All articles
Perspectives

AI in Lighting: What It Does Well, and What the Evidence Does Not Yet Support

Archlior
AI in Lighting: What It Does Well, and What the Evidence Does Not Yet Support

Lighting has used "smart" features for decades. A sensor turns lights off in an empty room. A photocell dims a row of fittings near the window. What is new is the claim that machine learning can do these jobs better, and that generative image models can show you a lit room before it exists.

Some of that is real. Some of it has not been tested in the way a specifier would want. Here is where the published evidence stands.

Start with the baseline: what plain controls already deliver

Before asking what AI adds, it helps to know what ordinary sensor-based controls achieve. A meta-analysis from Lawrence Berkeley National Laboratory pooled 240 energy-saving estimates from 88 papers and case studies. Its best estimates of average savings were:

  • 24% for occupancy sensing
  • 28% for daylight-linked dimming
  • 31% for personal tuning (users setting their own levels)
  • 36% for institutional tuning (a facilities team setting maximum levels)
  • 38% for combinations of strategies

Two details matter. The authors found that simulations overestimate daylighting savings by at least 10% compared with measured buildings. And the study dates from 2011, so it describes conventional controls, not machine learning. Savings also depend heavily on building type and baseline, so these averages are context, not a target for any product. They do show why the comparison baseline matters: an AI controller should be compared with a well-commissioned conventional system, not with lights left on all day.

Where machine learning plausibly helps: controls that learn from people

The research idea is occupant-centric control: a system that learns when and where people are, and what light they actually choose, then adjusts to match. A 2024 review in Building Simulation (Yuan and colleagues) covers this field. Its conclusion is measured. Despite extensive research on occupant behaviour, the authors say there are still significant challenges in using it for control, including the lack of a bridge between the information that monitoring collects and the controls that act on it. They also stress the importance of testing through experiment. These are the authors' conclusions about the literature they reviewed, up to 2024.

Reinforcement learning, where a controller improves by trial and error, was reviewed in Energy and Built Environment (Han and colleagues, 2021). The authors write that the examination of occupant behaviour using reinforcement learning is "not well established" and that how occupants interact with the learning agent is "still unclear". That is a 2021 assessment of one research area, and it is not a verdict on every controller sold today. We have not reviewed the performance of any specific 2026 product.

What it does suggest is that any performance claim, such as "our system saves 40%", should be checked like any other: measured in a real building, against a defined baseline, over a full year.

Generated images: convincing light is not measured light

Generative image tools can produce a convincing render of a lit room in seconds. The question for a lighting designer is whether the light in that image is physically right: does the shadow fall where the luminaire would put it, and is the brightness plausible?

A benchmark by Giroux and colleagues ("Shedding Light") tested 16 image generation models. It was posted as a preprint on arXiv in September 2026 and the authors state it has been accepted for publication in ACM Transactions on Graphics (SIGGRAPH Asia 2026 issue, December 2026), so it is a preprint of forthcoming work, not yet in print. The researchers asked each model to insert a plain object into a real photograph, then measured the direction, colour and distribution of the light on that object against a known reference. Findings include:

  • The best models had a median error in light direction of about 11 degrees (FLUX.1 Fill at 11.38 and Z-Image Turbo at 11.92). The weakest of the models singled out were roughly 25 to 35 degrees (HiDream 24.77, Hunyuan 28.17, SD 3.5 34.84).
  • Accuracy had not improved over time, even though newer models are larger and trained on more data.
  • Light arriving from the top and back was handled worst, with a median error near 42 degrees.
  • Models tended to predict light as coming from more frontal and lower directions than the truth, and did not respond strongly to the stated surface brightness of the inserted object, which the authors describe as struggling with the ambiguity between surface colour and lighting.

The limits of this study are worth stating. It tests indoor scenes with a single point light source, and it measures light direction and distribution, not lux values on a work plane. It covers the 16 models tested, not every tool or future model, and it does not test architectural rendering products as such. What it shows is that, for these models and this test, a good-looking image did not guarantee physically accurate lighting. Our reading is that this makes generated images a poor basis for decisions such as beam angles or glare checks, where a photometric calculation is the safer source.

Where simulation stands, for comparison

Physically based simulation has its own error margins, and knowing them gives a useful benchmark. Researchers at Pacific Northwest National Laboratory (Safranek and Davis, 2022) compared a Radiance-based tool with measurements in a controlled room lit by LEDs. Mean absolute errors were 1.4% for horizontal and 3.8% for vertical illuminance, which they considered reasonable for architectural simulation. Predictions of spectral quantities, which matter for circadian metrics, showed larger errors, and the authors described those as concerning and in need of further study.

That is one tool tested in one laboratory room, so it is an example, not a general guarantee. The point is that validated physical simulation can be checked against measurements and its error stated, which is what a generated image has not yet been shown to offer.

A practical checklist for specifiers

When a product or tool claims to use AI in lighting:

  1. Ask what it measures and what it controls. A sensor-driven dimming profile is different from a generated picture.
  2. Ask for results measured in a real building, with a stated baseline, not a simulation.
  3. Ask what the saving is measured against. A claim should name the baseline, since the conventional-controls averages above vary widely by building and strategy.
  4. Keep the photometric file (IES or LDT) and a standard calculation as the source of truth for levels and glare.
  5. Check what happens when the system fails. A learned controller still needs a sensible default and a manual override.

What we do not know

Within the sources reviewed here, we did not find evidence that learning controllers outperform well-commissioned conventional controls across everyday buildings, and the reviews describe the research as still developing. Newer work may exist that we did not review. Likewise, within the benchmark's scope, generated images were not shown to be photometrically reliable. The summary is that AI has narrow, measurable uses in lighting, and that broader claims should be tested against real-building data.

Sources