Light, Logic and Learning

It was November 1995. A television production for a continental audience, the Artisan™ lighting console by Vari*Lite, operating on the “What You See Is What You Get” principle. I was in the thick of it all, despite having limited ability to read and write musical notation.

Out of that situation emerged, as a pragmatic technical workaround, a textual and logically structured sequence of instructions, a bespoke command sequence that functioned without notation and yet allowed me to execute a complex lighting production with precision and reproducibility. No formula. No standard method. But an idea that today forms the beating heart of modern AI systems.

This staging score arose out of necessity: to establish scenic reference points and textual descriptions. Every lighting moment was connected to an action-oriented description: “When he raises his arms, breathe in, then emphasise.” This score was not improvisation, but a structured document that captured transitions, nuances and dramaturgy in a form I could execute. It bridged artistic intention and technical realisation. And it worked.

It defined states and transitions in a language that was close enough to the reality of the system to be executed, whilst remaining sufficiently removed from pure technology to lose nothing in artistic quality. Looking back, it seems anything but coincidental that precisely this balance between structured input and interpretable output forms the fundamental principle of today’s great language models.

What I sensed at the time, without being able to name it: that in the moment when light ceases to be a mere technical reaction and becomes a structured quantity within an interpretable system, something fundamental occurs. During my work on Michael Jackson’s appearance on Wetten, dass..?¹, this feeling became a certainty that has never left me since. The conversion of human ideas into structured data that a technical system can interpret. That is not a technical detail. That is the fundamental principle.

Some moments in our lives only reveal themselves as turning points years later. This one took nearly thirty years to do so.

Since then, the following question has presented itself: what will happen when machines no longer content themselves with executing visual tasks, but become capable of understanding them? This distinction seems important to me, since the two are often conflated, which prevents both realists and visionaries from conducting a meaningful discussion.

The question is not whether machines will one day design better than humans. That question is idle and misses the point entirely. The real question is what happens when a system that knows neither fear of the dark, nor fatigue after a late-night rehearsal, nor disputes with the director, takes on precisely those aspects that have hitherto been the exclusive preserve of human intuition. What then remains that only a human can do? And what emerges anew when the machine takes over the work? These questions cannot be answered from behind a desk. They require practice, experimentation, failure and renewed attempts.

What Already Works and What Does Not

What works today: AI-supported tools for lighting planning, enabling lighting visualisations to be created and shared more rapidly. Tools such as Midjourney² or similar image generators are increasingly used by lighting designers to create mood images, not as a final product, but as a means of communication with directors, set designers and producers. This is no futuristic scenario. It is established practice, including in my own work. And as ever: in marketing materials, the technical promises of certain AI tools regularly outpace what can actually be verified under production conditions. What functions in demonstration environments does not necessarily hold up under real-world conditions.

What has crystallised out of this as an independent methodological step, I refer to as the inverted “design workflow”. In the design process, the path runs from concept to technical realisation: idea, sketch, visualisation, programming, production and pictorial documentation. The inverted design workflow reverses this sequence. First, the visual document is produced, generated by an AI and refined through iteration until it corresponds to the artistic intention. Technical parameters are then derived from this image: colours, luminance distributions, beam characteristics, positioning. The result comes first; the path to technical realisation is traced backwards. This is no mere diversion. It is a new approach that offers a concrete and accessible point of entry, particularly for lighting designers who have hitherto had little experience with AI tools.

There are also solutions where AI increasingly automates time-consuming tasks, such as the management of luminaire libraries (fixture databases), 3D modelling, or suggestions for patch configurations. Initial attempts at the automatic generation of cues based on musical and/or textual analysis are already observable. Thanks to its command-line interface, the grandMA3 lighting console series offers an exceptionally user-friendly structure for automated input: almost all functions can be controlled via text commands. For an AI system, this is no minor detail. It is direct access to the system’s logic, and it perceptibly reduces the complexity of interaction between the operator and the lighting desk.

The implications of this accessibility extend well beyond the control layer. Thanks to the MCP server (Model Context Protocol server) and its associated interfaces, it is possible to establish a data pipeline connecting unstructured text with the lighting control system: textual descriptions of fixtures are converted into structured tables, from which scripts are generated and imported directly into the lighting console, including patch data, entries in the fixture library, and group and preset structures. The MCP server acts as a standardised intermediary layer that enables both sides, the language model and the control system, to communicate in a shared syntax. What currently demands hours of manual work thus becomes an automated task: structured production data derived from a natural-language input³.

Today, the transfer of expertise between operators occurs primarily through informal channels: verbally, via recordings of performances, or through the exchange of lighting console files (show files) without standardised annotation. An AI-supported knowledge database that would allow this expertise to be captured in a structured form would represent a significant advance, irrespective of whether the system develops generative capabilities or not.

In this context, the TILL Index⁴, Theoretical Indexing of Lighting Levels, holds particular significance: it is a method for evaluating and optimising mobile lighting systems, seeking to consolidate parameters such as location, lighting task and audience, together with a weighting factor determined by the nature of the event, into a standardised index. The method could serve as a structural foundation for such a knowledge base. The ZentrumNeueLichtkultur in Frankfurt am Main addresses related questions in the context of systematising lighting quality and could, in the longer term, serve as an institutional reference for standardised parameters. Structurally, however, a fundamental problem is apparent: the events sector still lacks a reliable and publicly accessible database upon which serious AI development could be built. There is no standardised collection of annotated performance files. There is no open library of recorded lighting sequences furnished with descriptions of their dramaturgy. The Faculty of Electrical Engineering, Media and Information Technology at HAW Hamburg conducted an analysis of MA2 show files to identify structural patterns in real production files, a methodological approach that demonstrated how heterogeneous data that is ostensibly standardised actually proves to be in practice⁵.

The problem of missing data standards becomes particularly apparent with regard to photometric properties. When one wishes to carry out precise photometric calculations, one frequently encounters a significant obstacle: values in operating manuals are not only non-harmonised, but are sometimes incorrectly labelled in technical data sheets. The calculation of the radiometric unit “radiant intensity” (W/sr) for a Robe MegaPointe fixture was only possible because I was able to obtain the necessary raw data through my studies at Technische Universität Ilmenau⁶. This is an area in which artificial intelligence may be able to bring greater transparency.

The ‘LEaT con 25’ trade show in Hamburg gave me the first opportunity to present my TSCHEMBOT⁷ project (Trigger Smart Cue Handling Execution Manager) to the public, an AI tool for the MagicQ lighting control system by Chamsys, enabling the creation of lighting sequences within the lighting console using natural language. In the course of the discussion, something became evident: many of those present were already using AI tools themselves. The lighting designer using Midjourney for concept sketches. The programmer asking the language model Claude to suggest cue structures. The chief lighting technician querying ChatGPT before going to rehearsal. This happens. In live productions, in renowned venues, on international tours. But no one speaks of it openly. Why not?

Because this profession has built its identity upon craftsmanship, intuition and experience accumulated on stage, for decades. Admitting to the use of AI would mean undermining that self-image: that of a lighting designer as an artist who senses what the light should accomplish, and who consults a language model on the side. The two are irreconcilable. So we stay silent. This is not a new phenomenon. It is the characteristic pattern of every industry undergoing transformation that has not yet integrated that change into its identity. The real task, as LEaT con 2025 made clear to me, is not implementation. It is about breaking taboos. Not about “getting started”, but about stopping the concealment.

What still remains to be developed is a fully autonomous system capable of conceiving a performance from A to Z, from dramaturgy to the marker list. Conceptually plausible. But not yet production-ready. The input data would be the command syntax of the system: groups, presets, cues, synchronisation, and a defined objective, a particular lighting effect or dramaturgical function. The output data would not be a list of commands, but a learnt decision logic: the result of an iterative process in which the system independently determines which inputs produce which effects. An AI model that, through reinforcement learning, not merely mimics existing design principles but independently determines how light must be used to achieve a particular effect. The rules would then no longer be inputs, but results of the learning process. That step is yet to come.

Rotterdam and the Question That Opens the Room

I presented this concept in spring 2026 at the CUE trade show in Rotterdam, in a workshop held under the title “AI-Powered Lighting Control: TSCHEMBOTi”⁸.

The responses were varied. On one side, there were those who immediately understood what I was speaking about, precisely those who are already experimenting with AI tools and recognise their potential. On the other, there was an evident scepticism that I respect and have no wish to dismiss: the question of whether a system that feels nothing, experiences nothing and has never been immersed in the darkness of a rehearsal is truly capable of giving light meaning. This question is not technical in nature. It is artistic. And it does not deserve a hasty answer. “So, can AI create an atmosphere?” Responses ranged from a categorical “No” to a cautious “Yes”, with a wide grey area in between. My own answer: as someone who creates with light, I regard the term “atmosphere” as extraordinarily elastic. In my work as a lighting designer, I have always been able to rely on two means that appear deceptively simple, yet rank among the most effective in the entire repertoire precisely because of that simplicity: the blackout and the mirror ball. Not in spite of their simplicity, but because of it.

For the blackout and the mirror ball are not tricks. They are coded reference points whose effect does not derive from their technical execution, but from the collective memory of those present in the room. Everyone knows what a blackout means. Everyone knows the mirror ball. Their effect is infallible, because they require no explanation. The atmosphere is not created by the technology: it emerges from the tension between what the light does and what the people bring with them. An AI tasked with creating an atmosphere must understand this context, not merely calculate it. The lighting designer decides what significance the mirror ball should hold in precisely this moment.

It can be deployed as irony, nostalgia, liberation or commentary. The audience perceives the effect but perceiving is not the same as shaping. It is the capacity to create this difference that makes lighting design an artistic endeavour. It is not a matter of knowing the effect produced by the means employed. It is a matter of knowing the effect desired in this precisely defined moment and why. This empirical knowledge cannot be documented; not because it is kept secret, but because it resists written capture. It is physical, contextual and cumulative. That is not a compelling argument against AI. It is a real boundary that must be named not to impede progress, but to understand what a system is capable of, and what it is not.

The Role of the Lighting Designer in Transition

What does all this mean for our role? This question is not rhetorical. It lies at the centre of every conversation I have had in recent months at trade shows, in workshops, in exchanges with colleagues. And it cannot be answered in a single sentence. For decades, this profession has been defined by the combination of three competencies: aesthetic sensibility, technical expertise, and the ability to align these two in a production context. What AI systems are changing is the weighting of these three components. Technical expertise does not become redundant, but certain tasks can be delegated. Patch structures, fixture libraries, initial cue drafts: what once demanded hours of manual work is now a matter of formulation, not merely of execution. This does not mean the lighting designer will have less work. It means the nature of that work is changing.

The executor is replaced, and I use this term deliberately, by the curator. The one who describes, selects, discards, foregrounds. The one who decides what the system should show and what it should not. The one who establishes the connection between artistic intention and the syntax of instructions. This role is not new. It has always been part of the profession. But it is now moving to the centre. And it demands competencies that are not systematically conveyed in the traditional training curriculum: the precise linguistic description of a decisive moment, structural thinking in formulating inputs, an understanding of the logic of a model whose outputs one evaluates without knowing its internal workings. At the same time, the relationship between the individual artist and the ensemble is shifting.

Those who have hitherto conceived their work as soloists will henceforth collaborate with a non-human partner who brings no shared context: no joint rehearsals, no exchanges with the director, no memory of the last tour. One must explain to this partner everything that one would normally take for granted. What helps in this case is not mastery of the technology, but an understanding of one’s own work and the capacity to describe it. To this is added a development that is rarely spoken about: the shift from keeper of secrets to translator. The empirical knowledge described in the preceding section, that which resists documentation, was for decades a source of professional authority. Those who possessed it were taken into account; those who did not were excluded. AI systems do not alter this equation by replacing empirical knowledge, they are not capable of that, but by creating a new and equally valuable competency: the ability to articulate that knowledge in such a way that a system can act upon it. The art of description complements intuition. It does not replace it.

And finally: what remains distinctly human? Not execution. Not routine. Not repetition. But the decision as to what significance a particular light should carry in a particular moment and why. As with the aforementioned mirror ball or the blackout. The moment of letting the audience go. These decisions can be structured, described and shared with a system but it is the human’s task to make them. That is no consolation prize. It is the realistic distribution of responsibilities between a system that commands syntax and a person who vouches for meaning.

Theatre and Concert: Two Distinct Domains

Those who take this shift in roles seriously cannot avoid distinguishing between the different contexts in which it unfolds. It would be a mistake, in this debate, to conflate theatre and concert. The demands differ fundamentally. The same applies to the possibilities and limitations of AI. In the world of concerts, the conditions for AI deployment are favourable. Repeatability is a foundational principle. The musical structure of a concert defines a clear temporal framework. Visual effect is primary and directly measurable. An AI that generates cue structures from audio data whilst drawing on a library of aesthetic references is, from a conceptual standpoint, closer to practical AI application in this domain than in theatre.

In theatre, different rules apply. A production lives through variation, through the moment when an actor interprets a scene differently. Theatre demands a degree of contextual intelligence that today’s AI systems do not possess. One must respond to a living situation, not to a fixed musical state. In theatre, the lighting designer is not merely an executor: they listen, exercise empathy and think alongside others. This quality cannot be replicated from training data. At least not today. Perhaps not tomorrow. Perhaps never. That is not an argument against AI in theatre. It is an argument for a realistic definition of its boundaries and for keeping the human in the loop for as long as those boundaries exist.

Sustainability and Efficiency: An Underestimated Argument

An aspect that is raised far too rarely in discussions within the industry. AI could lead to significantly more resource-efficient execution of lighting productions. Shorter programming times. Fewer rehearsals. Reduced energy consumption, lower travel costs, reduced production expenditure. When an AI system generates an initial functional reference structure within a matter of minutes, which an operator then refines rather than rebuilding from scratch, this has concrete implications for the production ecosystem. That is not speculation. It is a logical consequence.

Efficiency presupposes a clearly defined objective, however. And precisely here lies an often-underestimated risk: if the objective is poorly defined, if “more looks” counts as the criterion for success rather than “the right light at the right time”, then the AI system optimises for quantity rather than substance. The result would not be better lighting design, but more lighting design. Quantity over quality. The capability of the machine depends entirely on how precisely the human defines the objectives. That is not a weakness of the technology, but a requirement of the person who uses it.

Moreover, an AI that can analyse position data, intensity values and beam characteristics in real time can deploy light with greater precision. Fewer fixtures. More precise calibration. A lower overall energy consumption. It is the logical consequence of the optimisation capability of systems that make their decisions not out of habit, but according to the principle of efficiency. This optimisation presupposes a database that reflects reality, however. Here a structural problem becomes apparent that receives little attention: the data provided by manufacturers regarding fixtures is often inconsistent, measured under ideal conditions, and insufficiently standardised to permit meaningful comparison. An AI system that optimises on the basis of this data diverges from reality. The consequence: AI-supported measurement systems are not an option, but a prerequisite for any well-founded, AI-based planning decision.

In concrete terms, this means: a mobile measurement system integrated directly into the production process, automatically calibrating luminaires during the rigging phase. This is not a laboratory exercise, but a routine procedure. The measured actual values, luminous flux, colour temperature, beam geometry, are fed back immediately into the patch model of the control system, replacing the manufacturer’s nominal values. Deviations are logged. Variations over the production period are tracked. The system relies not on the data provided by the manufacturer, but on what is actually installed. This closed data loop forms the foundation for making reliable, AI-supported decisions in practice.

The Self-Learning, Self-Correcting System

The conceptually most significant development is not the automation of individual control steps, but the establishment of a closed loop. A system that does not content itself with acting, but continuously verifies whether its actions are achieving the desired effect and corrects itself on the basis of that verification.

The technical foundation of this model: the simultaneous capture of two levels of information. On the one hand, the user interface of the lighting console, indicating which commands the system has executed: which sequences are running, which settings are in effect, and which groups have been addressed. On the other, the physical effect of light in the space, captured by a camera filming the actual scene. These two data streams are transmitted in parallel to the AI system.

This dual input channel is not a luxury, but a necessity: DMX values alone are insufficient to describe the actual lighting situation. A fixture whose DMX brightness value is set to maximum can, in practice, exhibit considerable variation, attributable to the ageing of the light source, calibration differences, or the physical conditions of the environment. The physical reality must therefore be captured and verified independently. That is not a detour; it is a necessity.

The system logic functions as a closed control loop. The model sends control commands, observes the resulting lighting effect, compares it with the defined target state, and makes corrections where deviations occur. Iteratively. Continuously. A Model Context Protocol server acts as a standardised intermediary layer that normalises the flow of information between the control system and the AI model, rendering it interpretable for both sides.

The learning principle is deliberately constrained: the system is supplied only with the syntax of the command system. No design knowledge. No aesthetic rules. No predefined dramaturgy. The application of this syntax is learnt entirely autonomously. That is not simply a minor detail of the system architecture. It is the foundational design principle of the system. It is precisely this that distinguishes a system that reproduces knowledge from one that develops competency, through experience, through iteration, through the constant alignment of intention and effect.

What distinguishes this model from a conventional automation approach is its adaptability. If a fixture fails, a luminaire shifts position, or the camera detects a change in the state of the space, the system does not respond by displaying an error message, but by performing a recalculation. Not fault tolerance as a development objective, but adaptability as a foundational prerequisite. Whether this suffices to assert itself in a living theatre production is a separate question. In the world of concerts, where repeatability is a foundational principle, the answer is self-evident.

Limits, Risks and Open Questions

A show file is not merely a control file. It is the concrete realisation of conceptual decisions, aesthetic judgements, dramaturgical considerations and technical solutions. It is, in a sense, the work. The question of to whom this work belongs when an AI system has been involved in its creation has not yet found a satisfactory answer. Claims overlap: the AI developer who provided the foundation; the lighting designer who directed the system; the production company that commissioned the work; the client who paid for it. Which of them can claim a show file as their own work? This grey area is not purely theoretical in nature: it is of decisive commercial and legal significance.

For lighting designers, two aspects are of particular importance. First: if AI systems have been trained on existing productions (on show files, documented productions, archived lighting data) this may give rise to rights of further use that are not currently represented by any institutional body. Second: if a lighting concept created with the assistance of AI is eligible for protection, the question arises as to who can assert this protection and according to what criteria royalties are calculated.

There is also the risk of homogenisation. If lighting designers worldwide use the same AI tools, and these tools are based on the same training data, a new form of aesthetic convergence emerges. Not imposed. Not mandated. But structurally conditioned by the infrastructure of the tools themselves. This is a development of which those who create with light must become aware before it materialises. For when the machine smooths the path to the result, many will follow, and the distinctions that make lighting design an art will disappear, unnoticed.

Throughout its history, the field of lighting design has received new impetus whenever vocational reorientation, self-directed learning and the transgression of boundaries have been possible. AI tools can open precisely this space, or close it, if they prescribe from the outset the aesthetic model that the designer should properly be developing independently. The question is not whether the tool is powerful. The question is who holds it and whether that person knows what they are creating with it.

And finally: the question of responsibility. An AI system with access to intensity values, position data and beam characteristics could monitor photobiological exposure limits, but who is responsible when the system makes an error? This question remains, for the time being, unanswered. It must be resolved before such systems are deployed. This is not a precondition for research, but a precondition for implementation.

Outlook: And Three Things Lighting Designers Should Do Today

The reshaping of lighting design by AI is not a question of “whether”. It is a question of “how” and of the speed, with which those who create with light will learn to act proactively, rather than wait passively.

What drives me is not a fascination with technology for its own sake. It is the conviction that the conjunction of human perceptual experience and the machine’s capacity to structure can produce something that neither could achieve alone. It is not the machine that replaces the designer. It is not the designer who ignores the machine. But what emerges when both genuinely act in concert, when human intuition and mechanical precision do not preclude one another, but complement each other? For me, this conviction has long since ceased to be a mere hypothesis. It is the result of years of work at the intersection of technology, design and art.

First: actively integrate AI tools into one’s own workflow. Not as a diversion, but as a tool that must be tested, questioned and understood, for only those who know it can shape it. Those who wait for the technology to be “fully mature” are waiting too long. Maturity comes through practice, through engagement, through the collision of system logic and artistic demands. Familiarising oneself with imperfect tools at an early stage is not a disadvantage: it is the moment when lighting designers can still contribute to shaping those tools.

Second: refine one’s own language in relation to light. If one is unable to describe one’s creative intentions with precision, one cannot effectively direct an AI system either. The command of prompting does not begin with the machine. It begins with one’s own reflection. The ability to translate light into words that a system can interpret and this very ability is what distinguishes a good conversation with a director, a precise concept presentation, or a compelling portfolio. No new craft. An older one that suddenly acquires new relevance.

Third: seek out the debate within the industry rather than avoiding it. The scepticism of colleagues is not an obstacle, but a corrective factor. It is precisely in the dialogue between advocates and doubters that what truly has substance crystallises. The interplay between enthusiasm and resistance is not a one-sided process. It is the process through which those who create with light learn , not through consensus, but through constructive friction.

This exchange must continue. And it will continue, with or without consensus, with or without strategy, driven forward by those who cannot wait and corrected by those who will not wait. Whether those who create with light wish it or not, this debate will transform them. The only question is whether they will be its authors or its product. Light, logic and learning: that is not a title. That is a programme.