The Song Cannot Explain Itself

In a September 25 report for 404 Media, Jason Koebler describes prosecutors playing two versions of an AI-generated love song during the murder trial of Caleb Flynn, who is accused of killing his wife. According to the report, investigators recovered audio files and lyrics from his phone; the songs concerned his relationship with another woman.

The courtroom spectacle is the easy story. The harder one is the distance between hearing a performance and knowing what, exactly, its maker contributed.

A performance with several authorship layers

Koebler quotes forensic investigator Joseph Wilhelm describing the recovered files as apparently AI-generated and characterizing the versions as differently pitched emotionally: one upbeat-sad, the other sad-sad. He also quotes messages attributed to Flynn that describe making the song with an AI voiceover.

Those messages complicate the convenient label “AI-generated.” In the account, Flynn says he finished the lyrics, describes loading a piano rendition into software he identifies as ElevenLabs, and complains that the result altered the outro and made the vocals more rock-like than he wanted. He says he tried to sing with passion and the AI went “a little extreme.”

That is a reported description of a workflow, not independent technical verification of which product or model produced which component. It does, however, identify the distinction worth preserving: words, melody, vocal performance and automated transformation are not necessarily the work of the same author.

A listener encounters them together. An investigator trying to establish what a person did needs to separate them.

Nothing about synthetic vocals makes a file automatically irrelevant. Someone can choose to create, save or send a generated artifact. Those choices may matter independently of whether the voice is theirs. But the emotional intensity of that voice cannot, by itself, tell us how strongly the person felt something. The performance is an output, not a calibrated instrument for measuring intent.

Keep the explanation attached

This account rests on 404 Media’s reporting, not an independently reviewed court transcript or forensic extraction. The report describes the judge permitting playback. That does not establish the full basis for admission, the parties’ arguments about the files or what weight a fact-finder will give them. Nor does the existence of the songs establish guilt.

The institutional question is therefore not “Can AI testify?” A recording does not become a witness because its singer sounds human. The question is whether the explanation accompanying it is precise enough to challenge.

A useful standard would distinguish the recovered file from any playback copy, identify what connects it to a device or account, and separate documented human contributions from inferred ones. Where the generation process is unknown, that uncertainty should travel with the artifact rather than disappear beneath the chorus. These are proposed safeguards, not procedures this reporting confirms the court followed.

That is the coherenceist demand here: an institution should not purchase a clean story by silencing the ambiguities that affected people need to contest. Accountability means keeping the route from object to interpretation visible, including where that route runs out.

The technology can make expression easier, richer or simply stranger. It does not make interpretation effortless. Once a generated song enters a consequential proceeding, its synthetic confidence is precisely what deserves no special deference.

The voice can arrive polished. The explanation still has to do its paperwork.