How decaf Detects and Hides AI-Generated Posts on LinkedIn and Reddit

Reddit and LinkedIn feeds are increasingly filled with AI slop. A post that reads like a genuine personal insight often turns out to be 400 words produced in thirty seconds and published purely for engagement. decaf now filters these posts on both platforms, and this explains exactly how it works and why no approach is perfect.

The limits of AI detection

Firstly, no AI detector is perfect. Academic research is consistent on this point. Commercial AI detectors flag human-written text as AI-generated at alarming rates, with false positive rates ranging from under 1% to over 50% depending on the tool (University of San Diego Legal Research Center, 2025).

There's more to worry about. Studies show these detectors disproportionately flag non-native English speakers and neurodivergent writers because both groups naturally use repetitive phrasing. That's exactly what AI detection heuristics key on (Liang et al., Patterns, 2023).

This shapes decaf's entire approach: wrongly hiding a real person's post is worse than letting an AI post through. Every design choice reflects that judgment.

The two-gate system

decaf uses a two-part detection model based on approaches recommended in recent academic literature (Kobak et al., Science Advances, 2025; Reinhart et al., PNAS, 2025).

Gate one: hard flags. Machine artifacts that almost never appear in genuine human writing. Find one and the post is hidden immediately, no scoring needed.

Gate two: soft signals with a high bar. Multiple weak signals that accumulate into a score. A post is hidden only when the score reaches 6 AND at least two distinct signals have fired.

The "two signals minimum" rule is the precision safeguard. No single soft signal, however strong, can hide a post by itself.

Hard flags

These are traces left by ChatGPT and similar tools when people copy-paste. Their false positive rate is essentially zero because they simply don't appear in organic human text:

• ChatGPT tracking parameters: utm_source=chatgpt.com, utm_source=openai
• Citation placeholder junk: turn0search0, turn0image1 from pasted ChatGPT citations
• Escaped markup: oaicite, oai_citation, :contentReference
• Lenticular bracket artifacts: 【85† and similar patterns
• The footnote-return symbol: ↩
• Refusal phrases that survive: "as an AI language model", "I cannot fulfil that request"

Soft signals

Posts shorter than 25 words are skipped entirely. Short text doesn't contain enough signal and produces the most false positives.

Note: The scores and weights shown below are for demonstration purposes. As we experiment with different detection techniques and as AI writing patterns evolve, these values will likely change.

Typography and formatting

LLMs produce specific typographic patterns at statistically high rates (Wikipedia, Signs of AI Writing):

One em dash alone scores 4, which is below the threshold. One em dash plus one arrow scores 8 with two distinct signals, which clears both gates. That's intentional. Strong tells pair to flag a post. Neither alone can do it.

The rightwards arrow is worth calling out. ChatGPT routinely uses it to imply causation: "problem → solution", "input → output". It's rare enough in genuine posts to carry real signal weight.

Rhetorical templates

AI models reuse the same argument structures regardless of topic. Research confirms this "voice" is statistically distinct from human writing (Reinhart et al., PNAS, 2025):

Excess vocabulary

LLMs over-use specific words. Kobak et al. (2025) identified them through corpus analysis. Wikipedia editors documented similar patterns. decaf uses a focused 20-word list:

delve, intricate, underscore, showcase, pivotal, tapestry, testament, realm, landscape, leverage, robust, multifaceted, elevate, holistic, streamline, synergy, paradigm, groundbreaking, transformative, actionable

Each match scores +1, capped at +4 total. This prevents a legitimate technical post from being wrongly flagged just for using genuine vocabulary. Sentence-starters like "Moreover" and "Furthermore" each score +1.

How the threshold works

A post is hidden only when the score is 6 or higher AND at least 2 different signals have fired.

What decaf doesn't do

No third party products are used. These would be expensive and slow. decaf doesn't use perplexity scoring or neural network classifiers like DetectGPT or RoBERTa. Although these may produce better recall, they require network access or compute that conflicts with decaf's privacy-first approach.

A versioned ruleset

The weights and vocabulary list are named constants designed to be re-tuned as AI models evolve. Research shows that tells like em dash overuse peaked in 2023-2024 and are actively suppressed in newer models (Kobak et al., 2025; Juzek & Ward, ACL Findings, 2025). The filter will drift, and the architecture accounts for that.

References

• Wikipedia. Signs of AI writing.
• Kobak, D., et al. (2025). Delving into LLM-assisted writing in biomedical publications. Science Advances.
• Reinhart, A., et al. (2025). Do LLMs write like humans? PNAS, 122(8).
• Liang, W., et al. (2023). GPT Detectors Are Biased Against Non-Native English Writers. Patterns.
• University of San Diego Legal Research Center. (2025). Generative AI Detection Tools.
• Juzek, T., & Ward, M. (2025). Why Does ChatGPT "Delve" So Much? ACL Findings.