Is QuillBot AI Detector Accurate?

Overview

QuillBot’s AI Detector can identify many unedited AI-generated passages in benchmark and controlled tests, but its score is not a dependable verdict about who wrote a particular piece of text. The best quantified evidence is a strong result on the RAID benchmark; smaller real-text tests also show that performance can change sharply after paraphrasing or human editing. No independent study establishes one general accuracy rate for the current detector across ordinary writing, languages, disciplines, text lengths, and mixed human–AI use.

In practical terms, QuillBot can be useful as a preliminary signal or a way to locate passages for closer review. It cannot prove that a person used AI, establish how much help they received, or verify that text is human-written simply because it returns a low AI score. QuillBot itself cautions against using its detector alone for decisions affecting academic standing or a person’s career. (The Technology Behind Our AI Detector | How It Works - Quillbot AI)

What the QuillBot AI Detector does

QuillBot describes its detector as a proprietary classifier that examines word patterns and sentence variety, produces confidence categories and phrase-level explanations, and can label text as human-written, AI-generated, or mixed. The company says its system uses a DeBERTa-v3-large dual encoder and multiple classifiers, and that it has been trained on AI-written and paraphrased text. These are descriptions supplied by the vendor; QuillBot has not made the model itself or a complete independent technical audit available on that page. (The Technology Behind Our AI Detector | How It Works - Quillbot AI)

An AI detector does not inspect a writer’s drafting history or identify which application produced a sentence. It classifies features of the submitted text against patterns learned from examples. A result therefore concerns a model’s assessment of the text as submitted, not a verified account of the writing process. The London School of Economics makes the same distinction in its guidance: detectors cannot tell the difference between extensive AI substitution and a writer who did substantial work before using AI to improve a final draft. (Staff FAQs on the use of GenAI in teaching, learning and assessment)

Figure: What a detector score can and cannot establish. The diagram summarizes QuillBot’s description of its text analysis and institutional guidance that a result should be considered with context, not treated as proof. (The Technology Behind Our AI Detector | How It Works - Quillbot AI; Staff FAQs on the use of GenAI in teaching, learning and assessment)

How accurate is QuillBot AI Detector?

The clearest numerical result comes from RAID, a research benchmark designed to test generated-text detectors across varied generators, writing domains, decoding methods, and adversarial attacks. RAID’s research paper describes more than six million generated texts spanning 11 models, eight domains, 11 attacks, and four decoding strategies; it also reports that detector performance can fall when the model, domain, sampling method, or text is changed. (RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors)

The current RAID leaderboard lists QuillBot with an area under the ROC curve (AUROC) of 99.62, a true-positive rate (TPR) of 99.62% at a 5% false-positive rate, and a TPR of 98.1% at a 1% false-positive rate. TPR here means the share of AI-generated samples the detector identifies at the stated false-positive operating point on that benchmark. It is not the same thing as 99.62% universal accuracy, and it does not mean a submitted essay has a 99.62% chance of being AI-written. (AI-text detection on RAID All settings with attack (test) benchmark leaderboard · SOTA2 Research)

The studies measure different things on different texts, so their percentages should not be combined or used to rank tools as if they came from one common test. The RAID leaderboard is a comparatively broad benchmark, while the direct experiments provide context-specific snapshots. (AI-text detection on RAID All settings with attack (test) benchmark leaderboard · SOTA2 Research; How Sensitive Are the Free AI-detector Tools in Detecting AI-generated Texts? A Comparison of Popular AI-detector Tools; Evaluating the Performance of AI Text Detectors, Few-Shot and Chain-of-Thought Prompting Using DeepSeek Generated Text; An exploratory study on the effectiveness of AI detection tools in identifying AI-generated articles - Houston Methodist Scholars)

Why the 99% figure needs context

A detector’s true-positive rate and false-positive rate depend on the chosen threshold and on the test corpus. A detector tuned to catch more AI-generated passages may also mark more human passages as AI-like. RAID makes this trade-off visible by reporting TPR at fixed FPR values instead of reducing performance to one headline number. Its authors also emphasize that shared, varied benchmarks matter because evaluations on narrow or easy datasets can make a detector appear more reliable than it is under changed conditions. (RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors; AI-text detection on RAID All settings with attack (test) benchmark leaderboard · SOTA2 Research)

A score displayed as a percentage should not automatically be read as the probability that a writer used AI. QuillBot describes its output as confidence scoring, but its public technology page does not publish enough calibration information to interpret every score as a verified authorship probability. The actual meaning depends on the model, threshold, training examples, and the population and text type being assessed. (The Technology Behind Our AI Detector | How It Works - Quillbot AI)

Can QuillBot be detected as AI?

Sometimes, but the answer depends on what “QuillBot” means and which text is being checked. QuillBot includes both a paraphrasing tool and a detector. AI-generated writing paraphrased through QuillBot may still be classified as AI-like by some detectors, but paraphrasing can also change the result enough that a detector misses the AI origin. A detector cannot reliably identify QuillBot as the particular tool used; it evaluates the final wording rather than a tool-use record.

A direct 2025 experiment illustrates this inconsistency. Researchers created a short article entirely with ChatGPT 3.5 and paraphrased it using QuillBot’s free standard mode. QuillBot’s detector reported 100% AI for the original and 0% AI for the QuillBot-rewritten text. The authors tested only one original text, so the finding demonstrates a possible failure mode; it does not establish that QuillBot paraphrasing always evades detection. (How Sensitive Are the Free AI-detector Tools in Detecting AI-generated Texts? A Comparison of Popular AI-detector Tools)

A separate 2025 study tested 49 human-authored question-and-answer pairs and 49 matched DeepSeek-generated responses, then created additional samples using paraphrasing and humanization. In that study, humanization reduced QuillBot’s accuracy to 58%. The sample and transformation design were specific to DeepSeek, so that result should not be projected onto every kind of QuillBot-edited text or the detector’s current live version. (Evaluating the Performance of AI Text Detectors, Few-Shot and Chain-of-Thought Prompting Using DeepSeek Generated Text)

The reverse inference is also invalid: a positive QuillBot score does not establish that QuillBot was used. AI-like features can appear in ordinary human writing, and the tool’s analysis does not reveal which writing application created or revised the text. A low score is no proof that AI was absent either.

False positives, false negatives, and reliability limits

A false positive occurs when human-written text is flagged as AI-like; a false negative occurs when AI-generated text is classified as human or is not flagged. Both matter. A false positive can prompt an unjustified accusation, while a false negative can create false confidence that a text was written without AI assistance.

A peer-reviewed 2026 study of oral and maxillofacial surgery articles offers one positive signal for QuillBot: among 24 published human-written articles, it did not classify any as AI-written. The same study included 12 AI-generated articles produced using ChatGPT, DeepSeek, Gemini, and Copilot, but its public abstract does not report a full QuillBot detection rate across those AI examples. The study is limited to one specialist literature population and a small sample; zero flags in 24 papers is not evidence that false positives never occur. (An exploratory study on the effectiveness of AI detection tools in identifying AI-generated articles - Houston Methodist Scholars)

There is also a broader fairness concern. Stanford researchers tested seven AI detectors against essays by U.S.-born eighth-graders and English learners who wrote TOEFL essays. Across the detector group, 61.22% of the TOEFL essays were classified as AI-generated, and 89 of 91 were flagged by at least one detector. The result is not QuillBot’s individual false-positive rate; it shows that features associated with less varied or less complex English can cause problems for detectors as a class. (AI-Detectors Biased Against Non-Native English Writers | Stanford HAI)

Performance can also change after paraphrasing, editing, changing the AI model, or using a different genre. RAID researchers found vulnerabilities across detectors to attacks, unseen generators, and generation settings. The practical consequence is that an evaluation of untouched model output does not by itself establish performance on student essays, translated writing, edited drafts, short answers, or text produced by a newer model. (RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors)

QuillBot’s help center advises users to submit at least 80 words when the detector does not produce a result because the text is too short. This is a product operating requirement, not a guarantee that every text above that length can be classified accurately. (Why is my text shown in red without displaying any result? – Quillbot Help Center)

How reliable is QuillBot for school, work, or publishing decisions?

QuillBot can help identify passages that merit closer examination, but it is not reliable enough to serve as stand-alone proof of misconduct or authorship. A detector only sees the final text; it cannot determine whether a writer used a grammar aid, received permitted AI help, drafted independently and later edited with AI, or submitted generated text with little revision. LSE advises that neither an AI-like label nor a human-like label resolves those questions, and says a positive result should not be the sole evidence in an academic-integrity investigation. (Staff FAQs on the use of GenAI in teaching, learning and assessment)

For an instructor, editor, or manager reviewing a result, a more defensible process is to check the applicable policy, examine the relevant passages, compare with prior writing where appropriate, and ask the author to explain their sources or process. Any follow-up should consider other evidence rather than treating a detector score as a conclusion. This approach matches QuillBot’s own warning not to rely on its detector alone when a decision could affect academic standing or a career. (The Technology Behind Our AI Detector | How It Works - Quillbot AI; Staff FAQs on the use of GenAI in teaching, learning and assessment)

For writers checking their own drafts, a score can be a prompt to inspect clarity or sentence patterns, but repeatedly editing only to lower a detector score is not a reliable way to establish authorship. Keep drafts and version history where relevant, follow the stated rules for AI assistance, and disclose tool use when required. For anyone uploading confidential work, institutional privacy rules and the service’s data terms should also be checked before submission.

Conclusion

QuillBot’s detector performs strongly on the RAID benchmark, and a small specialist study found no false AI flags among its 24 human-written articles. But independent tests also show that paraphrasing and humanization can sharply reduce detection, while no published evidence supplies a single accuracy rate that applies to every writer, task, language, and current detector version. The direct answer to “is QuillBot AI Detector accurate?” is therefore: it can be useful for screening, but it is not conclusive. Treat a QuillBot result as an uncertain text classification—not proof that a specific person did or did not use AI.