Identifying "AI Tells"

---

Sponsor Spotlight: 🚀 ClickHouse Managed Postgres provides up to 10x performance boosts for I/O-intensive tasks utilizing local NVMe. It also features seamless, native integration with ClickHouse to enable real-time analytical capabilities. [Explore more ↗]

---

The Mission of The Consensus

The primary objective of The Consensus is to champion high-quality technical documentation and writing within the realm of software infrastructure. To achieve this, we utilize several distribution channels:

• At a Glance: A daily curated highlight on our homepage.
• Weekly Newsletter: A free deep-dive into the best of the week.
• Third-Party Feed: A public, filtered stream of quality technical content.

However, a survey conducted by Cynthia Dunlop regarding the preferences of technical readers revealed a stark truth: people generally prefer "good" AI writing prefer even "bad" human writing over polished AI-generated content.

"It infuriated me to click on a random cool-sounding article in our feed just for the actual writing to grate." — Phil Eaton

This is particularly jarring when the content is produced by organizations that publicly advocate for the "responsible" and "temperate" use of AI in their communications.

The Solution: An AI Tells Detector

While using a premium tool like Pangram for every single entry in the feed is financially unsustainable, I have spent the last month developing an internal AI tells detector.

How it Works

Unlike most detectors, this is not a machine learning model. Instead, it relies on a set of static heuristic rules. These rules were developed by analyzing the linguistic drift between writing produced before 2023 and writing produced after 2023.

# Conceptual logic of the AI Tells Detector
def check_ai_tells(text):
    score = 0
    ai_markers = ["delve", "tapestry", "in the rapidly evolving landscape", "comprehensive guide"]
    for marker in ai_markers:
        if marker in text.lower():
            score += 1
    return score # Higher score = more "AI-like"

Important Nuances

It is critical to understand that this tool does not definitively prove an article was written by an AI. It simply flags content that resembles AI writing.

• Cost: Very low.
• Accuracy: Imperfect (it has even flagged my own writing).
• Purpose: A signal, not a verdict.

The Workflow

The following diagram illustrates how the detector influences the visibility of an article:

Impact on Promotion

We now apply a strict filter for our promoted content. If an article falls within the $\text{Top } 1\%$ (the 99th percentile) of "AI tells," it faces the following restrictions:

• No placement in the "At a Glance" homepage section.
• No distribution via social media channels.
• No "sparkle" icon (which denotes a highly interesting piece).
• No inclusion in the weekly newsletter.

The Philosophy of Voice

The core frustration stems from the fact that most AI content sounds identical—essentially "stock" Claude or "stock" ChatGPT. If companies invested the effort to define their own unique character, rules, and preferences for their AI, the result would be far more palatable.

While we believe in the right of every company to choose its communication style, we reserve our active promotion for work that is thoughtfully crafted by a human.

Note: We aren't banning AI content entirely. You can still find these articles in the feed, or specifically filter for them using the sparkle filter or the search query aitells 90.

---

About the Author: Phil Eaton

Phil is a veteran of the industry with a background including:

1. Postgres product work at EnterpriseDB.
2. Co-founder and marketing lead at TigerBeetle.
3. Engineering Manager at Oracle.

He currently manages the Software Internals Discord and the Software Internals Email Book Club.

---

Community & Opportunities

Developer Spotlight: 🌟 Deepthi Sigireddi — A seasoned developer with over 25 years of industry experience. [Read the interview].

💼 Open Roles

💰 Funding & Acquisitions

📅 Upcoming Events

• ScyllaDB P99 Conf (Remote) — October 21, 2026 (Systems Performance)
• Bavarian Database Day (Munich) — October 26, 2026 (Databases)
• DistrictCon Year 2: DisCo Together (Washington, DC) — February 6, 2027

---

Note: JavaScript is required to fully experience The Consensus.