Four ways to mark AI content without the left stripe
Decision (Sept 26, 2026): Marginalia is the site-wide style. The other three are kept in docs/assets/ai-voice.css and can be previewed on any page with ?ai=tabs, ?ai=perforated, or ?ai=tiles, or switched permanently by changing one line in docs/assets/ai-voice.js.
Every option keeps the non-negotiables from the research: AI content is recognizable without color alone, carries a text label, names its source, and shows whether a person has reviewed it. Each option is shown in the same three places: a tutor reply (screen A), an AI draft block in the builder (screen B), and an AI summary on the results page.
AI speaks the way a scholar annotates a book: in the margin, in the reading serif, with footnoted sources. There's no box and no stripe. A small italic glyph in the gutter and a small-caps name carry the identity, and the citations become real footnotes, which fits an academic audience.
Rule: AI text is set in Fraunces with a gutter glyph ("※" for tutor, "¶" for drafts); human text stays in Plex Sans. Provenance sits in the left margin.
Tutor reply · screen A
Why isn't switching to F1 enough here?
※
course tutor · hint 1 of 2
Think about what F1 is computed from.1 If the same customers sit in both sets, which inputs to F1 are already contaminated?
1 Week 3 slides, p. 42 Reading 3.2, §2
AI draft block · screen B
AI draft, from Week3_slides.pdf p. 4–7 · v3
A headline metric is a claim about the typical useraverage person. Fairness questions are about who the average ignoreshides.
AI summary · results
¶
summary · written by the tutor from your answers
You were solid on subgroup reporting and proxy features. The gap is person-level leakage.1 Next: the 4-minute "Subgroup metrics" chunk, then two new review cards.
Each AI block is a tinted card with a small tab at the top, like a file folder or recipe card. The tab holds the label and the state. It feels tactile and a little playful, and it scales: human-written blocks get a neutral tab, so every block announces who wrote it the same way.
Rule: Every content block has a tab. Violet paper means AI, warm paper means human. The tab text says who wrote it and whether a person has checked it.
Tutor reply · screen A
Why isn't switching to F1 enough here?
Tutor · hint 1 of 2
Think about what F1 is computed from. If the same customers sit in both sets, which inputs to F1 are already contaminated?
Week 3 slides, p. 4Reading 3.2, §2
AI draft block · screen B
Written by you
In the loan example, the team reported 94% accuracy…
AI draft · Week3_slides.pdf p. 4–7
A headline metric is a claim about the typical useraverage person. Fairness questions are about who the average ignoreshides.
AI summary · results
Summary · from your answers
You were solid on subgroup reporting and proxy features. The gap is person-level leakage. Next: the 4-minute "Subgroup metrics" chunk, then two new review cards.
Drafts look like tear-off slips: a dashed, perforated edge with ticket notches and a small rubber stamp. The shape itself means "not final yet." When an author accepts a block, the perforation turns into a solid edge and the stamp changes to Reviewed. Review state becomes something you can see at a glance down a long course.
Rule: Perforated edge = unreviewed AI output. Solid edge = kept. Tutor chat uses a borderless tinted bubble with a mono stamp line, since chat is never "kept."
Tutor reply · screen A
Why isn't switching to F1 enough here?
Tutor · hint mode
Think about what F1 is computed from. If the same customers sit in both sets, which inputs to F1 are already contaminated?
Week 3 slides, p. 4
AI draft block · screen B
AI draft · unreviewed
A headline metric is a claim about the typical useraverage person. Fairness questions are about who the average ignoreshides.
from Week3_slides.pdf p. 4–7
Reviewed · Dr. Okafor
Which change would you make first before trusting the 94% figure?
AI summary · results
Summary · from your answers
You were solid on subgroup reporting and proxy features. The gap is person-level leakage. Next: the 4-minute "Subgroup metrics" chunk, then two new review cards.
This one is built from the project's name. AI is signed with a small four-tile mosaic mark, and AI cards carry a faint cluster of tiles fading in from the top-right corner, as if the content were laid in from a larger mosaic. Citations use a small diamond tile. It's the most brand-owned option: it can only belong to Tessera.
Rule: The mosaic mark plus a text label always appear together. Corner tiles appear on AI cards only, never on human content, and never behind body text.
Tutor reply · screen A
Why isn't switching to F1 enough here?
Think about what F1 is computed from. If the same customers sit in both sets, which inputs to F1 are already contaminated?
Week 3 slides, p. 4
AI draft block · screen B
AI draft · Week3_slides.pdf p. 4–7
A headline metric is a claim about the typical useraverage person. Fairness questions are about who the average ignoreshides.
AI summary · results
Summary · from your answers and tutor chat
You were solid on subgroup reporting and proxy features. The gap is person-level leakage. Next: the 4-minute "Subgroup metrics" chunk, then two new review cards.
Kept for later: tiles for identity, perforation for state
The options solve two different problems, and the strongest system combines them instead of choosing one:
Who made this? Use the mosaic mark and a text label (option 04). It's brand-owned, it works at 14px in a chip, avatar, or tab, and it replaces the generic sparkle for good.
Has a person checked it? Use the perforated edge that becomes solid when kept (option 03). This is the only option where review state is visible without reading any text, and review coverage is the core promise of the product.
Where did it come from? Use diamond-tile citations in chat, and marginalia-style footnotes in long-form reading (option 01), where they suit an academic audience.
Tutor reply
Why isn't switching to F1 enough here?
Tutor · hint 1 of 2
Think about what F1 is computed from. If the same customers sit in both sets, which inputs to F1 are already contaminated?
Week 3 slides, p. 4
AI draft block
AI draft · unreviewed · Week3_slides.pdf p. 4–7
A headline metric is a claim about the typical useraverage person. Fairness questions are about who the average ignoreshides.
AI summary
Summary · from your answers and tutor chat
You were solid on subgroup reporting and proxy features. The gap is person-level leakage. Next: the 4-minute "Subgroup metrics" chunk, then two new review cards.