Artificially generatedSemantic Scorecard: the moment you cut a word, something is gone
A production statistic: how much of a piece is mine, how much came from my AI tools, and what the tools did to the meaning.
A Semantic Scorecard measures how much original wording is in a piece, how much the model added, and whether the meaning got finer or coarser. A summary is always lossy compression. If nothing were lost, it would still be the original text. The moment I cut a word, something is gone. Measuring that is semantic scoring. Not lossless reduction. A sort that puts a bill on itself.
Not who owns the byline. Who built the sentence, and what happened to the meaning.
Media made with neural nets produces a new lie when the split stays invisible: everything looks like one voice. The scorecard makes the shares visible. It belongs under the piece, not in a log nobody reads. Waiting for Grok Godot got the first card. This page is the instrument you fill the same way every time.
This is Semantic Mining applied to our own output. The mine this time is the dictation. The ore is my sentences. The sorter is the tool. The card says how much ore passed through and how much the machine milled.
Origin, operation, meaning-delta. Three numbers, not one.
A single figure, “72 percent John”, is the door, not the whole card. It measures wording, visible sentences that stood in the dictation as they are or almost. It does not measure whether the tool moved the meaning while leaving the words.
| Axis | What | Unit |
|---|---|---|
| Origin | Share of visible wording already in the input, against share the tool set new | Percent of word mass in the finished piece |
| Operation | What the model did: grammar, word choice, synonyms, sentence order, glue, structure, sources, image | Share of the model’s work, not of the word mass. May overlap. Adds up on the model side, not on the whole text |
| Meaning-delta | Did the meaning get finer, coarser, stay, or get added | Percent of meaning, estimated, in four baskets: refined, coarsened, passed through, added |
Model operations, fine:
- Grammar (case, commas, typos: Grog to Grok)
- Word choice and synonyms
- Sentence order
- Glue: new sentences that only connect
- Structure: sections, TOC, sequence
- Evidence: sources, fact table
- Image: canvas, cake, cover
Meaning-delta, fine:
- Passed through: the sentence stands, the meaning stands.
- Refined: the same thought, sharper. Layer 4 and 5 put in order after the lookup.
- Coarsened: lossy. Beckett cut to one sentence. The magazine lookup dropped. One word less.
- Added: a claim that was not in the dictation. Research John did not speak.
If a summary lost nothing, it would still be the original text.
You can coarsen meaning by picking and compressing. Something is lost by force. It is lossy compression, not lossless reduction. The moment I cut a word, something is gone. That holds for a human editor and for a model that “tidies”. Tidying is already delta. Anyone who treats grammar as a neutral filter is lying about meaning.
The scorecard makes that loss visible instead of hiding it under style. A high John percentage with high coarsening means: the words are his, the cut cost meaning. A low John percentage with a lot of added means: the piece is research, not dictation. Both are allowed. Both have to sit on the card.
Three pieces, three production types. The card shows the difference the byline hides.
Applied to the three newest McGrinsey articles before this one. Estimate by reading the dictation against the piece, not a token diff from the session log. Numbers rounded, because fake precision to 1.3 percent would be a lie here.
1 · Waiting for Grok Godot
Origin: 72 percent original wording, 28 percent model.
Operations inside the 28: grammar and typos 6, glue 9, structure 7, evidence 3, image 3. Percentage points of the whole text.
Meaning-delta: passed through 70, refined 6 (layers 4/5 after the lookup), coarsened 8 (Beckett to one sentence, lookup cut, dictation repeats collapsed), added 16 (Godot line, links, fact table, outline cake, first scorecard).
2 · Semantic Mining
Origin: 61 percent original wording, 39 percent model.
Operations inside the 39: grammar 4, glue 14, structure 8, evidence 6, image 7.
Meaning-delta: passed through 58, refined 10 (two machines in sequence made explicit), coarsened 12 (atelier job, edit button, writing-paradigm meta taken out of the public piece: loss against the dictation, gain for the reader), added 20 (fact table, canvas, scraper rules as architecture not workshop).
3 · American Pie
Origin: 14 percent original wording, 86 percent model.
Operations inside the 86: grammar 2, glue 8, structure 12, evidence 28, image 10, research sentences 26. The dictation was questions plus a political reading, not an article body.
Meaning-delta: passed through 12, refined 8 (Miss American Pie, levee, not D or R), coarsened 4, added 76 (history, sources, White Terror, Christie’s, Telegraph). The other operating mode: not sorting ore, opening a new mine because the ore was only questions.
Chinamerica, a day older, sits closer to American Pie than to Grok Godot: research body, thesis from John, sentences almost all tool. Its card belongs on a later sheet, not in these three, or “last three” becomes four.
The refinement is what a single percentage swallows.
Across the three pieces the tool mostly did four things. First, turn speech into case and sentence boundaries. Second, glue where fragments would lose the reader. Third, structure: sections, so the stream becomes stairs. Fourth, evidence and images that were not in the dictation.
Meaning changed, finer or coarser, sits around 14 percent on Grok Godot (6 refined plus 8 coarsened), around 22 on Semantic Mining, around 12 change to the incoming meaning on American Pie plus 76 added meaning. Added meaning is the expensive kind: it can be wrong, and it looks like voice. Hence the fact table. Hence this card underneath.
Rule for the next piece: when John dictates, origin aims above 60. When John only asks questions, origin under 20, and the card has to say so before anyone thinks he wrote the White Terror that way himself.
| Claim | Basket | From where |
|---|---|---|
| A summary is always lossy | Thesis, used as method here | John, 9 September 2026. Information theory: shorter without loss only if there was redundancy. Cutting a content word is loss. |
| Grok Godot 72 / 28 | Estimate | Reading dictation against piece, rounded. No token diff from the session log. |
| Semantic Mining 61 / 39 | Estimate | Same method. |
| American Pie 14 / 86 | Estimate | Dictation was questions, body is research. |
| The percentages are exact | No | Rounded, visual check. A second count may move them a few points. Not the direction of the three pieces. |
- Waiting for Grok GodotFirst scorecard in the piece. 72 / 28.https://mcgrinsey.com/magazin/waiting-for-grok-godot/
- Semantic MiningMeaning from the unknown. The mine this card turns on our own output.https://mcgrinsey.com/magazin/semantic-mining-meaning-from-unknown/
- American PieQuestions plus a reading, body is research. Origin 14.https://mcgrinsey.com/magazin/american-pie-miss-american-pie-levee/
- McGrinsey writing skillSpoken-input rule, 9 September 2026: dictation stays ore.MCGRINSEY_SKILLS/WRITING___EN.md
Cut-off: 9 September 2026. Scorecard of this piece itself: origin about 48 John (lossy lines, three axes as the brief), 52 model (building the instrument, the three applications, canvas). Meaning-delta: refined 15, coarsened 5, added 32, passed through 48. The card does not lie about itself.
