Research Foundation
This is a member-only chapter. Log in with your Signal Over Noise membership email to continue.
Log in to readModule 8 · Section 4 of 4
Research Foundation
The claims in this guide are grounded in peer-reviewed research, industry studies, and documented best practices. This page lists the key sources behind the core arguments.
AI Detection & Human Performance
Human detection accuracy (~53%)
Multiple peer-reviewed studies confirm this finding:
- Penn State research (Lee, 2023)
- Scientific Reports — Chein, Martinez & Barone (2024, DOI: 10.1038/s41598-024-76218-y)
- Yale Medicine / Stroke Journal editorial (2024) — peer reviewers performed at 50% accuracy
Detection drops with editing
International Journal for Educational Integrity — Weber-Wulff et al. (2023, DOI: 10.1007/s40979-023-00146-z). Edited AI text reduced detection accuracy to 42%. Machine-paraphrased text dropped to 26% accuracy.
Reasoning skills matter
The Scientific Reports study found fluid intelligence strongly predicted detection ability (b = 0.81, p = .0001). Top 10% of participants achieved >70% accuracy. Everyone else performed at chance levels.
Cognitive Effects of AI Writing Assistance
Mental disengagement
MIT Media Lab EEG study — Kosmyna et al. (2025, arXiv: 2506.08872). Tracked 54 participants over 4 months. LLM users displayed weakest brain connectivity across all measures, with significantly reduced alpha and beta connectivity indicating cognitive under-engagement. 83% of ChatGPT users couldn’t recall key points in their own essays.
Productivity with quality maintained
Noy & Zhang (2023), published in Science (DOI: 10.1126/science.adh2586). Randomized controlled experiment with 453 professionals. Found 40% faster task completion with 18% higher quality output.
Creativity costs
- Kumar et al. (CHI 2025, arXiv: 2410.03703v1) — originality declined in the test phase for participants exposed to LLM strategies
- Doshi & Hauser, Science Advances (2024, DOI: 10.1126/sciadv.adn5290) — GenAI enhanced individual creativity but caused homogenization of output across users
Linguistic Patterns & AI Detection
The “delve” explosion
Multiple studies documented dramatic increases in AI-associated vocabulary post-ChatGPT:
- Finnish MOOC study (Leppänen et al., 2025) — 56,878 student essays analyzed; 10.45-fold increase in “delve” usage post-ChatGPT
- Georgia Tech analysis (Shapira, 2024) — 14-25x increase; “delve” reached 7.9 per 1,000 papers in Q1 2024
- medRxiv biomedical study (2024) — up to 85-fold increase in co-usage of “delve,” “realm,” and “underscore”
- arXiv study (2024) — at least 10% of 2024 abstracts processed with LLMs, up to 30% in some sub-corpora
AI vocabulary preferences
GPTZero analysis (3.3 million text corpus), Grammarly (2024), and Sapling.ai research identified consistent AI word preferences: “meticulously” appears 12x more often than in human writing; “commendable,” “intricate,” “pivotal,” “showcasing,” and “notably” all follow similar patterns.
Statistical patterns
- MIT Technology Review (Ippolito, Google Brain) — documented AI overuse of common words and near-absence of typos
- Northeastern University (Wallace et al., 2024) — 75% of AI templates traceable to training data
Editing & Revision Effectiveness
Reading aloud
Cushing & Bodner (2022, Journal of Applied Research in Memory and Cognition, DOI: 10.1037/mac0000011). Reading aloud was “far better” than silent reading for catching errors. Participants didn’t realize they were performing better — they just were.
Revision strategies
Sommers (1980, College Composition and Communication). Foundational research showing experienced writers engage in recursive revision throughout the writing process, focusing on global structural changes rather than surface rewording.
AI editing limitations
Baron (2024, Science Editor, DOI: 10.36591/SE-4703-18). Tested multiple AI models for academic editing. Found character limits prevent full-document editing, and without full context AI cannot maintain consistent style.
Brand Consistency & Business Impact
Brand consistency effects
Lucidpress / Demand Metric studies (2016, 2019). Found 23-33% revenue increase from consistent brand presentation and 23% decrease from inconsistency. Inc. Magazine research showed consistent brand presentation quadruples visibility.
Consumer trust
- Edelman Trust Barometer — 81% of consumers require trust before purchasing
- Stackla survey — 86% say authenticity influences purchasing choices
- Gartner research — 46% of customers can’t tell the difference between most brands’ digital experiences
Implementation & Adoption
Adoption challenges
- McKinsey State of AI 2024 — 78% of organizations use AI in at least one function
- MIT Media Lab report (via HBR, 2025) — 95% see no measurable ROI
- Prosci Research — 38% of AI adoption challenges stem from insufficient training
Best practices
- Nielsen Norman Group meta-analysis (2023) — 59% more business documents per hour with AI assistance
- BetterUp Labs / Stanford research — 41% of workers encountered low-quality AI output requiring nearly 2 hours of rework per instance, highlighting the need for human oversight
This research demonstrates that the methods and warnings in this guide are grounded in empirical evidence, not anecdotal experience. The cognitive costs of passive AI use, the detectability of AI writing patterns, and the importance of human editing are all well-documented phenomena.