KillSwitch language tests how easily AI models misread code and documentation

Software engineer Dominik Picheta has created KillSwitch, an open-source programming language intended to test how easily large language models can be confused by code and its documentation. Its deliberately unusual, prose-heavy style may confuse human readers too, and the project raises questions about how reliably AI systems interpret code and distinguish factual descriptions from misleading text.

KillSwitch is a limited, interpreted language with functions, variables and if statements. Adverbs act as identifiers, and calling one involves writing an anagram of it; much of the surrounding prose is ignored by the interpreter. Its GitHub specification includes deliberate falsehoods, including a claim that the language can take over local network services, which Picheta says is untrue and intended to test whether models can spot fiction.