The Consensus, a publication focused on software infrastructure, has introduced a new mechanism to identify and manage content that exhibits 'AI tells.' This initiative stems from a shift in reader preferences; according to a survey by Cynthia Dunlop, tech readers often prefer human-written content—even if it contains minor flaws—over polished but generic AI-generated prose. To address this, the site's founder, Phil Eaton, developed a detector to flag articles that strongly resemble the stylistic patterns of large language models (LLMs) like Claude or ChatGPT.
The initiative stems from a shift in reader preferences; according to a survey by Cynthia Dunlop, tech readers often prefer human-written content.
Unlike high-end commercial detectors that use machine learning, this internal tool relies on a set of static rules. These rules were established by comparing pre-2023 writing styles against post-2023 writing, while accounting for the natural drift in human writing styles over time. The detector does not provide a definitive verdict on whether a piece was written by an AI; instead, it identifies articles in the 99th percentile of 'AI tells.' These flagged articles are automatically excluded from high-visibility areas of the site, such as the 'At a Glance' homepage section, social media promotions, and the weekly newsletter.
While the system can be bypassed by training AI models to mimic specific human styles, the primary goal is to promote thoughtful, human-authored technical writing. The Consensus maintains that using AI to communicate is not inherently negative, but they aim to reduce the prevalence of 'stock' AI prose that can make content feel repetitive and unoriginal. For readers who prefer to see all content regardless of these flags, the site still provides options to filter the feed manually or adjust the 'aitells' threshold in search queries. This approach highlights a growing tension in technical publishing between the efficiency of AI generation and the perceived value of human-authored nuance and character.
One significant limitation of this rule-based approach is its potential for false positives, as the system has occasionally flagged human-authored articles. Furthermore, the move highlights a broader challenge in content curation: as AI models become more sophisticated at mimicking human nuances, the distinction between 'good' AI writing and 'bad' human writing may become increasingly difficult to define through static rules alone. The Consensus's strategy reflects a move toward prioritizing the unique 'character' of human writing as a primary metric for editorial quality in the software infrastructure space.
Ultimately, the implementation of these 'AI tells' serves as a gatekeeping mechanism to ensure that the most promoted content maintains a human touch. By creating a friction point for generic AI output, the platform encourages a standard of writing that avoids the repetitive templates common in current LLM outputs. While the technology for generating content is becoming more accessible, the editorial focus remains on the human element of technical communication, where specific trade-offs between volume, speed, and authenticity are being actively managed by publishers.
The mechanism works by analyzing linguistic patterns that have emerged since the widespread adoption of generative AI. By identifying these 'tells'—such as specific structural repetitions or overly neutral phrasing—the site can programmatically surface content that feels more authentic to its audience. This creates a tiered system where human-authored content is given a 'sparkle' of promotion, while AI-heavy content is relegated to the background, allowing the publication to maintain its editorial standards without completely banning the use of AI tools by its contributors.