Gemini Fingerprint

Technical description

F3 scores a document on two components and normalises the sum to the 0 to 5 scale. The lexical component sums weight × occurrences over a per-language dictionary of Gemini-associated phrases weighted 1 to 5 and concentrated on synthesis and structuring markers (crucial, essential at 2; in essence, at its core, to summarize, here are some at 3; here's a breakdown, key takeaways, executive summary, let's break this down at 4; that's a great question at 5); a matched phrase that also belongs to the shared cross-model generic set is multiplied by 0.25 first. The structural component adds fixed and graded points for a tripartite intro-bullets-conclusion shape (+3), a Markdown table (+2), mixed heading styles where both ## headings and bold-as-heading lines appear (+2), an invented-DOI pattern (+4), em-dash density min(2.0 + (r − 3) × 0.3, 4.0) at a rate r of at least 3 per 1000 words, and heavy bold-header use scored as min(n × 0.5, 3.0) for n of 4 or more bold section headers. The raw total R maps to the reported score as min(5.0, R / 15 × 5). Twelve language dictionaries are available; the document language selects one.

How it works

The implementation is deterministic and runs at Layer 1 over compiled regular expressions.

Lexical scoring. The active per-language dictionary maps each phrase to an integer weight from 1 to 5, rising with how distinctive the phrase is of Gemini output. The weight concentrates on the model's synthesising, presentation-oriented register: the conversational openers that's a great question (5) and absolutely (4), and the structuring and summarising markers here's a breakdown, key takeaways, executive summary, let's break this down (4) alongside in essence, at its core, to summarize (3). Each case-insensitive match adds its weight times its occurrence count, and a phrase held in the shared cross-model set has its weight multiplied by 0.25 first, leaving the Gemini-specific residue. Matches of weight 4 or more are reported at warning severity, lighter matches at informational severity.

Structural scoring. Six signatures contribute. A tripartite layout, a short introduction of at most two sentences, a middle containing a bullet list, and a closing prose paragraph, adds 3 and captures the model's default report template. A Markdown table adds 2, reflecting the habit of returning tabular summaries. Mixed heading styles, where the document uses both ## headings and standalone bold lines as headings, add 2. A reference matching the invented-DOI pattern adds 4, since fabricated identifiers are a recurring Gemini citation artifact. Heavy bold-header use is scored on a curve: four or more standalone bold section headers add min(n × 0.5, 3.0), so four headers contribute 2 and the contribution rises by 0.5 per additional header to a ceiling of 3.

Em-dash density. As in the other fingerprints, em-dash density is scored on a graded curve. The em-dash count divided by the word count and multiplied by 1000 gives a rate r; once r reaches 3 per 1000 words (with at least two em-dashes present), the contribution is min(2.0 + (r − 3) × 0.3, 4.0).

Aggregation. The lexical sum and the structural contributions are added into a raw score R, reported as min(5.0, R / 15 × 5). The raw score and the detected phrases with their counts and effective weights are returned in the metadata.

Score thresholds

Score	Meaning
0 to 1	Sparse synthesis vocabulary, continuous prose, no tables or bold-header layout.
2 to 3	A concentration of structuring markers, or one layout signature such as a table, mixed headings, or a handful of bold headers.
4 to 5	The synthesis register and several layout signatures co-occur: tripartite template, tables, dense bold headers, and high em-dash density, or an invented DOI is present. Strongly consistent with unedited Gemini output.

Why this matters

Gemini favours a report register over an essay register. Tuned to return organised, presentation-ready answers, it opens by restating the task, breaks the body into bullets and tables, labels sections with bold headers, and closes with a summary, a layout that reads as a briefing rather than as argued prose. In a manuscript those habits surface as a tripartite intro-bullets-conclusion shape, tabular summaries, and a proliferation of short bold section labels, each of which is uncommon in continuous academic writing and so carries signal. The synthesising vocabulary, here's a breakdown, key takeaways, executive summary, marks the same orientation at the lexical level. The invented-DOI signature is the sharpest of the set: the model has a documented tendency to manufacture plausible-looking identifiers, and a reference whose DOI matches the fabricated-identifier pattern is strong evidence of generated citation rather than transcription. F3 isolates the Gemini-specific part of this profile by discounting the vocabulary and formatting it shares with other assistants through the cross-model generic discount, and it hands the fabricated-citation question to the dedicated verification indicators (G1 and L3) for confirmation.

Limitations

The vocabulary and density thresholds were calibrated against 2024-2025 Gemini output and require periodic recalibration as the model and its users adapt. The lexical signal yields to paraphrase or to replacement of the highest-weight markers, and the layout signatures vanish when Markdown is stripped during conversion to plain text, so a document that began as a structured report but was flattened will register only the residual vocabulary. In the other direction, genuinely report-style human writing, executive summaries, technical briefings, slide decks rendered as text, produces tables, bold headers and tripartite structure for reasons unrelated to any model, which is why the indicator reports resemblance to a profile and weighs each signature rather than treating any one as decisive. The invented-DOI check flags the pattern of a fabricated identifier and is a screen rather than a verification; whether the cited paper exists is settled by the citation indicators. The dictionaries are most developed for English and thinner across the other eleven languages, while the structural and em-dash checks are language-independent.

Theoretical background

F3 combines lexical fingerprinting with layout analysis. The lexical weights follow the excess-vocabulary logic shared across the F-series, restricted to the synthesising, breakdown-oriented register characteristic of Gemini rather than the assistant vocabulary common to all systems, which the cross-model generic discount subtracts. The structural strand draws on the observation that feedback-tuned models trained to be helpful converge on presentation conventions, and that the conventions differ by model: where Claude nests headings and hedges, Gemini tabulates and labels. The fabricated-identifier strand connects to the broader literature on citation hallucination, in which generated references frequently carry well-formed but non-existent identifiers; F3 contributes the surface-pattern detection of that artifact, leaving existence checking to the Layer 3 verification stack.

References

1. Kobak D, González-Márquez R, Horvát EÁ, Lause J. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances. 2025. https://arxiv.org/abs/2406.07016
2. Liang W, Zhang Y, Wu Z, Lepp H, Ji W, Zhao X, Cao H, Liu S, He S, Huang Z, Yang D, Potts C, Manning CD, Zou J. Quantifying large language model usage in scientific papers. Nature Human Behaviour. 2025. DOI: 10.1038/s41562-025-02273-8 https://www.nature.com/articles/s41562-025-02273-8
3. Thelwall M, Kousha K. Have LLM-associated terms increased in article full texts in all fields? arXiv preprint arXiv:2604.07565. 2026. https://arxiv.org/abs/2604.07565
4. Ansari MS. Compound deception in elite peer review: a failure mode taxonomy of 100 fabricated citations at NeurIPS 2025. arXiv preprint arXiv:2602.05930. 2026. https://arxiv.org/abs/2602.05930