe to share in this context.

But most people helping a friend this way wouldn't think to do any of that. And the friend on the receiving end almost certainly wouldn't think to ask. When you use AI on someone else's behalf, the privacy decisions you make aren't just yours, and the gap between how carefully you handled it and how carefully the average person would handle it is worth naming, I'm still trying to come up with a good term for it.

Now, back to the experiment.

The response was genuinely good. ChatGPT identified the room's palette from the photo. Soft gray upholstery, warm wood floors and beams, black hardware, sage and blue pillow accents. It gave five color directions with reasoning for each, including: deep navy, soft olive, camel/cognac leather, creamy neutrals, and charcoal.

And then it did something that felt helpful but is worth examining. It gave me a "designer pick."

Top two options: moss green fabric or camel/cognac leather. "If you want cozy lake house, go green. If you want elevated modern retreat, go cognac leather."

Clean. Decisive. Easy to forward to a friend who just wants to know what to buy.

But ChatGPT has never been to this lake house. It doesn't know my friend's budget, whether they have kids or pets, how much direct sunlight hits that spot in the afternoon, or what else is already in their shopping cart. It rendered a confident purchase recommendation from a single photograph and a one-sentence prompt.

That's not a reason to distrust the answer. It's a reason to remember that AI tools are optimized to be helpful and conclusive, and "conclusive" can sometimes look a lot like "go buy this specific thing." A tool that ends every response with "here's exactly what I'd choose" is a tool that's very easy to follow without questioning.

I kept going. I told ChatGPT that the chairs would face the fireplace, with their backs to the window.

That detail matters in interior design. Chairs facing inward become visible from the entryway. The back of the chair is what people see first. It changes how you think about silhouette, shape, and color.

ChatGPT acknowledged this. It said "that placement actually changes the design decision in a useful way," narrowed its list to three options, and added a note about choosing chairs with visually interesting backs like barrel shapes, wood-framed styles, leather sling designs.

Except when I looked at the revised list, nothing had actually been removed from the original. Cognac leather, moss green, and slate blue were all already there. The update changed the confidence and framing of the recommendation. The options themselves didn't change.

That's worth paying attention to, especially when you're using AI to help someone who is trusting you to filter what's actually useful from what just sounds useful.

What Worked

• Accurate visual analysis from a single photo
• Well-organized reasoning for each color option
• Conversational, confident responses that felt like talking to someone who knew what they were doing
• Explicitly acknowledged new context rather than ignoring it

What to Watch For

• Uploading someone else's home photo requires their permission and adding any photo should always come with a scan for personally identifiable information. Many people will skip both steps
• Confident "designer pick" framing can nudge toward a specific purchase without full information about budget, lifestyle, or practical constraints
• "That changes things" doesn't always mean things actually changed
• When you're passing AI advice to someone who isn't evaluating it themselves, the responsibility to notice falls entirely on you

The Takeaway

Using AI to help non-technical people with real problems is one of the most valuable things these tools can do and one of the higher-stakes ways to use them. My friend isn't going to notice that the "updated" recommendation was already in the original list, or think twice about the fact that a photo of their home now lives in an AI company's systems. That's my job.

The advice was genuinely useful. I still passed most of it along and they now have beautiful, camel-colored leather chairs that they love to complement the room. But I had to audit the privacy implications, fact-check the "updated" recommendation against the original, and filter out the confident purchase nudge before any of it was ready to hand to someone who was simply trusting me for help.

If that much interpretation is required before an AI output is safe to pass to another person, that's not the user's job to manage. That's a gap the tool hasn't closed yet.