The Uncanny Valley of AI Architectural Rendering: Why "Too Perfect" Can Backfire

By MNKI Team · Published · Updated

The Uncanny Valley of AI Architectural Rendering: Why "Too Perfect" Can Backfire

Quick answer: In our view, trust rather than image quality is now the real issue with AI architectural renders. RIBA's survey of its members found that 74% of practices use AI in at least some projects, and 17% of respondents agree their designs are better because of it. A 2025 peer-reviewed study found participants judged real versus AI images correctly 63.7% of the time, and increasingly relied on an image looking "too perfect" rather than on specific flaws. The practical takeaway: keep control of materials, light and proportion, check the output against your source drawings, and consider telling clients when imagery is AI-generated.

Written by the MNKI team. Last verified: September 29, 2026. MNKI makes an AI rendering tool, so we have a commercial interest in this topic. Statistics come from the sources listed at the end, and where evidence is thin we say so.

What the "uncanny valley" is

In a 1970 essay, the Japanese roboticist Masahiro Mori described what became known as the uncanny valley: as something artificial becomes more human-like, our comfort with it rises, then drops sharply when it is almost, but not quite, convincing. The idea comes from robotics. Applying it to architecture is an analogy, and we are not aware of peer-reviewed research that measures an uncanny-valley effect specifically in architectural visualization. What follows is evidence about AI-image perception and professional attitudes, plus our own editorial observations, and we label which is which.

Adoption is high; belief that AI improves design is not

  • RIBA AI Report 2026 (UK). RIBA reports that nearly three-quarters of practices (74%) now use AI in at least some of their projects, up from 59% in 2025 (RIBA's press release words the measure slightly differently). RIBA's press release reports that 17% of respondents agree their designs are better because of AI, and 77% agree AI "can never replace human creativity". These are survey responses from RIBA's members, not measurements of design quality, and we have not reviewed the full report.
  • Chaos and Architizer survey (2026). In a survey of nearly 800 professionals (about half in North America) run by a rendering-software vendor (Chaos) with Architizer, 74% of respondents (a different measure from RIBA's 74% of practices) said they were likely or very likely to increase their use of AI over the next 12 months, and nearly 86% of current AI users reported measurable time savings in design and visualization tasks. The sample was self-selected and the sponsor sells visualization software, so read it as indicative rather than representative.
  • Autodesk State of Design & Make 2026. Autodesk, a software vendor, surveyed 2,500 leaders across AECO (architecture, engineering, construction and operations), design and manufacturing, and media and entertainment (so not architecture-only). Security was the top concern at 60%; accuracy came next at 49%, up from 39%, which Autodesk attributes to scepticism of output from relatively untested models.

Put together: RIBA members report widespread use, one vendor-run survey reported time savings by most current AI users, and concerns about accuracy are rising among the leaders Autodesk surveyed. Those are different surveys with different samples, so we do not combine or compare their numbers.

What research says about "too perfect" images

A 2025 study in Frontiers in Artificial Intelligence (Högemann, Betke and Thomas; details as summarised on the paper's open-access page) asked 104 German-speaking participants to judge 50 images of landscapes, architecture and interiors, 25 real and 25 generated by five text-to-image models. Overall accuracy was 63.7%. The authors report that participants increasingly relied on an image looking "too perfect" rather than on specific technical flaws.

Two caveats. First, this study is about detecting AI images, not about how clients react to architectural proposals. Second, an older study gives a different kind of evidence: a 2018 study of historic virtual architecture (150 subjects) found photorealistic renderings scored highest for visual appeal and perceived accuracy, while watercolour renderings were best for recognising architectural style. That study used historic virtual buildings rather than client proposals and did not test uncanny-valley effects, so it neither supports nor refutes the idea. The evidence that "too perfect" renders lose client trust is, at present, mostly opinion and qualitative observation.

Things to check in an AI render (an editorial checklist)

This checklist is our own editorial suggestion. It is not a measured finding, and we have not run a study:

  • Materials you cannot specify. A floor that reads as generic "wood-ish" rather than a finish someone could order.
  • Lighting that is consistent but characterless. Shadows fall correctly, but nothing suggests a time of day, orientation or climate.
  • Proportions that drift. A window or door quietly larger or smaller than your sketch implied.
  • Surfaces that are too clean. Real materials show grain, wear and irregularity.
  • Context that was never yours. Invented neighbours, planting or street furniture that do not exist on the site.

How to reduce the risk

  1. Fix the tell, not the whole image. If one material reads as generic, mask just that area and regenerate it with inpainting, instead of re-rolling the entire render. See how inpainting compares across tools.
  2. Anchor materials to something real. Name the material in your prompt, use a reference image, and generate a material palette for discussion. Always check real products with suppliers. See our mood board and palette guide.
  3. Use a stricter strategy when fidelity matters. MNKI describes its Strict strategy as deterministic, closely following your input geometry with minimal interpretation; Balanced is its recommended default; Creative is for exploration. See how MNKI's controls work.
  4. Compare against the source. Overlay or flick between the render and your sketch or plan and check openings, massing and levels. Our sketch-to-render guide has a review checklist.
  5. Keep AI for concept communication. AI renders are suited to getting a client to react to a direction. They are not dimensioned drawings or construction documents.

Disclosure and the rules that are arriving

Trust is partly a disclosure question, and guidance is still being developed (RIBA says it is drafting its own). This is general information, not legal advice.

  • AIA. The American Institute of Architects' AI Firm Toolkit recommends disclosing AI use when it played a substantive role in a deliverable the client will rely on, and gives AI-generated renderings presented as project visualizations as an example. It says a clear note in a transmittal or in the deliverable itself is often sufficient, and links this to the AIA Code of Ethics rule against misleading clients (that link is the toolkit's own reading of the rule).
  • EU AI Act. Article 50 applies from 2 August 2026. It requires providers of systems that generate synthetic images to mark outputs in a machine-readable way (with a grace period to 2 December 2026 for systems already on the market), and requires those who use such a system to produce a "deep fake" to disclose it, with a lighter disclosure duty for evidently artistic or creative works. A deep fake is defined as content resembling existing persons, objects, places, entities or events that would falsely appear authentic. The European Commission's FAQ does not say whether a visualization of a proposed, unbuilt building falls under that definition, so do not assume either way. If you work in or for the EU, take legal advice.
  • Content credentials (C2PA). The C2PA explainer (version 2.2; newer versions of the specification exist) describes a standard for recording where content came from, including AI actions. It also states that provenance alone cannot tell you whether the content is true, and credentials can be stripped, although the explainer describes "durable Content Credentials" that use watermarking and fingerprinting to allow recovery. They help, but they do not replace honest labelling.
  • RIBA. RIBA says it is developing responsible-use guidance and an "AI Overlay" for its Plan of Work; we did not find published RIBA guidance on disclosing AI visuals to clients at the time of writing.

A simple habit: label AI-generated visuals as such, state what they are meant to communicate (design intent, not construction accuracy), and keep the source drawings available.

FAQ

Why do some AI-rendered buildings look fake even when they are realistic?

Possible reasons, in our editorial view and not measured, include one inconsistent detail (a generic material, characterless light, drifting proportions) or a surface that looks unnaturally clean. In a 2025 study, participants increasingly relied on an image looking "too perfect" when judging whether it was AI-generated.

Can AI renders be used in client presentations?

Yes, for concept and design-intent communication. Consider labelling them as AI-generated, check them against your source drawings and do not present them as dimensionally accurate or construction-ready.

How do I make an AI render look more authentic?

Ideas to try (untested by us): fix specific weak areas with inpainting, name real materials and use reference images, and keep the render anchored to your own sketch or plan rather than a text description alone.

Is there research on the uncanny valley in architectural visualization?

We are not aware of peer-reviewed research measuring it specifically for architectural visualization. The closest evidence is general research on how people perceive AI images.

Do I have to disclose AI-generated renders?

It depends on your jurisdiction, contracts and professional rules. The AIA toolkit recommends disclosure for AI-generated renderings presented as project visualizations, and Article 50 of the EU AI Act applies from 2 August 2026; whether it covers a given architectural render depends on its definition of a deep fake, which the European Commission's FAQ leaves open. Take legal advice if you are unsure.

Sources