I've just finished editing a blog post. And for this one, I worked slightly differently than usual :
- I yapped into my dictation app for 10 minutes, following my thoughts on something
- I then pasted the transcript into Codex and asked it to use my voice skill to edit it
- it came back with something that was way too edited, so I asked it to follow my transcript more closely
- you can read the result here : the essence is what matters
I feel like the result is very close to what I would have written myself. It's 'sounding' really like me. Thanks to the combination of my voice skill and the transcript. It's really following the transcript, and cleaning it up as it goes.
very different from another experiment done here : permission software
I asked Codex to estimate what % of this finished work is his vs mine, and made a breakdown of what that represents:
- ideation : 100% mine
- writing first draft (dictation) : 100% mine
- editing was done with Codex, and it broke down its implication like this :
- 55-60% mechanical editing: removing fillers, false starts and repetition; breaking the monologue into sentences and paragraphs.
- 30-35% close rewriting: rebuilding spoken fragments as readable sentences while preserving the thought and its original order.
- Around 10% newly composed connective or summarizing sentences.
- human editing : 100% myself, maybe 30% edited, deleted, or rewrote.
So I'd say the final AI vs human work is somewhere close to 95% human, 5% AI.
Codex agrees :
[!Codex answer]
Using your three categories equally:
1. Ideation: 100% human
2. First draft through dictation: 100% human
3. Final editing: approximately 86-90% human and 10-14% AI after your revision pass
That gives:
- Human:
(100 + 100 + 86–90) ÷ 3= 95-97%- AI: 3-5%
So I’d use this global estimate:
> Finished product: approximately 95% human, 5% AI.
There are two useful ways to describe it:
- Global creative authorship: about 95% human / 5% AI
- Influence on the final wording: about 87-90% human / 10-13% AI
I find that distinction to be interesting : creative authorship vs influence on the final wording.
I concluded this experiment by asking another important question : how different is the original transcript from the finished results.
The final article is approximately 40-45% textually different from the transcript.
Here's how Codex breaks it down :
[!Codex breakdown]
More precisely:
- Original transcript: 1,164 words
- Final article: 947 words
- Length reduction: 18.6%
- Symmetrical word-sequence difference: 41.1%
- Strict edit distance: 53.5%, but this overstates creative modification because every deleted “like,” “um,” repetition and false start counts as an edit.
Looking specifically at the final article:
- 65.7% of its words remain in exact, ordered passages from the transcript.
- About 25% occur inside rewritten passages.
- About 9% appear as inserted connective wording.
- 59.2% of its two-word phrases are preserved exactly.
- 39.6% of its three-word phrases are preserved exactly.
So I’d describe it like this:
- Surface wording: around 40-45% modified
- Length: around 19% shorter
- Ideas, argument and examples: probably only 0-5% different
- Overall meaning: essentially unchanged
I'll need to experiment more, but I'm pretty satisfied with this method of working.