A client arrives with an AI-generated interior on their phone and asks whether the space can look like that. It is a fair question, and the honest answer starts somewhere unexpected: that image contains no light. It contains a picture of light — a distribution of bright and dark pixels that a model has learned to produce because it resembles the millions of photographs it was trained on.
Nothing in it carries an illuminance value, a beam angle, a glare rating or a maintenance factor. On a commercial project those are the quantities that determine whether the scheme certifies, whether the space is comfortable to work in a year later, and whether the energy report passes. Here is what the current generation of image models genuinely does well, where it breaks down, and why that boundary is structural rather than a matter of waiting for a better model.
What AI gets right
The strength is atmosphere. Generative models have absorbed an enormous quantity of well-photographed architecture, and they reproduce the feel of a lit space with real fidelity — the warmth of a 2700 K restaurant at night, the way a wall wash flattens or reveals a stone texture, the mood difference between a room lit from the ceiling and a room lit from its perimeter.
That is not a small thing. For years the hardest part of a concept conversation was that the client could not see the night-time version of a space until it was built. A rendered sketch closes that gap in minutes rather than days, and it closes it in a language the client already speaks. Used at concept stage, AI imagery is a communication tool of genuine value: it makes an abstract lighting intent arguable. The client can point at something and say not that, which is far more useful than a nod at a written description.
It is also useful for exploring options quickly. Producing six variations of a facade treatment to see which direction is worth developing is exactly the kind of low-stakes, high-volume work that suits a generative tool.
Where it stops
The failures are consistent, and they are the ones a lighting designer notices in the first few seconds.
Illuminance has no units. A model can make a room look bright. It cannot tell you the space sits at 320 lux maintained for typical screen-based office tasks under AS/NZS 1680.2.2. There is no calculation underneath the image — no room geometry, no surface reflectances, no utilisation factor, no light loss factor. The apparent brightness is a stylistic choice the model made, not a result.
Luminaire output and the light patch do not agree. Point a 26° spot at a wall from four metres and the pool of light has a predictable size and edge. In generated images the fitting and its beam routinely contradict each other — a narrow spot casting a wide diffuse wash, a downlight producing a patch that could only come from a much wider distribution. The model has learned that fittings and bright areas appear together; it has not learned the photometry that connects them. Holding that geometry steady is a discipline of its own.
Shadows are not self-consistent. Multiple sources in one room produce shadows that should agree on where the light is coming from. Generated scenes commonly show shadows falling in incompatible directions, or an object lit from a source that does not exist in frame. Once noticed, it is difficult to stop noticing.
Glare is invisible. Discomfort glare is a calculated quantity — discomfort-glare assessment, such as the glare evaluation method described in AS/NZS 1680, checked against the limit for the task. An image cannot express it. A generated office can look serene while carrying a luminaire layout that would be genuinely unpleasant to sit under, because the model has no representation of a viewer position or a field of view.
Maintenance does not exist. The maintained illuminance in AS/NZS 1680 is the level held at the end of the maintenance cycle, after lumen depreciation and dirt accumulation. AI imagery is always day one, always clean, always full output. The design question — what will this look like in three years — has no counterpart in the model.
Why the gap is structural
It is tempting to read these as bugs that a larger model will fix. The gap runs deeper than that.
A lighting calculation engine solves a physical problem. Light leaves a source with a measured intensity distribution, travels, reflects off surfaces with known reflectances, and arrives at a plane where illuminance is computed. Change the ceiling height and every downstream number changes, because the numbers were derived rather than asserted.
A generative image model does something else entirely: it produces a plausible array of pixels conditioned on a description. It has no source, no surface, no plane. Ask it for the lux level on a desk and the question has nowhere to land: the generated image does not contain the physical and photometric data required to answer that question. Improving the model makes the picture more convincing. It does not add a quantity that was never in the pipeline.
This matters practically because a more convincing wrong answer is more dangerous than an obviously wrong one. An image that looks right invites everyone to stop checking.
What a trained eye checks first
When assessing whether AI truly understands light, a trained lighting designer first checks whether objects and luminaires are positioned logically and follow the correct perspective, for example, ensuring a ceiling light is not mistakenly placed on a wall. They also examine whether light beams have a realistic shape, direction, spread and intensity; whether each type of lighting is appropriate for the space and its intended function; and whether the scale and proportions of the objects and luminaires are believable. These details reveal whether an AI-generated image demonstrates a genuine understanding of lighting or merely looks convincing at first glance.
Where AI belongs in the workflow
The line sits at the point where an output becomes a number that someone relies on.
Concept, mood, client conversation, rapid option generation — AI earns its place, and the working method for getting good results out of it is a discipline in itself. Everything downstream of that runs on calculation: illuminance targets to relevant Australian Standards (e.g. AS/NZS 1680 for interior lighting), cable sizing to AS/NZS 3008, voltage drop and protection to AS/NZS 3000, lighting power density against Section J7 of the NCC. Those figures go into documentation, get checked by a certifier, and carry professional liability. They come from AGI32 or DIALux EVO, and a human signs them.
Used that way the technology is straightforwardly useful. The failure mode is treating a concept image as a specification — approving a scheme from a picture, then discovering during documentation that the fitting count, the energy budget and the glare performance were never compatible with what everyone agreed to.
FAQ
Can AI design lighting for a commercial project?
It can generate concept imagery that communicates a lighting intent, which is a real contribution at the front of a project. It cannot produce the illuminance calculations, luminaire schedules, circuit design or compliance documentation that a commercial project requires. Those are calculated in photometric software against the relevant Australian Standards (e.g. AS/NZS 1680), AS/NZS 3000 and the NCC, and signed by a person.
Are AI renders accurate enough to approve a design from?
No. A generated image carries no illuminance values, no beam angles and no glare rating, so there is nothing in it to approve against. Treat it as a conversation about intent, then have the intent verified by calculation before it becomes a decision.
Will AI replace lighting designers?
It is already accelerating part of the work — early concept visualisation that used to take days. The calculation, the standards compliance, the coordination with the electrical design and the judgement about what a space actually needs are not tasks a generative model performs. The role shifts toward the parts that require accountability.
What software actually calculates lighting?
Photometric calculation software such as AGI32 and DIALux EVO are widely used in Australian practice, with ElumTools used where the calculation sits inside a Revit model. They work from luminaire photometric files, room geometry and surface reflectances, and produce illuminance results that can be checked against relevant Australian Standards like AS/NZS 1680. Where each tool belongs across the whole workflow is set out in a separate guide.
SL Design Consultant is an independent lighting and electrical design consultancy in Sydney, delivering lighting and electrical design across commercial, hospitality, retail, heritage and public projects. Our product-neutral recommendations are guided by design intent, technical performance and project requirements.