ThesisDraft vs. Sapling: A Free Detector Tuned for a Different Kind of Text

What Sapling and ThesisDraft Are

Sapling is an AI writing assistant for customer-facing teams — sales replies, support messages, autocomplete inside a helpdesk. Alongside that it publishes a free AI detector that returns an overall score and per-sentence highlighting, which is how most people outside its core market encounter the company.

ThesisDraft’s AI check exists for one situation: a student about to submit a thesis, in English or German, who wants to know what a detector will see. Free for 1,500 words with no account, €4.99 once for the full document, and a published false-positive rate beside every verdict.

Side by Side

Register Is Not a Detail

A detector learns what machine writing looks like relative to the human writing it was shown. Change the human side and the whole calibration moves.

Business and support writing is short, varied, conversational and personal. Thesis writing is long, formal, deliberately uniform, heavy with conventional transitions, and repeats its terminology on purpose because inventing synonyms for a technical term is bad practice. Every one of those properties reads as machine-like to a detector calibrated on informal text.

That is not a theoretical objection. Published research has found detectors falsely flagging a majority of non-native English writers, and the mechanism is the same. It is also how we broke our own detector once: an early version scored an AUC of 0.184 — worse than a coin flip — against 236 pre-2022 arXiv abstracts, because a family of lexical-richness features scored dense human academic prose as machine-like. Those features now carry zero weight, and we only found the problem because we tested against academic text specifically.

The False-Positive Rate

Measured 9 August 2026, on human academic documents published before ChatGPT existed:

• English over 400 words — about 3 false positives in 100.
• English under 400 words — about 2 in 100.
• German over 400 words — about 4 in 100.
• German under 400 words — no high-probability verdict at all.

Published with its limits: in-sample long-form results, small AI corpora, and a German corpus of expository Wikipedia prose. The scripts are in the repository and listed on the accuracy page.

Per-Sentence Scores, and What They Are Good For

Sapling highlights sentence by sentence, and so do we. It is genuinely the most useful view a detector can offer — but it is easy to over-read.

A single sentence is short, and short text is where every detector is least reliable. Our own measurements make that concrete: at under 400 words in German, no threshold strong enough to catch AI costs fewer than roughly 7 false accusations in 100, which is why we refuse to issue a strong verdict there at all. A sentence is far shorter than 400 words.

So sentence highlighting should be read as where to look, not what is true. In practice the useful pattern is a run of adjacent flagged sentences with uniform length and stock transitions — that is a paragraph worth rereading. One flagged sentence in a page is noise.

What Sapling Does Better

• Instant, no-cost access for a quick paste, with no word budget to think about.
• Real-time writing assistance inside business tools — a different product, and a good one for its market.
• An API and integrations for teams embedding detection in a workflow.

The Verdict

Use Sapling for a free look at a short business or general-purpose passage.

Use our AI check when the text is a thesis: measured on academic prose, calibrated separately for German, with the error rate stated next to the verdict and a report you can keep. The first 1,500 words are free, no account.

Run the free AI check · See our measured error rates