QuillBot: AI Paraphrasing & Detection

Updated 7 July 2026

QuillBot is a mainstream AI-assisted tool that rephrases text and assigns an AI-percentage score to indicate the likelihood of machine generation.
It functions both as a paraphraser and an AI-content detector, demonstrating high detection scores on original texts but reduced accuracy on humanized paraphrases.
The tool is central to adversarial research and AI writing studies, highlighting the need for careful calibration across languages, domains, and evaluation settings.

QuillBot is a mainstream paraphrasing system that, in current research, is treated both as an AI-assisted writing service and as an AI content detector. One study describes it as “known as a paraphrase tool” that also provides “AI content detection services and features,” together with grammar checks, citation generation, and summarizations; the same work uses the online QuillBot service, as of March 2025, both to paraphrase DeepSeek outputs and to score text with an AI-text percentage between 0 and 100 (Alshammari et al., 23 Jul 2025).

1. Functional scope and roles

Research literature assigns QuillBot two distinct operational roles. First, it functions as a paraphrasing system: text is pasted into the service, a rewrite mode is selected, and the system returns reworded output. Second, it functions as an AI-text detector: text is submitted to its AI Content Detector, which returns an AI-percentage score used as a proxy for the probability that the text is machine-generated. In the DeepSeek evaluation study, QuillBot is explicitly central to both sides of the experiment, because it is used to generate adversarial paraphrases and is also evaluated as a detector against those same classes of attack (Alshammari et al., 23 Jul 2025).

This dual role distinguishes QuillBot from tools that are treated only as detectors. In the same DeepSeek study, AI Text Classifier, Content Detector AI, GPT-2 detector, GPTZero, and Copyleaks are treated as pure detectors, whereas QuillBot is treated as both a transformation service and a detection system. This makes QuillBot unusually important in detector-robustness work: it is not merely one system among several, but part of the adversarial pipeline itself (Alshammari et al., 23 Jul 2025).

In broader comparative evaluation, QuillBot is also studied purely as a detector. A 2026 comparative framework benchmarks QuillBot against supervised neural models and other online tools, including ZeroGPT, GPTZero, Originality.AI, Sapling, IsGen, Rephrase, and Writer. In that framework, the detector is treated as a black box whose thresholds and calibration are not transparent, and its outputs are converted to binary Human/GenAI decisions for evaluation (Buttaro et al., 19 Mar 2026).

2. Experimental uses as paraphraser and detector

In the DeepSeek study, the total testbed consists of 294 answer samples built around 49 question–answer pairs. The corpus includes 49 human-written answers, 49 DeepSeek-original answers, 49 DeepSeek-paraphrased answers, and 147 DeepThink-related answers divided into original, standard-paraphrased, and humanize-mode conditions. QuillBot is used directly to paraphrase 49 DeepSeek responses, and it is also evaluated as one of the six detectors on the relevant conditions (Alshammari et al., 23 Jul 2025).

Sample type	Count	QuillBot’s role
Human-written answers	49	Detector
DeepSeek-original answers	49	Detector
DeepSeek-paraphrased answers	49	Paraphraser and detector
DeepThink-original answers	49	Detector
DeepThink-paraphrased (Standard)	49	Detector against paraphrasing attack
DeepThink-paraphrased (Humanize mode)	49	Detector against humanization attack

The same paper notes two implementation constraints that matter for interpretation. The detector is used with its default behavior, no custom thresholds or modes are reported, and QuillBot requires at least 80 words for detection. The evaluation is also explicitly time-bound: the reported results were recorded in March 2025, and the authors caution that tool behavior may change over time (Alshammari et al., 23 Jul 2025).

A separate computing-education study uses QuillBot only as an attack generator. There, 10 English ChatGPT submissions are paraphrased through QuillBot, producing a “ChatGPT+QuillBot” subset used for resilience testing. No human-written text is paraphrased, no Spanish text is paraphrased, and the paper does not report multiple QuillBot modes or parameter sweeps; QuillBot is treated as a black-box paraphraser with a single configuration (Orenstrakh et al., 2023).

3. Detection metrics and reported performance

The DeepSeek evaluation defines an “average AI detection score” rather than standard confusion-matrix accuracy: $\text{Average AI Detection} = \frac{\sum X_i}{N},$ where $X_i$ is the AI-percentage returned for sample $i$ , and $N$ is the number of samples. For recall analysis, the paper applies a 50% decision threshold: if QuillBot’s AI percentage is greater than 50%, the sample is treated as predicted AI (Alshammari et al., 23 Jul 2025).

Across that protocol, QuillBot shows near-perfect recognition on original DeepSeek text and very strong performance on standard paraphrases, but substantial degradation under humanization. A later comparative study, using different datasets and standard accuracy/recall reporting, presents a more variable picture across languages and domains. The contrast is important because the metrics are not identical across the two studies (Alshammari et al., 23 Jul 2025).

Setting	Reported QuillBot result	Notes
Human-written text, DeepSeek study	100% human detection	No false positives reported
DeepSeek-original	98.4% average AI detection; 100.00% AI recall	49 TP, 0 FN
DeepSeek-paraphrased	95.49% average AI detection; 97.96% AI recall	48 TP, 1 FN
DeepThink-original	95.4% average AI detection	Strong on unmodified text
DeepThink-paraphrased (Standard)	≈84.2% average AI detection	Reported with a typographical artifact as 84.1.9%
DeepThink-paraphrased (Humanize mode)	58% average AI detection	Major degradation
dtEN (balanced English)	65.0% accuracy; 100.0% Human; 34.4% GenAI	Strong Human bias
dtITA (Italian, GenAI-only)	76.7% accuracy = 76.7% GenAI recall	23.3% labeled Human
ART{MH} (Italian thematic)	100.0% accuracy; 100.0% Human; 100.0% GenAI	Perfect on this small test

Within the DeepSeek study, QuillBot is in the top tier on original DeepSeek text together with Copyleaks and GPTZero, and it is best-in-class on DeepSeek-paraphrased text at 95.49%, slightly ahead of Copyleaks at 93.54% and GPTZero at 92.61%. On DeepThink-paraphrased (Humanize mode), however, Copyleaks is stronger at 71%, while QuillBot falls to 58% and GPTZero to 52% (Alshammari et al., 23 Jul 2025).

In the comparative neural-detector framework, QuillBot behaves conservatively on the balanced English benchmark: it achieves 100.0% Human recall but only 34.4% GenAI recall on dtEN. It performs better on the Italian single-class GenAI benchmark at 76.7%, and it is perfect on the small thematic Italian ART{MH} dataset. This suggests that QuillBot’s observed reliability is strongly dataset-dependent and may vary with language, domain, and calibration regime (Buttaro et al., 19 Mar 2026).

4. Adversarial paraphrasing and evasion research

Recent work distinguishes simple paraphrasing from detector-aware adversarial paraphrasing. In the formalization of “Adversarial Paraphrasing,” the paraphraser is modeled as $\mathcal{P} : \mathcal{X} \to \mathcal{X}$ and the detector as $\mathcal{D} : \mathcal{X} \to [0,1]$ , where lower detector scores mean “more human-like.” Instead of sampling the next token only from the paraphraser’s distribution, the framework filters candidates with top- $p$ and top- $k$ , scores each candidate continuation with the detector, and selects the token with minimum detector score. The paper explicitly frames this as “QuillBot + a detector in the loop,” in contrast to ordinary paraphrasing that does not optimize against detector feedback (Cheng et al., 8 Jun 2025).

That distinction matters empirically. The same paper states that simple paraphrasing is analogous to what QuillBot does: a single round of paraphrasing without access to any detection score or iterative adversarial optimization. Under that baseline, paraphrasing can even make detection easier for some modern detectors: T@1%F increases by 8.57% on RADAR and 15.03% on Fast-DetectGPT. By contrast, adversarial paraphrasing guided by OpenAI-RoBERTa-Large reduces T@1%F by 64.49% on RADAR and 98.96% on Fast-DetectGPT, with an average T@1%F reduction of 87.88% across eight detectors (Cheng et al., 8 Jun 2025).

A separate education-focused detector study shows the effect of QuillBot as a practical evasion tool. On 10 English ChatGPT submissions, threshold-based accuracy after QuillBot remains 100% for GLTR, but falls from 100% to 60% for GPT-2 Detector, from 100% to 50% for CopyLeaks, from 100% to 40% for CheckForAI, from 100% to 40% for Originality.AI, from 90% to 30% for GPTKit, from 60% to 20% for AI Text Classifier, and from 70% to 20% for GPTZero. The same study reports that, when a paraphraser such as QuillBot is utilized, average accuracy on ChatGPT-generated data falls to approximately 49.17% (Orenstrakh et al., 2023).

Taken together, these results do not support a single uniform conclusion about paraphrasing. Ordinary paraphrasing can be highly damaging to many deployed detectors, but not necessarily to all of them; detector-aware humanization is much stronger than simple paraphrasing; and QuillBot occupies the baseline end of this spectrum unless it is augmented with detector feedback. This suggests that robustness depends on the detector’s training distribution, the evaluation metric, and whether the paraphrasing process is detector-agnostic or detector-aware.

5. QuillBot in AI-assisted writing research

User-study research on AI-assisted writing tools provides a useful interactional context for QuillBot. The literature distinguishes a push paradigm, in which the system proactively offers suggestions, from a pull paradigm, in which the user explicitly requests text generation. The same work states that QuillBot’s current design is much closer to pull: the user selects text and requests a paraphrase or continuation, and the system returns multiple sentence suggestions or variants that can be accepted or modified. User acceptance is formalized as

$D = \frac{A}{N} \times 100,$

where $A$ is the number of accepted suggestions and $X_i$ 0 is the number of suggestions or generation requests; efficiency is measured as

$X_i$ 1

where $X_i$ 2 is word count and $X_i$ 3 is time in minutes (Pereira et al., 2023).

In that study, pull-based assistance yields higher acceptance than push-based assistance on both familiar and unfamiliar tasks: 54.03% versus 35.20% on the familiar task, and 62.50% versus 40.90% on the unfamiliar task. Efficiency is nearly identical on the familiar task, with 30.48 WPM for pull and 31.55 WPM for push, but diverges sharply on the unfamiliar task, where pull reaches 54.93 WPM and push 24.74 WPM. Participants also report that AI assistance helped them diversify ideas, keep writing clear and concise more quickly, enjoy collaboration with the tool, and retain a sense of ownership over the final text (Pereira et al., 2023).

This interactional evidence does not describe QuillBot directly as an evaluated product, but it is directly relevant to QuillBot-like systems. It suggests that QuillBot’s paraphrasing workflow aligns more closely with user-initiated generation for idea expansion, rewriting, and overcoming writer’s block than with system-initiated micro-completion. It also suggests that low-friction interfaces and editable suggestions are central to perceived usefulness and authorship.

The literature is consistent in treating QuillBot’s outputs and scores as useful but non-definitive. The DeepSeek study states that QuillBot and Copyleaks showed high detection scores and no false positives on the human set, yet also emphasizes that not all current online detection systems are 100% reliable and that performance declines notably on DeepThink-paraphrased text, especially under humanization. The authors therefore recommend selecting detectors based on users’ requirements and not relying solely on QuillBot in high-stakes adversarial settings (Alshammari et al., 23 Jul 2025).

Other evaluations reinforce this caution. In the comparative neural framework, commercial tools including QuillBot show “substantial variability and limited reliability” across datasets, languages, and domains; in the education study, detector behavior is also degraded by paraphrasing and by non-English text. Methodologically, these conclusions are qualified by modest dataset sizes, black-box calibration, language limitations, and time-bounded tool versions. In the DeepSeek study, only 49 base questions are used; in the neural comparison, test sets contain 60 samples per dataset; and the recorded QuillBot behavior corresponds to specific evaluation dates rather than a stable, versioned API contract (Buttaro et al., 19 Mar 2026).

Research on paraphrasing systems of the same broad class provides two technical reference points for future interpretation. One is the unsupervised paraphrasing pipeline based on task-adaptation, self-supervision, and Dynamic Blocking, where Dynamic Blocking prevents the model, after emitting a source token, from outputting the subsequent source token on the next generation step. That work reports state-of-the-art performance on QQP and ParaNMT and robustness to domain shift, and it demonstrates transfer to other languages without additional finetuning (Niu et al., 2020). A plausible implication is that QuillBot-like systems can be analyzed not only as interfaces but also as decoding systems that balance semantic preservation and surface-form dissimilarity.

A second reference point is prompt optimization for QuillBot-like systems. “PromptQuine” treats prompts as token sequences that can be pruned into effective but often unintelligible subsequences, and reports that such “gibberish” prompts match or surpass state-of-the-art automatic prompt optimization techniques across classification, multiple-choice question answering, generation, math reasoning, and jailbreaking tasks (Wang et al., 22 Jun 2025). This does not identify QuillBot’s internal prompting strategy, which remains undisclosed in the cited studies. It does, however, indicate that future QuillBot-like systems may depend increasingly on latent prompt search, detector-aware decoding, and other black-box optimization procedures whose behavior is only partially legible from the user interface.

In research terms, QuillBot is therefore best understood not as a single-purpose paraphraser, but as a composite object in current AI-text studies: a rewrite service, a detector, an adversarial instrument for probing detector robustness, and an instance of the broader design space of pull-based AI-assisted writing systems.