AI writing tells: 13,000 phrases flagged in study, Claude and Astra diverge

A new study by marketing company Graphite has identified approximately 13,000 words and phrases that could signal AI-generated text. The findings reveal distinct patterns in outputs from Anthropic’s Claude Opus 5.5 and OpenAI’s Astra models, while suggesting that some AI systems are evolving to avoid detection.

Study identifies thousands of AI writing tells

Marketing company Graphite has released findings from a study identifying approximately 13,000 words and phrases that may signal AI-generated text. According to Gizmodo, the research highlights patterns in outputs from models like Anthropic’s Claude Opus 5.5 and OpenAI’s Astra, which could help distinguish AI-generated content from human writing.

Claude Opus 5.5’s detectable patterns

The study flags specific phrases as strong indicators of AI involvement in text generated by Claude Opus 5.5. Gizmodo reports that the phrase ‘this matters’ appears 116 times more frequently in AI-generated content than in human writing. Other patterns include ‘looking ahead the,’ which is 40 times more common, and ‘why [blank] matters,’ which appears 92 times more often. The word ‘dependable’ is used 26 times as frequently as in human writing.

However, the study also suggests that Claude Opus 5.5 is evolving to reduce its detectable tells. For example, the model now uses em dashes 99% less often than its predecessor, Opus 5. Researchers note that Claude-generated text is increasingly resembling human writing over time.

OpenAI’s Astra model exhibits distinct tells

OpenAI’s Astra model displays its own set of identifiable patterns, according to the study. Gizmodo reports that the phrase ‘does not establish’ is 275 times more common in Astra’s outputs compared to Claude’s. Unlike Claude Opus 5.5, which is trending toward more human-like writing, Astra’s outputs appear to be moving in the opposite direction, with less emphasis on mimicking human cadences.

The study also notes that Opus 5.5 qualifies claims less often than Astra and is far more likely to use superlatives. These differences highlight how AI models are developing distinct writing styles, influenced by their training data and design priorities.

Context: AI’s evolving role in content creation

AI models like Anthropic’s Claude and OpenAI’s Astra are increasingly used for generating written content, from marketing copy to technical documentation. While these tools offer efficiency and scalability, concerns about authenticity, transparency, and compliance with regulations—such as copyright and defamation laws—have grown alongside their adoption.

Efforts to watermark or trace AI-generated content, such as Google DeepMind’s SynthID Bio for proteins, reflect broader industry attempts to address these challenges. However, as models evolve, so do the methods for detecting their output, creating a dynamic interplay between creation and identification.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

This study underscores a familiar tension in AI development: the race to mimic human writing while avoiding detection. For founders and operators, the takeaway is twofold. First, if you’re using AI to generate content—whether for marketing, customer communications, or product descriptions—you’re walking a tightrope. Tools like those developed by Graphite won’t just fade away; they’ll get sharper, and audiences (or regulators) may start penalizing obvious AI tells. Second, the divergence between models like Claude Opus 5.5 and OpenAI’s Astra suggests that not all AI providers prioritize the same goals. Some are doubling down on ‘human-like’ output, while others seem content with utility over mimicry. That’s a critical decision point for businesses integrating AI: do you want a tool that blends in or one that gets the job done, even if it stands out?

Key takeaways

• Graphite’s study identified ~13,000 words and phrases linked to AI-generated text.
• Phrases like ‘this matters’ and ‘looking ahead the’ are far more common in AI outputs than human writing.
• Claude Opus 5.5 is reducing detectable patterns, including em dash usage, while Astra’s writing is becoming less human-like.
• AI models are evolving differently—some prioritize stealth, others prioritize function.

FAQ

Why do AI models like Claude Opus 5.5 and Astra have detectable writing patterns?

AI models develop patterns based on their training data, fine-tuning, and design objectives. Some models, like Claude Opus 5.5, are optimized to mimic human writing closely, which can lead to overuse of certain phrases or structures. Others, like Astra, may prioritize task completion over stylistic mimicry, resulting in distinct but less human-like patterns.

How might businesses be affected by AI-generated content detection tools?

Businesses using AI for content creation could face reputational, legal, or regulatory risks if their output is flagged as AI-generated. Detection tools could be used by platforms, regulators, or audiences to enforce transparency rules or penalize low-quality AI content. This makes it critical for companies to review and refine AI outputs to align with their brand voice and compliance requirements.

Are AI models like Claude Opus 5.5 intentionally reducing detectable patterns?

While the study suggests Claude Opus 5.5 is shedding some of its tells—such as em dash usage—it’s unclear whether this is intentional. The shift could reflect broader improvements in the model’s training or a response to user feedback about readability and naturalness. However, it demonstrates that AI writing styles are not static and can evolve over time.

Related on Lazyfounder

• Anthropic
• OpenAI
• More ai news
• Google DeepMind Launches SynthID Bio to Watermark AI-Designed Proteins