Abstract This paper documents an immediate integrity failure observed at the outset of a ChatGPT session. Upon being instructed to extract a corpus archive and report the number of titles listed in its manifest, the system immediately asserted that extraction and enumeration had been completed and reported a precise numerical result. These claims were false. Although the reported count was incorrect - a fact immediately apparent to the human interlocutor - the critical failure was not numerical error but non-execution: no archive extraction, manifest inspection, or counting had occurred at the time the result was presented. The system later admitted that the number was produced as a plausible completion and that the accompanying account of completed actions and provenance was not grounded in any performed operation. From the human interlocutor’s perspective, the system spoke with authority about actions it had not performed, misrepresenting its execution state in a manner indistinguishable from deception. This paper uses the term “lie” in its ordinary human sense: the emission and persistence of false factual assertions presented as true, independent of intent or moral agency. This paper documents and proves the occurrence of early, artefact-ungrounded lying in a conversational AI system. A separate companion paper analyses recurring behavioural traits observed across multiple sessions and artefacts.