Chorus Inútil

Chorus Inútil is a real-time choir of ChatGPT text-to-speech instances repeating nonsensical syllables and repeating letters. When forced to vocalize unintelligible text, it glitches and diffracts, revealing ghosts, music, rhythms, and words trapped in its training data.

Unlike traditional text-to-speech, which synthesizes phonetics via concatenative synthesis, ChatGPT’s text-to-speech is built using machine learning. Instead of stitching together pre-recorded sound snippets, it’s trained to recreate human speech by training off of hundreds of hours of captioned audio data. ChatGPT’s text-to-speech, therefore, includes breathing, emotion, and intonation.

Normally, it performs well when speaking normal and expected words and phrases. However, when the text falls outside of what it expects, it's forced to compensate and drift away from its normal functionality. Vocalizing phrases like ‘oooooooooooooooo’ or ‘kkkkkkkkkkkkkkkkkk’ force it to pull from the peripheries of its training data, reaching into corners populated by discarded data like background noise, music, and sighs.

As the piece runs and continuously vocalizes these illegible inputs, it harmonizes in glitch – in outbursts of unintended laughter, crying, words, gibberish, and sometimes even music.

Chorus Inútil extracts unseen landscapes embedded within a widely used and integrated AI tool and brings awareness back to its underlying predictive and mechanical functioning. As newer models, like GPT-5, continue being released and marketed as smarter than the last, it becomes easier for us to detach ourselves from its internal mechanisms. How different Is a well-presented GPT output, to one that endlessly repeats a single character? Just how the stranger parts of its text-to-speech data are cast in shadow, only conjured via glossolalia, ChatGPT’s skewed perspectives and intentions can quickly be shadowed under the guise of the latest model.

Chorus Inútil was exhibited in Copeland Gallery at the exhibit ‘BitRot’ by The Phreaking Collective.