Participial tack-on padding
- Id
- participial-tackon
- Status
- Active
- Severity
- high
- Detection
- statistical
- Evidence grade
- peer-reviewed
- Languages
- en
- Added
- 2026-08-15
- Updated
- 2026-08-17
Currently signals low-effort writing.
What it is
A sentence finishes. Then a comma arrives, and an -ing clause says the same thing again in the register of analysis. Reinhart and colleagues counted present participial clauses in GPT-4o output at 5.3 times the human rate, with a Cohen's d of 1.38, the largest effect in their grammatical table. That figure sits in the paper body rather than the abstract, and it compares two corpora. It licenses nothing about one sentence.
Why it reads as machine-made
The clause has the shape of a conclusion and none of the weight. It hangs off the main clause, adds no fact, and lets the sentence end on a note of consequence the writer never paid for. Wikipedia files the same habit under superficial analysis. Two of them in a paragraph and the prose starts reading like captions under photographs.
Specimens
The council approved the new cycle lane, marking a significant step forward for the city's transport plan. The decision was welcomed by residents, reflecting a growing appetite for safer streets. The scheme was described as a priority, underlining a commitment that has been building for some time.
The council approved the new cycle lane in March. Building it closes the northbound lane outside the market hall for eleven weeks from September, and the bus route moves two streets east while it does.
Three sentences, three trailing clauses, and every one of them restates the clause it hangs off. Delete all three and nothing is lost. The repair says what the approval costs. Figures in this repair are invented for the specimen.
How it is detected
- Metric
- trailing-participial-clauses-per-1000-words
- Threshold
- unset
- Direction
- above
- Threshold basis
- No threshold ships. The published number is 5.3 times the human rate with a Cohen's d of 1.38, and it is a ratio between corpora printed in a paper body, not a per-document cutoff. Nobody has published a human distribution for this rate, so any line drawn here would be invented. Measure the draft, compare it against the same writer's earlier work, and treat a jump as a reason to reread. That is a FeedSquad review trigger and not a finding.
Who writes this way legitimately
Academic and technical writers use trailing -ing clauses that carry a real result, and the clause is then load-bearing: a valve failed at 40 bar, flooding the pump room. The room flooded, which the main clause never said. The test is whether the clause introduces a noun, a figure or an event absent from the sentence it hangs off. Delete it; if nothing is lost, it was padding, and if a fact is lost, it was a sentence in the wrong clothes.
Model attribution
Reinhart and colleagues measured GPT-4o and Llama 3 70B Instruct. The instruction-tuned models diverged from human prose further than the base models did, which is the opposite of what instruction tuning is usually assumed to do. No study assigns the habit to one vendor.
Platform notes
- wikipedia
- Listed under superficial analysis on the signs-of-AI-writing page, where the shortcut WP:SUPERFICIAL points. The page treats the trailing clause as text that performs analysis without doing any.
Status history
| Date | Status | Rationale |
|---|---|---|
| 2026-08-15 | Active | Reinhart measured 2024-era models and no later study reports the gap closing. The construction is still carried on Wikipedia's signs-of-AI-writing page with no decay note, and vale-ai-tells ships 68 surface forms of it. |
Sources
- 01Reinhart et al., Do LLMs write like humans? PNAS 122(8) (arXiv:2410.16107)peer-reviewedaccessed 2026-08-14
- 02Wikipedia: Signs of AI writingcommunityaccessed 2026-08-14
- 03vale-ai-tells, 111 machine-checkable rulescommunityaccessed 2026-08-14
- 04tropes.fyi pattern directorycommunityaccessed 2026-08-14
CC BY 4.0 / The AI Tells Index, feedsquad.com/ai-tells