INVERSE AI

An intelligence that listens before it reads.

Every human being learns to speak before learning to read. Most people who ever lived never learned to read at all. What is strange, in human history, is an intelligence that only knows how to read.

The weavers of the Atlas say you judge a rug by its reverse side. The front can deceive; the back shows every knot. Today’s AI is a beautiful surface. Turn it over, and you see how it is made: from text rather than voice, in English rather than the world’s languages, judged by measures designed elsewhere. Inverse AI starts from that side, and weaves again.

Three inversions

Today’s AIInverse AI
1. Start from the voiceThe voice is a noisy version of a text to be recovered. No transcription, no learning.The machine learns like a child, by listening to thousands of hours of untranscribed speech before seeing any text. Accent, rhythm and melody are kept.
2. The machine adaptsSpeak slowly, articulate, don’t mix languages. If it fails, you spoke badly.No one speaks “badly”. Switching languages becomes a skill to learn, not an error to correct.
3. Speakers judgeSuccess means matching a standard transcription, on average.Panels of speakers judge meaning, register and irony. Results are published by variety.

Four properties, one principle

  • The voice structures everything. If a question is asked with worry, the machine must be able to hear it.
  • Knowledge stays alive. It grows each time a community validates or corrects it; earlier versions are never erased.
  • Difference is never corrected. A regional variant, a borrowed word, a switch of language: information to learn, not noise.
  • It remembers who speaks. To a grandmother of the Souss, it does not answer in administrative Arabic.

No language of passage

What the machine learns, the representations it builds and the references used to judge it stay in the original language. Translation remains possible, but only as a service at the output, on request, always labelled. An Amazigh tale can be translated into English to share it with the world. It is never passed through English to be understood.

LUMINA: follow one voice through the system

LUMINA turns Inverse AI into a plan engineers can build and researchers can criticize. Throughout, the community decides, and a provenance registry traces every step.

  1. Collect. A collector from the next village, trained and paid by the local association, records Lalla Fadma in her kitchen. Offline, uncompressed. Four consents, asked aloud in Tachelhit.
  2. Keep at home. The recording goes to her community’s own space, with its own keys and rules. The original is sealed, read-only.
  3. Listen like a child. A speech model keeps learning from untranscribed Moroccan speech: Darija, Tachelhit, Tamazight, Tarifit, Haketia.
  4. Recognize languages, moment by moment. The switches between languages are kept: a choice of language is a message.
  5. Transcribe without correcting. Writing follows the sounds actually spoken, across Arabic script, Tifinagh, Latin and Arabizi.
  6. Understand with local categories. Kinship, kinds of land, the sacred and the states of water are described as speakers think them, in open formats that outlive any machine.
  7. Answer aloud. In her Tachelhit, with the respect owed to her, and a check: wach fhamti mezyan? “Did you understand well?”

The right to take your voice back. The system is trained in separate pieces, one per community. If a community withdraws, only its piece is retrained, at a fraction of the cost, and tests confirm nothing of it remains.

What Inverse AI is not

  • Not a rejection of writing. Text stays useful; it is no longer the master.
  • Not a rejection of large models. What we refuse is English as the working language, the target or the reference.
  • Not a retreat into identity. It helps a farmer of the Souss read a document in French, starting from his own language.
  • Not a substitute for people. No machine will replace a grandmother telling a tale. It is a link in the chain, not the chain.

An honest note. LUMINA is an architecture, not yet a product. The white paper sets out hypotheses and how to test them; it does not yet report results of its own.