As AI rendering has become standard practice across architecture, a new skill has quietly developed among design-literate clients: spotting the tells. The principals of larger developers, the more engaged residential clients, the competition jurors who receive 200 AI renders in a single round — they're learning what to look for. Understanding what they're looking for is now part of the professional render workflow.
This is actually useful information. Not as something to evade, but as a quality checklist.
The common tells of a low-quality AI render
Furniture from the wrong era or style. The AI has a strong learned prior for what "a nice living room" looks like — and it's often a contemporary European interior from approximately 2019. If your design calls for something specific, a hallucinating AI will override it with its trained aesthetic default. A Scandinavian sofa in a render for a brutalist apartment is a tell.
Walls that bow or have imprecise angles. AI-generated spaces often have a very slight curvature to surfaces that should be planar — walls that are 99% flat rather than exactly flat. This is subtle but immediately identifiable to anyone with spatial training. It reads as "something is wrong" before the viewer can articulate what.
Outlets, switches, and fixtures in wrong positions. The AI places these according to its training data rather than your design. A light switch that's centred between a door and a wall junction, rather than at the logical mounting height and position, is a consistent tell. The model knows these objects exist but not where they belong in your specific design.
The wrong window behaviour. Windows are among the most scrutinised elements in any architectural render. If the reflections in a window don't correspond to the modelled exterior, or the view through the window shows a landscape inconsistent with the building's context, clients notice. The reflections in particular tend to be generic rather than specific.
Textures that look painted-on. Well-calibrated material textures have depth, variation, and scale-appropriateness. AI-generated textures sometimes have a slightly flat quality — they look like a photograph of material rather than material. Stone in particular can look like a photographic print of stone rather than actual stone in a space.
Overlit highlights. A very specific tell: blown-out highlights on surfaces near windows that have a synthetic quality. Traditional rendering handles this through physically accurate exposure. AI rendering approximates it and sometimes produces highlights that feel slightly too intense or too uniformly smooth.
The "AI smooth" quality. Hard to define precisely, easy to recognise once you've seen it: a certain softness to edges, a slight over-smoothness to surfaces, an absence of the micro-variation that real materials have. It's getting much less common as models improve, but it's still present in lower-quality outputs.
Why geometry drift is the most revealing tell
Of all the identifies above, geometry drift — walls moved, proportions changed, spaces that don't match the floor plan — is the most damaging professionally. It's also the most recognisable to the clients most likely to scrutinise your work.
A principal who has reviewed the floor plan and elevations will immediately notice if the render shows a room that's 3.8m wide when the drawing says 4.5m. A structural column that doesn't appear in the render when it appears in the documentation creates confusion about what's actually being built.
This isn't just an aesthetic problem. A render that doesn't match the design is a documentation problem. It creates expectations that don't align with the actual project, which surfaces as a difficult client conversation later.
The tells of a high-quality AI render
Understanding what makes an AI render fail helps define what makes one succeed. The best AI renders have:
Correct geometry that matches the model. The most important thing. If a client who has seen your drawings looks at the render and the spatial relationships match exactly — ceiling height, room width, window positions — the render is doing its primary job.
Named, specific materials. "Calacatta Viola marble slab, polished" looks different from "white stone countertop." Materials named specifically — by species for timber, by quarry name for stone, by manufacturer for tile — pull from a more precise region of the AI's training data and produce more recognisable results.
Lighting that behaves correctly. Shadows that fall where physics dictates, reflections that correspond to the modelled geometry, light that appears to come from a single coherent source. The AI doesn't have to simulate physics exactly — it just has to produce an image that appears to follow physical rules.
Correct furniture scale. This is a surprisingly strong tell when it goes wrong. A dining chair that looks like it seats someone 2.2m tall, or a sofa that would occupy half the floor area of the room as modelled, breaks the spatial credibility of the render immediately.
No anachronistic objects. The AI's training data has a temporal distribution. Objects from the current material culture — contemporary furniture, current technology, accurate signage — produce more credible renders than generic mid-period objects.
What this means for your workflow
The practical implications are direct: use a geometry-fidelity tool; specify materials by name rather than description; use lighting presets rather than vague prompt language; have a second set of eyes review the render before it goes to the client.
The geometry-fidelity point is the most important. Tools that are engineered to preserve your input geometry — Maquete's architecture-specific prompting is designed specifically for this — eliminate the most damaging category of tells before the render is generated. The tells that remain are about material quality and lighting accuracy, which can be refined through iteration.
The useful inversion
There's something clarifying about clients learning to spot AI renders. The things that reveal a render as AI — geometry drift, wrong material specificity, incorrect lighting behaviour, anachronistic objects — are exactly the things that made bad traditional renders bad, too.
The quality standard hasn't changed. What's changed is the path to meeting it. A good AI render is a render that accurately represents the design, uses correct spatial relationships, specifies real materials specifically, and applies coherent lighting. A bad one doesn't.
This was always the standard. It just used to be easier to fail quietly.
How can you tell if a render is AI generated? The common tells are: furniture that doesn't match the design intent, walls or proportions that don't match the documentation, outlets and fixtures in implausible positions, textures with a flat "printed" quality, and a slight over-smoothness to surfaces. High-quality AI renders from architecture-specific tools eliminate most of these.
Why do AI renders sometimes look fake? Most AI rendering tools use general diffusion models that prioritise visual quality over geometric accuracy. The model adds furniture, adjusts proportions, and applies its own aesthetic defaults when the input doesn't match its training data expectations. Architecture-specific tools with high conditioning strength and geometry-fidelity prompting prevent this.
Does geometry preservation matter for client renders? Yes, significantly. Clients who have reviewed floor plans and elevations will notice if the render shows different proportions or positions. Beyond aesthetics, a render that doesn't match the design creates documentation problems — it sets expectations that don't align with the actual project.
How do I make AI renders look more realistic? The highest-impact changes: use a geometry-fidelity tool that preserves your actual walls and proportions; specify materials by name rather than generic description; use a specific named lighting preset rather than a vague mood prompt; render at 4K and avoid obvious post-processing. The realism comes from accuracy, not from visual effects.