The Static Horizon

Justin Burmeister and I built Sift over twenty-five years. Two decades of records, half-finished sessions, debates about whether a chord or note was sitting right. By the time he died in 2020, we had more than 150 tracks, a catalog that represented something we had spent most of our adult lives assembling. In 2017, I was introduced to the first AI persona of a deceased human being - someone who had “recreated” their father after he had passed, using writings and recordings. That idea never sat right with me - how can we prove whether the AI output is even aligned with the person who died, never mind actually what that person would have thought?

Because I am an idiot, it took me almost six years to realize I had the opportunity to test that concept, in a limited way, with what Justin and I had created together.

The technical details, for those interested … I fine-tuned MusicGen, a generative audio model built on Meta's AudioCraft library, on the full Sift archive. To keep things local and ensure that nothing went up to the cloud, I ran it on my Macbook Pro M4 Max with 64GB of unified memory. The model trained across 10 epochs - 30 hours, and I spent a lot of time generating audio and listening to the output.

Some of it was genuinely impressive - and eerie at times. Tracks that sounded like things we might have abandoned in 2011. Timbres I recognized. A certain way the arrangements settled. The model had learned something real about what Sift sounded like.

But, that's also where the problem starts.

In the last months before he died, Justin and I had been recording demos of new material. Our last album, Denouement, was a love letter to our influences, but our new material was a move in a much different direction - odd time signatures, microtonal and polytonal sections, and other things that were not entirely new for Sift, but not heavily used in the past. It was still exploratory, rough in the way demos are when moving across new ground. But there was also a real direction to it - something more than just trying out brand new skills.

When I fed the model text prompts describing those qualities, describing microtonal passages and asymmetric rhythmic figures, it produced nothing remotely aligned with what we were producing - and this wasn't a misconfiguration. It was because the training data didn't contain that material. The model had never seen it.

Autoregressive audio generation works by extending patterns across what's already there. When the pattern doesn't exist in the source, there's nowhere specific to go - it’s monkeys on typewriters.

Researcher David Cropley, writing in the Journal of Creative Behavior in 2025, puts it clearly: large-scale generative models are constrained by what he calls a mathematical ceiling, one that binds them to continuation rather than conceptual break. What he describes as "Big-C" creativity, the kinds of pivots that define how a serious artist develops over time, falls outside what these systems can do structurally. The training data is the horizon, and the model can't see past it.

I've started referring to this as the static horizon (a term borrowed from cosmology, but that works here). It's what we get when a generative system is working at full capacity and still can't move as expected or directed. The outputs are convincing, the aesthetic is intact, but the whole thing is ultimately a loop. High fidelity to a past that no longer has a future destination.

There are also real psychological stakes here, and I want to be mindful about them. Two studies published in Frontiers journals in 2025 examined what happens to people who use posthumous AI simulations of people they've lost. One described a state the researchers call "suspended closure in grief," where a simulation offers just enough emotional continuity to prevent someone from fully accepting that the person is gone. The other looked at "affective outsourcing," the way people delegate their grief processing to digital systems, and identified a real tension between how vulnerable users are in that position and the fact that the machine has no intention whatsoever. It's producing outputs, not caring (it is, literally, unable to care). That reality matters, even when the outputs may feel meaningful.

It was emotional hearing some of the AI-generated “Sift” tracks. There were moments that felt like finding something in a drawer I'd forgotten about. One of those cassettes from a long gone rehearsal space that was recorded on a boom box in the late-90s. But it was also clear that I was hearing an archive expander - a good-enough echo. The model was producing material that sounded like our mid-career work because that's all it had been exposed to. It wasn't carrying anything forward. It was holding something in place.

And that creates a specific kind of distortion. Not just a technical one. When you generate convincing mid-period Sift material in 2026, you're making an implicit claim about who Justin Burmeister was as an artist. The claim is technically accurate in a narrow historical sense, and it's also quietly false, because the person it represents was not standing still. He was somebody who had grown restless with the boundaries of what we'd built together and was in the middle of working out what came next. That version of him gets no representation in the model. It’s as if he died twice.

Morris and Brubaker, writing in the CHI 2025 conference proceedings on AI afterlife systems, call this a failure of temporal framing. The AI can approximate a persona, but it can't evolve across time. Any claim that a generated output captures what Justin would have made is a statistical guess dressed up as knowledge. It's fidelity to a catalog, not to a human being. To me, it’s a ghost that’s always with us, but it never evolves - and runs a real risk of slowing our own evolution.

I also want to be clear that I don't think the work I did is worthless or that the technology doesn’t deliver real value. There may be legitimate uses for skilled historical mimicry, archival ones, educational ones. But they need to be named accurately.

Posthumous AI generation is not continuation. It's retrieval with a convincing surface. When it gets framed as something more - when it gets presented as creative momentum or extension, it does something I think is genuinely harmful: it competes with the more honest, more messy, more human reality. That's a loss and the model didn't capture the loss. It glossed over it.

Spitale and colleagues, in a 2025 preprint on ethical design of AI afterlife systems, argue that any responsible implementation has to commit to truthfulness, meaning an honest acknowledgment of the limits of what the system can actually know about the person it's reconstructing. That standard must be taken seriously.

The real artifact of this project isn't the generated audio. It's what the model couldn't reproduce - those rough, unfinished demos sitting at the edge of a dynamic horizon. The model's failure to capture them is, paradoxically, likely the most accurate thing that came out of this experiment. It shows that Justin's creative life had momentum and was expanding, evolving - even as he closed in on his fifth decade of life.

He was a person who kept moving and evolving. A generative model trained on almost everything he created with Sift couldn't predict where he was going, because where he was going hadn’t been captured in static form and given to the model.