xonecas.io
696
Cs

coder-skills / files

coder-skills / human-writing / references / research-notes.md

5.3 KB
1# Research Notes for `human-writing`
2
3This file captures the research basis behind the skill. Keep the
4main `SKILL.md` practical; keep detailed evidence here.
5
6## Wikipedia article and source
7
8### Wikipedia: Signs of AI writing
9Source page:
10- https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
11- Source wikitext: https://en.wikipedia.org/w/index.php?title=Wikipedia:Signs_of_AI_writing&action=raw
12
13Key takeaways used in the skill:
14- The page treats AI tells as **patterns**, not proof. A single word or
15  one awkward sentence is not enough.
16- Many strong tells are about **generic significance claims**,
17  **superficial analysis**, **promotional wording**, **vague
18  attribution**, and **rigid essay structure**.
19- It flags repeated "AI vocabulary" clusters such as `additionally`,
20  `align with`, `delve`, `underscore`, `pivotal`, `vibrant`,
21  `showcase`, and `tapestry`.
22- It highlights broader structural habits: rule-of-three phrasing,
23  overused em dashes, Markdown-heavy output, canned collaborative
24  phrasing, and generic "challenges / future outlook" conclusions.
25- It also includes a useful warning section on **ineffective
26  indicators**: blandness, transition words in isolation, and formal
27  prose are not reliable proof by themselves.
28
29How it shaped the skill:
30- Focus on removing formulaic phrasing and generic uplift, not on
31  banning individual words absolutely.
32- Emphasize specificity, judgment, and natural rhythm over detector
33  gaming.
34- Add boundaries so the skill does not encourage false claims of human
35  authorship.
36
37## Research on distinctive AI style
38
39### Russell, Karpinska, Iyyer (ACL 2025)
40Source: https://aclanthology.org/2025.acl-long.267/
41
42Why it matters:
43- Frequent ChatGPT users were strong detectors of AI-generated text and
44  outperformed many automated detectors.
45- Their explanations relied not just on word lists, but also on
46  **formality, originality, and clarity**.
47
48How it shaped the skill:
49- The skill focuses on overall feel and judgment, not only lexical
50  substitutions.
51- "Human" writing should sound selective and situated, not just less
52  detectable.
53
54### Juzek and Ward (COLING 2025)
55Source: https://aclanthology.org/2025.coling-main.426/
56
57Why it matters:
58- Documents lexical overrepresentation in LLM writing, with words such
59  as `delve`, `intricate`, and `underscore` appearing far more often
60  after LLM adoption.
61- Treats this as a measurable linguistic drift, not just a meme.
62
63How it shaped the skill:
64- Include a short watchlist of repeated AI-favorite wording.
65- Frame the issue as **clusters and repetition**, not one forbidden
66  token.
67
68### Kobak et al. (Science Advances 2025)
69Source: https://www.science.org/doi/10.1126/sciadv.adt3813
70
71Why it matters:
72- Shows abrupt post-2022 increases in excess vocabulary across more
73  than 15 million biomedical abstracts.
74- Reinforces that LLM writing leaves a measurable lexical footprint at
75  scale.
76
77How it shaped the skill:
78- Prefer plain verbs and concrete nouns over polished but generic
79  vocabulary.
80- Warn against abstract noun piles and repetitive stylistic markers.
81
82### Rudnicka (Scientific American, 2025)
83Source: https://www.scientificamerican.com/article/chatgpt-and-gemini-ai-have-uniquely-different-writing-styles
84
85Why it matters:
86- Argues that major chatbots have distinct **idiolects**, much like
87  humans do.
88- Implies that AI tells shift over time and differ by model.
89
90How it shaped the skill:
91- Avoid hard-coded, forever bans.
92- Teach broader habits: specificity, rhythm, audience matching, and
93  selective judgment.
94
95### Saha and Feizi (2025)
96Source: https://arxiv.org/abs/2502.15666
97
98Why it matters:
99- Detectors often flag even lightly AI-polished human text as AI.
100- This creates false accusations and shows the limits of detector-first
101  thinking.
102
103How it shaped the skill:
104- Optimize for better writing, not for beating detectors.
105- Keep the skill centered on readable, honest prose.
106
107## Official style guidance on sounding human without sounding sloppy
108
109### Google developer documentation style guide - Voice and tone
110Source: https://developers.google.com/style/tone
111
112Key takeaways used in the skill:
113- Aim for a voice that is conversational, friendly, respectful, and
114  natural.
115- Sound like a **knowledgeable friend**, not a pedant and not a clown.
116- Avoid buzzwords, placeholder phrases, choppy repetition, and
117  long-winded sentences.
118
119### Digital.gov - Plain Language Web Writing Tips
120Source: https://digital.gov/resources/plain-language-web-writing-tips
121
122Key takeaways used in the skill:
123- Get to the point quickly.
124- Use contractions, active voice, familiar words, and short paragraphs.
125- Put the most important information first.
126
127### 18F Content Guide - Be concise
128Source: https://guides.18f.gov/content-guide/our-approach/be-concise/
129
130Key takeaways used in the skill:
131- Be specific, informative, brisk, and human.
132- Use contractions and short sentences.
133- Vary sentence length so writing stays readable without sounding
134  robotic or clipped.
135- Avoid adjective-heavy spin and caveat-bloated grammar.
136
137## Practical synthesis
138
139Across the sources, the same advice keeps showing up:
140
1411. Lead with the point.
1422. Prefer concrete nouns and verbs over abstract uplift.
1433. Use contractions when natural.
1444. Vary sentence rhythm.
1455. Match the audience's register.
1466. Avoid repeated formulae, hype, and canned summaries.
1477. Keep clarity and honesty above "sounding human".