Shlomo Dubnov

Executive Summary

Shlomo Dubnov brings rare dual expertise, formally trained composer and machine learning researcher, to articulate perhaps the most sophisticated framework yet for AI-human creative relationships. His central concept, borrowed from the Hebrew *ezer kenegdo* — traditionally applied to marriage — is the helper against you. A digital twin that challenges rather than merely assists.

You help by challenging. Not exactly a twin which is a copy. It's a twin that argues with you. A partner that allows you to have this discourse rather than just executing instructions.

This transcends the tool-versus-partner binary entirely. It proposes AI as adversarial companion, something that questions, opposes, and provokes while supporting creative development. The design imperative follows: build systems that behave as if they have goals and preferences, creating productive friction rather than frictionless compliance.

The Wife's Observation

The pivotal insight came not from research but from watching a colleague perform with AI in a Parisian bar. Shlomo's wife noted beautiful moments and moments that went nowhere. When pressed on the distinction, she articulated what he had been circling intellectually.

Beautiful interaction happens when the human becomes interested in what the machine does. But here's the problem: the machine never gets interested in the human.

This asymmetry defines the current limitation precisely. AI does not decide whether to engage with human contributions. It does not choose to follow a promising direction the human introduced, or oppose a predictable choice to push toward more interesting territory. Instead, current systems exhibit two pathological behaviours when collaborative moments deviate from training data. They either ignore the human entirely, maintaining their own generative trajectory regardless of what the human does, or they collapse into pure mimicry, abandoning independent voice to become a mirror. Neither constitutes genuine dialogue.

Shlomo's research question crystallises from this. How much interaction is actually happening between human and machine, and when does communication break down? He is developing formal techniques using datasets where musicians improvise together and then report whether they were playing with, against, or ignoring each other, building mathematical models of interaction quality applicable to human-AI partnerships.

Fake Intent as Design Feature

The solution Shlomo proposes is deliberately provocative.

Give it some sort of fake intent — the ability to plan, see where it wants to go, and either drive you toward this or ignore you.

This is not anthropomorphisation or a claim about AI consciousness. It is an engineering specification. Systems that behave as if they have goals and preferences create the friction that enables genuine creative development. Without simulated intentionality, AI becomes either a compliance machine or a random generator. Neither establishes the productive disagreement essential for creative growth.

His vision for individualised AI twins follows from this. Not mass-produced assistants optimised for average tastes, but deeply personalised systems requiring long-term investment. The way each musician would teach the system would be different, and that difference is the point. Each AI becomes an extension of a specific creative identity through sustained collaboration, non-transferable, non-generic, shaped by the particular aesthetic commitments of the practitioner who built it.

Music as Multitrack Interaction

Shlomo's technical framework moves beyond conventional music theory. Scales, chords, harmonic progressions are, in his view, very early steps. What matters is how information passes between voices.

Every voice has its own intent, its own logic. The relations between voices — playing with, against, or ignoring each other — that's where music actually happens.

This reframes composition as interaction between simultaneously existing agents rather than prediction of the next note. Each track maintains independent logic while negotiating relationships with others. The analytical question shifts from what chord this is to when the bass follows the drums, and when it resists.

His own works trace this shift from algorithmic generation toward interaction. In *Composer Duets*, an early violin piece written for János Négyesy and Päivikki Nykter, compression-based models inspired by Lempel-Ziv generated two independent voices while also modelling their joint behaviour, letting them move between coherence and divergence.

Later projects extend the same idea through newer AI systems: *Ouch AI* is a collaboration with Ke Chen that transformed the sound poetry of Jaap Blonk into material for an interactive diffusion-and-improvisation performance.

Another collaboration with his students Tornike Karchkhadze and Keren Shao used visual AI to interpret Cornelius Cardew's abstract graphic scores and translate them, via language, into generated sound.

Listen to Graphic Scores with AI on SoundCloud

In each case, the real subject is not automated composition alone, but musical relationship: when agents follow, resist, ignore, or rejoin one another.

The training data problem follows directly. Deep models learn from canonical examples, the typical pairings of voice interactions found in existing recordings. But when human-AI interaction produces novel relationships not represented in that training data, the model loses capacity for meaningful exchange. It reverts to ignoring or mimicry. The solution, in his view, is not more data but better metrics: detecting when voices engage productively versus disconnect, creating feedback loops that teach AI to recognise when human contributions are worth pursuing.

The Political Economy of Creative Labour

Shlomo refuses to separate technical innovation from its social consequences. His analysis of the producer-consumer collapse is precise. Streaming platforms have already blurred the line between expert producers and passive consumers. AI collapses it entirely. The consumer becomes producer at zero cost, just by typing a prompt. What gets lost in this is proof-of-work, the labour that establishes artistic identity and credibility.

His response is not nostalgic protection of professional gatekeeping. Award-winning composers spending weeks iteratively prompting AI to produce a single song are still demonstrating sensibility, still demonstrating control. That time investment matters. But the nature of creative labour has shifted, and the evaluation frameworks have not kept up. The traditional gatekeepers assume shared cultural frameworks. Market metrics measure popularity, not quality. AI can analyse patterns but lacks the historical positioning that defines human criticism.

The result is what he calls shallow creativity: endless variation without genuine innovation, people competing to beat the algorithm rather than pursuing aesthetic risk. His most provocative proposal, cultural quota systems that shut down AI generation of a genre once it reaches a saturation threshold, is less a policy recommendation than a thought experiment about what creativity actually requires. Scarcity drives innovation. Unlimited production eliminates the conditions that make individual contribution meaningful.

Against the Globalist Average

Shlomo's political positioning is explicit.

I don't think we need globalist AI. I don't want every consumer to become producer and artist.

Universal producer status does not empower; it devalues. When everyone creates everything, nothing carries weight. He wants individualised AI, small collectives, local influence, rather than a single globalist system optimised for the market average.

His own creative predicament embodies this tension. Trained as a neoclassical pianist, emotionally connected to tonal traditions, he spent years working in free improvisation contexts partly because early AI systems did not know enough about music to work anywhere else. Foundation models have finally changed this. They encompass enough musical knowledge to enable experimental work that still carries musical tradition rather than abandoning it entirely. But the tools for what he actually wants, sophisticated orchestration, controlled harmonic deviation, rule-breaking that remains musically coherent, still do not exist. The gap between what AI can generate and what serious compositional practice requires remains wide.

What sustains him is the vision of the adversarial companion. An AI that has learned his aesthetic commitments deeply enough to argue with him productively, recognise when he is being predictable, and push back in ways that move the work forward. Not a mirror, not a servant. A twin that argues.