Harold Cohen (1928–2016) spent the better part of four decades doing something few of his painter peers would have considered a serious artistic pursuit: writing software. The result, AARON, is now routinely credited as the first genuine AI artmaking program — a claim worth scrutinising rather than simply repeating. Cohen began developing it in the early 1970s and continued refining it until his death, at a moment when “artificial intelligence” meant something closer to symbolic rule-systems than the self-learning models the term now conjures.
That distinction matters more than the marketing shorthand around AARON tends to allow. The software could not expand its own knowledge base; every parameter — figures, plants, spatial logic in its figurative period — was hand-coded by Cohen. What it possessed instead was a bounded but genuine autonomy: within those fixed constraints, AARON navigated millions of compositional possibilities Cohen had not explicitly dictated. It is this friction between authored limitation and unauthored outcome that gives the work its interest, and its difficulty. AARON was not generating novelty from nothing; it was exercising real choice within a closed system Cohen alone had built.
The more contentious question is what that arrangement means for authorship — and it’s here that easy narratives about “AI collaboration” tend to flatten something genuinely knotty. Cohen signed every collaborative work as its creator, never AARON. Yet the software has continued producing new plots posthumously, prompting the obvious institutional problem of what to call work made after only one half of the partnership remains active. The Harold Cohen Trust’s position, at least, refuses ambiguity: an authentic Cohen/AARON work required three elements simultaneously present — Cohen, the software, and the iterative process of adjustment between them. On that reading, the body of work effectively closed in 2016, and everything generated since is AARON’s output alone, not Cohen’s.
Whether that’s a satisfying resolution is a separate matter. It sidesteps rather than answers the deeper question raised by any generative system: at what point does a tool’s contribution to a composition constitute authorship rather than execution? Cohen’s own answer — that AARON was less a tool than an “other self” — suggests he never fully believed the distinction held in the first place, even as he continued to sign the work under his own name.
What is not in dispute is the scale of the technical undertaking. Cohen wrote AARON from scratch, without precedent, and without any assurance that a computer could be made to produce anything resembling deliberate compositional judgment. He did so after already establishing himself as a painter — a career he effectively set aside to test whether a machine could be given something like taste. Measured against the current AI-art landscape, where questions of authorship, autonomy, and dataset are being litigated in real time, Cohen’s decades-long experiment looks less like a historical curiosity and more like the first serious attempt to ask the question everyone is now asking again.