AI Tells: Gemini 4 Argon

Introduction

Google says Gemini 4 Argon has “stronger writing capabilities,” “answers questions more directly,” and “skips clichés.” But in our analysis of web articles, it retains many well-known AI tells.

In our AI tells study and Opus 5.5 update, we found thousands of AI tells that differ between model families and versions. In this paper, we extend our analysis to Gemini 4 Argon and compare its writing with GPT-6 Astra, Claude Opus 5.5, and Gemini 3.1 Pro.

We use the same methods as in our original study, comparing AI and human web articles on the same topics. We identify words, phrases, and frames (word patterns with underscores representing gaps of up to three words) that occur at least twice as often in AI writing as in human writing, after normalizing for the amount of text and applying frequency thresholds. We also compare overall word distributions and stylistic features, such as em-dash use.

Argon Has Thousands of Tells

We find 3,248 word and phrase tells in Argon, compared with 3,677 in Astra and 2,551 in Opus 5.5.

We find several recurring themes in Argon’s writing. The first two tables compare length-normalized usage rates between models, rather than with human writing. Large ratios can reflect very low usage in the comparison model.

“Furthermore” appears in 45% of Argon’s articles, compared with none of Astra’s and just one of Opus 5.5’s.

Together, these six stock phrases appear about 42 times as often in Argon as in Astra, after normalizing for the amount of text. “Fast-paced world” appears 332 times in Argon’s articles, compared with none in Astra’s and once in Opus 5.5’s. “Half the battle” appears 158 times as often as in Astra and 14 times as often as in Opus 5.5, after normalizing for the amount of text.

Astra qualifies claims more often than Argon with phrases such as “may help,” while Opus 5.5 flags importance more often with “why it matters.”

Argon Uses More Mannered Prose and Em Dashes Than Astra and Opus 5.5

Mannered prose substitutes metaphor and flourish for direct statements. We include Anthropic’s definition verbatim in our scoring prompt. We use Claude Opus 5 to score each article’s mannered prose from 0 to 100, using a subsample of 1,000 matched topics.

Argon’s mean score is 19.48, compared with 7.91 for Astra, 10.57 for Opus 5.5, and 6.65 for human writing. Argon’s score is about 2.9 times the human score, 146% higher than Astra’s, and 84% higher than Opus 5.5’s.

Argon also uses em dashes more often than the other two models. Its rate is about five times Astra’s, although it is still 38% below the rate in pre-ChatGPT human writing. Opus 5.5 almost never uses them.

Argon uses em dashes much more often than Gemini 3.1 Pro. The average strength of the 11 well-known tells we measure is also 13% higher, even though Argon’s overall word distribution is closer to human writing. We measure strength using mean absolute Cohen’s d, which captures differences between AI and human writing rather than frequency alone.

Argon’s Word Choice Is Closer to Human Writing Than Astra’s

These tells describe specific writing habits. We next compare the overall word distribution using Jensen-Shannon divergence. Lower values mean more similar usage. Argon’s word-distribution divergence from human writing is 44% lower than Astra’s, but 17% higher than Opus 5.5’s. The same ordering holds for two- and three-word phrases.

Divergence measures the overall word distribution, while the tell count measures how many words, phrases, and frames are disproportionately common. A model can be closer to human writing overall while still having thousands of tells or using specific AI tells more often. Neither measure evaluates writing quality.

Conclusion

Despite Google’s claim that Argon “skips clichés,” it retains some stock phrases that rarely appear in Astra or Opus 5.5. We find 3,248 word and phrase tells, including stock phrases, emphatic language, and formal transitions. As in our earlier studies, different models have different tells.

Argon uses more mannered prose and em dashes than Astra and Opus 5.5, though its em-dash rate remains below the pre-ChatGPT human rate. Argon’s overall word choice is closer to human writing than Astra’s, but not as close as Opus 5.5’s. These measures capture different aspects of writing and do not give the same ranking of models.

Explore Argon’s Tells

Select a comparison to explore words, phrases, frames, stylistic features, and LLM-summarized themes. The tables rank each side separately. For stylistic features, larger absolute Cohen’s d values indicate stronger differences between the two sets of articles.

Search for a term or frame below to compare its use across all eleven models.

Methodology and Limitations

We use the same methods as in our original study. We analyze n-grams and frames across 9,973 topics with both a human article and an article from each of eleven models. Mannered prose uses the same 1,000-topic subsample as the Opus 5.5 update. Adding Argon changes the shared article set slightly, so results for earlier models may differ from the previous studies.

The limitations of our original study also apply here. Our documents are web articles, and we generate the AI articles using a single, fixed prompt. Tells may differ in other contexts, such as conversations with AI models. The human articles predate ChatGPT, so some differences may reflect changes in writing style over time. Mannered-prose scores depend on the evaluator and prompt, and the evaluator is itself a Claude model. We do not recommend classifying authorship based on individual tells.