From time to time, writers come to me to read through their draft or layout their self-published fiction. The old man had hired me for such jobs, but that time, he told me, it would be different: he can't see anymore, writes blind, and this is not fiction. "It is a bit ridiculous, but I started going to therapy at nearly 90 years old, and I want my notes to be readable for someone else. I don't know. Maybe it can be useful. Can you type my journal?". He is a bit embarrassed by the request. I behave like an MD in front of a naked patient and pretend it's just like his previous books (auto fiction, family memoirs) while leafing through one of the notebooks.

It is legible once you get used to his writing, "but this is just 2022. It gets much worse in 2024", he warns me. I can decipher it, but typing everything seems like a waste of time. Cursive writing isn't legible to automatic text detection software, so I eventually settle with a hybrid solution: reading the journal out loud, having an AI transcribe it, and going over the output – it is quick enough to make the pay OK and he has a lot of notebooks to transcribe. The text is interesting (how many personal journals have you read as an adult?), enough to make the process worth my interest on top of my time. Besides, the transcription mistakes are curious.

For instance, the text is grammatically feminine (since it recognizes the reading voice as being that of a woman), despite being written by a man and using masculine grammatical forms. It also invents plausible French words (I had to check their non-existence in my biggest dictionaries), and it negates words (the rarest ones) that actually exist. It prefers to invent rather than use an odd word. It also ignores the more complex forms of French conjugation. The old man is very articulate, so all this adds to my workload, and I wonder how much leveling and normalizing is done by these tools (Is the everyday French used in office meetings even that rich? I don’t know).

My favorite type of mistake relates to so-called hallucinations. These models were trained on subtitled videos: they associated whatever was said with the subtitle, and when there was text but no voice, or voice but no text, they got the sense that they should get creative. Especially at the end of a video where you would normally find mentions such as "translated by Studio XYZ", "Credits: Jane Smith" or "Find more subs at moresubs.com" and “Follow me for more vids !". The involuntary silence I might add at the end of every recording (I record 10 minutes at a time, my voice strains beyond that) is then understood as meaningful, and the transcriber software tends to add things that do not exist (it invents, hallucinates, lies – pick according to where you stand in the Reality wars).

Here are the first few silent lies:

Sous titres par Marseille / Subtitles by Marseilles

Je m’en saisis au long de l’autre / I'll take it upon along the other

Ok. Je t’aime ! / OK. I love you !

Merci à vous. / Thank you all.

C’est bon ! / It's good !

Je vous ai dit. / I told you.

Sous-titrage M.S.A. / Subtitling MSA

Je vous ai dit que je vous ai dit ! / I told you that I told you !

J’ai été une autre vidéo. / I have been another video.

J’espère que vous allez bien. / I hope you are alright.

C’est la première fois. / It is the first time.

Metteur ! / Putter ! (or Placer !)

Je vous invite à la fin de la vidéo ! / I invite you at the end of the video !

Les vatiques. (One of those words that could be French. The closest existing word would be viatique : last rites)

C’est parti ! / Let's go !

Fiquez ! (Another plausible invention. It sounds like an old way to tell someone to care less about something : faire fi / fiquer)

Enfin… / Finally...

La décision d’un peu plus de temps. / The decision of a bit more time.

Je vous invite à tous. / I invite you to all.