Which AI detectors offer sentence-level highlighting instead of a single score? The 17 cards whose pages state it, and what each highlights
Seventeen of the 72 AI detectors in the index state on a vendor page that the result marks parts of the text, a sentence, a passage or a word, rather than returning one document score alone, by the input limits, plan, claim and caveat fields of every card read 8 to 10 November 2026: two colour every word (GLTR's four-colour rank map and IvyPanda's red-to-blue word map), twelve state sentence-level marks (GPTZero, Proofademic, Walter Writes, AIDetectPlus, Smodin, BrandWell, Quetext, PlagiarismCheck, Paperpal, Detecting-AI.com, StealthWriter and Phrasly) and three mark passages or segments (Pangram's AI Assistance view, Compilatio's highlighted passages and Turnitin's retained segments). Two more cards mark something other than AI text: Grammarly Authorship colour-codes typed against pasted text as a process record, and Sensity returns heatmaps on images and video. The other 53 cards state no highlighting on the pages the index read, which is not the same as having none.
A highlight is a product feature the vendor describes, read here as a statement and never as a measure of whether the marked sentences are the AI ones; no indexed study scores a highlight. GPTZero's own FAQ says the model is less accurate per sentence than per document, and PlagiarismCheck lets the user move the threshold that decides how many flagged sentences count. The minimum input ranking records the shortest text each tool will score at all, and the bullet points and code question asks what the same tools do with text that is not prose. Rows name the unit highlighted, the output form in the vendor's words, the page that states it and the read date. This page is one of the questions answered from the index; the AI detector index carries every card.
| Detector | Unit highlighted | Output form, in the vendor's words | Source | Read date |
|---|---|---|---|---|
| GLTR | Word | Each word coloured by its rank under GPT-2 117M: top 10 green, top 100 yellow, top 1,000 red, the rest purple, with three histograms over the whole text; no single score | gltr.io | 9 November 2026 |
| IvyPanda | Word | "a colour-coded word map with diagrams": "likely generated by an AI if it is mostly red with some orange words"; "human-made if our GPT essay checker adds plenty of orange, green, and blue"; no percentage | ivypanda.com | 10 November 2026 |
| GPTZero | Sentence | Document, paragraph and sentence-level classification, with the FAQ's caveat that accuracy "on the document-level classification will be greater than the accuracy on the paragraph-level, which is greater than the accuracy on the sentence level" | gptzero.me/faq | 8 November 2026 |
| Proofademic | Sentence | "Sentence-level AI highlighting" listed on the Essential plan beside a document score; the research page reports a 92.9% sentence-level F1 on a synthetic mixed-authorship set | proofademic.ai/pricing | 8 November 2026 |
| Walter Writes | Sentence | "Treat the score and sentence-level items as review signals and consider the text, writer, and use case before making a decision" | walterwrites.ai | 8 November 2026 |
| AIDetectPlus | Sentence | "the per-sentence reasoning matters more than the headline percentage" | aidetectplus.com | 9 November 2026 |
| Smodin | Sentence | A score without an account; "signing up adds sentence-by-sentence analysis" | smodin.io | 9 November 2026 |
| BrandWell | Sentence and phrase | "drills down into specific sentences and phrases to highlight what sounds and reads more robotic and more humanlike" (archived page) | web.archive.org, 9 August 2026 | 9 November 2026 |
| Quetext | Sentence | "line-by-line accuracy"; "confidence scores and highlights specific areas"; the API returns per-sentence AI probability | quetext.com | 9 November 2026 |
| PlagiarismCheck | Sentence | Sentence-level results with a threshold the user sets from 5% to 50% of flagged sentences; the API returns percent, strong_percent and likely_percent | plagiarismcheck.org | 9 November 2026 |
| Paperpal | Sentence | A three-band score (AI-written, human-written, human-AI blend) with sentence highlights; "highlights edited and paraphrased AI text at the sentence level" | paperpal.com | 10 November 2026 |
| Detecting-AI.com | Sentence | An overall score as the share of characters covered by AI-labelled sentences, with sentence highlights in three bands (75 to 100 strong, 30 to 74 possible, 0 to 29 low) | detecting-ai.com | 10 November 2026 |
| StealthWriter | Sentence | "Paste up to 500 words and get sentence-level feedback" on the no-account demo | stealthwriter.ai | 10 November 2026 |
| Phrasly | Sentence | A score for everyone; "text highlighting only for signed-in users" | phrasly.ai | 8 November 2026 |
| Pangram | Segment | "AI Assistance: Pangram will highlight segments that have been edited by AI rather than fully AI-generated", beside a document verdict | pangram.com | 8 November 2026 |
| Compilatio | Passage | "works passage by passage and returns a score for the estimated proportion of AI text with the passages highlighted" | compilatio.net | 9 November 2026 |
| Turnitin | Segment | A percentage "along with the segments identified highly likely written by AI", retained as part of the similarity report | guides.turnitin.com | 8 November 2026 |
Statements as the fields of each card quote them from the page linked, read on the date shown; 17 of 72 cards, counted 10 November 2026 in the data file.
Which AI detectors colour every word?
Two, and both are free. GLTR, the 2019 Giant Language Model Test Room, overlays a colour on each word by where GPT-2 117M ranked it as the next word, green for the top 10 predictions, yellow for the top 100, red for the top 1,000 and purple for everything else, and adds three histograms over the whole text; it returns no single score at all (gltr.io, read 9 November 2026). IvyPanda's GPT Essay Checker does the same with a different palette and a verdict in words: the result is "a colour-coded word map with diagrams", "likely generated by an AI if it is mostly red with some orange words" and "human-made if our GPT essay checker adds plenty of orange, green, and blue to the color palette", on a 4,500-character paste box with no percentage (ivypanda.com, read 10 November 2026). A word map shows the reader what the model found predictable; it does not tell the reader which sentences were written by a machine, and GLTR's own paper measured the colouring by how much it helped people, raising their detection of fake text from 54% to 72% (arXiv 1906.04043, read 9 November 2026).
Which AI detectors mark sentences?
Twelve state it, in three forms. Seven describe a sentence mark as part of every result: GPTZero's FAQ describes classification at document, paragraph and sentence level (gptzero.me/faq, read 8 November 2026); Quetext promises "line-by-line accuracy" and returns per-sentence AI probability through its API (quetext.com, read 9 November 2026); PlagiarismCheck returns sentence-level results with a threshold the user sets between 5% and 50% of flagged sentences (plagiarismcheck.org, read 9 November 2026); Paperpal and Detecting-AI.com each pair sentence highlights with a three-band score (paperpal.com and detecting-ai.com, read 10 November 2026); StealthWriter's demo gives "sentence-level feedback" on up to 500 words (stealthwriter.ai, read 10 November 2026); and BrandWell's archived page "drills down into specific sentences and phrases" (web.archive.org, read 9 November 2026). Three describe the sentence view as a caution or a reading aid: Walter Writes tells users to "treat the score and sentence-level items as review signals" (walterwrites.ai, read 8 November 2026), AIDetectPlus says "the per-sentence reasoning matters more than the headline percentage" (aidetectplus.com, read 9 November 2026), and Proofademic lists "Sentence-level AI highlighting" as an Essential plan feature (proofademic.ai, read 8 November 2026). Two gate the sentence view behind a sign-in: Smodin's FAQ says signing up "adds sentence-by-sentence analysis" and Phrasly offers "text highlighting only for signed-in users" (smodin.io and phrasly.ai, read 9 and 8 November 2026).
Which AI detectors mark passages or segments rather than sentences?
Three, each with its own unit. Pangram's AI Assistance view highlights "segments that have been edited by AI rather than fully AI-generated", a mark aimed at mixed authorship rather than at whole-AI text (pangram.com, read 8 November 2026). Compilatio "works passage by passage" and returns a proportion score "with the passages highlighted" in the Magister+ and Studium reports (compilatio.net, read 9 November 2026). Turnitin retains the AI percentage "along with the segments identified highly likely written by AI" inside the similarity report, and its FAQ adds that customers cannot delete the AI component alone (guides.turnitin.com, read 8 November 2026). A segment can be a sentence, a paragraph or a span the vendor's model chose; none of the three pages defines its length.
What do the other cards return?
A document score, a verdict, or a result on a file that is not text, by the pages the index read. Two cards mark something other than AI text and are listed here so a reader does not mistake them: Grammarly Authorship's reports carry "Typed by a Human analytics and colour-coded text", a record of how text entered the document rather than a classification of it (support.grammarly.com, read 9 November 2026), and Sensity returns "a binary verdict, suspicious or valid, with confidence scores and heatmaps" on images and video (docs.sensity.ai, read 9 November 2026). The remaining 53 cards state no highlighting on the pages recorded in their fields, among them Copyleaks, Originality.ai, Winston AI and Sapling, whose cards quote accuracy claims, limits and policies and no sentence view; a feature the vendor has but did not state on those pages is not entered, and the next recheck of every card is 8 December 2026 (methodology).
Cards marking something other than AI sentences, and the four most-tested cards with no highlighting statement
| Detector | What the page states | Source | Read date |
|---|---|---|---|
| Grammarly Authorship | "Typed by a Human analytics and colour-coded text" in every Authorship report; a process record of typed against pasted text, not an AI score | support.grammarly.com | 9 November 2026 |
| Sensity | "a binary verdict, suspicious or valid, with confidence scores and heatmaps" on uploaded images and video | docs.sensity.ai | 9 November 2026 |
| Copyleaks | No highlighting statement in the fields read; a testing methodology page with a model version and test date | copyleaks.com | 8 November 2026 |
| Originality.ai | No highlighting statement in the fields read; a 99%+ (accuracy) claim for 5.0 and a release history | originality.ai | 8 November 2026 |
| Winston AI | No highlighting statement in the fields read; a zero-false-positive claim across 27,743 human documents | gowinston.ai | 8 November 2026 |
| Sapling | No highlighting statement in the fields read; "Accuracy varies with the dataset, text length, and decision threshold" | sapling.ai | 8 November 2026 |
A highlighted sentence is the vendor's guess about a part of the text, made with less context than the document score, and GPTZero is the one vendor on this page that says so in its FAQ.
No detector score, highlighted or not, should be the only evidence in a misconduct, hiring or publishing decision.
Questions about sentence-level AI detection
Which AI detectors show which sentences are AI?
12 of 72 cards state a sentence-level mark on the pages the index read 8 to 10 November 2026: GPTZero, Proofademic, Walter Writes, AIDetectPlus, Smodin, BrandWell, Quetext, PlagiarismCheck, Paperpal, Detecting-AI.com, StealthWriter and Phrasly. 2 more colour every word (GLTR, IvyPanda) and 3 mark passages or segments (Pangram, Compilatio, Turnitin).
Is a sentence-level result more accurate than a document score?
No, by the one vendor statement on the point: GPTZero's FAQ says accuracy "on the document-level classification will be greater than the accuracy on the paragraph-level, which is greater than the accuracy on the sentence level" (gptzero.me/faq, read 8 November 2026). 0 indexed studies score highlights sentence by sentence; the independently tested ranking sorts on document-level records.
Does GPTZero highlight sentences?
Yes, by its FAQ's description of sentence-level classification, read 8 November 2026 (gptzero.me/faq); the same FAQ ranks that level as the least accurate of the three. 16 study records test GPTZero's document verdicts, none its sentence marks.
Which free AI detectors highlight without an account?
5 cards state a highlight with no sign-in: GLTR and IvyPanda (word colours), StealthWriter (sentence feedback on up to 500 words), Detecting-AI.com (sentence bands on up to 15,000 characters) and Quetext (line-by-line on its free 5,000 words), read 8 to 10 November 2026. Smodin and Phrasly give the score free and the sentence view to signed-in users only.
Which AI detectors highlight edited or mixed AI text?
3 state it: Pangram's AI Assistance view marks "segments that have been edited by AI rather than fully AI-generated" (pangram.com, read 8 November 2026), Paperpal "highlights edited and paraphrased AI text at the sentence level" (paperpal.com, read 10 November 2026) and Detecting-AI.com lists "Mixed drafts" and "Human-edited and AI-assisted writing" in its coverage table (detecting-ai.com, read 10 November 2026). The paraphrased text question sets those claims beside the study results.
Sources
All read 8 to 10 November 2026 on the date shown in each row; the source of every row is linked in the tables.
- GLTR demo: http://gltr.io/ (read 9 November 2026)
- Gehrmann, Strobelt and Rush 2019, GLTR: https://arxiv.org/abs/1906.04043 (read 9 November 2026)
- IvyPanda GPT Essay Checker: https://ivypanda.com/tools/gpt-essay-checker/ (read 10 November 2026)
- GPTZero FAQ: https://gptzero.me/faq (read 8 November 2026)
- Proofademic pricing: https://proofademic.ai/pricing/ (read 8 November 2026)
- Walter Writes AI detector API page: https://walterwrites.ai/ai-detector-api/ (read 8 November 2026)
- AIDetectPlus AI detector: https://aidetectplus.com/ai-detector (read 9 November 2026)
- Smodin AI Content Detector: https://smodin.io/ai-content-detector (read 9 November 2026)
- BrandWell AI Content Detector (archived 9 August 2026): https://web.archive.org/web/20260809041054/https://brandwell.ai/ai-content-detector/ (read 9 November 2026)
- Quetext AI Content Detector: https://www.quetext.com/ai-content-detector (read 9 November 2026)
- PlagiarismCheck AI detector: https://plagiarismcheck.org/ai-detector/ (read 9 November 2026)
- Paperpal AI detector: https://paperpal.com/tools/ai-detector (read 10 November 2026)
- Detecting-AI.com: https://detecting-ai.com/ (read 10 November 2026)
- StealthWriter AI detector: https://stealthwriter.ai/ai-detector (read 10 November 2026)
- Phrasly AI Detector: https://phrasly.ai/ai-detector (read 8 November 2026)
- Pangram home page: https://www.pangram.com/ (read 8 November 2026)
- Compilatio Studium: https://www.compilatio.net/en/product/studium/ (read 9 November 2026)
- Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 8 November 2026)
- Grammarly Authorship support article: https://support.grammarly.com/hc/en-us/articles/29548735595405 (read 9 November 2026)
- Sensity formats and limits: https://docs.sensity.ai/resources/formats-limits (read 9 November 2026)
Revision log: 2026-11-10: page created from the input limits, plan, claim and caveat fields of 72 cards at index version 2026.11.12
Related pages: Which AI detectors have a student plan? The four cards that price a student, academic or education tier for one person, and the nine that sell education to institutions only, Do AI detectors work on scanned PDFs? The 18 cards that accept a PDF upload, the four that say the file needs a text layer, and the three that OCR images instead, AI detectors ranked by independent testing Two records added in the fifth wave on 9 November 2026, Pegoraro 2023 and Fleckenstein 2024, changed nine counts and none: Pegoraro 2023 tested ZeroGPT, GPTZero, Copyleaks, Originality.ai, Writer, Content at Scale (BrandWell), DetectGPT, the GPT-2 Output Detector and GLTR, lifting GPTZero to 12, the GPT-2 Output Detector to 10, Copyleaks to 7 and Originality.ai and ZeroGPT to 6; Fleckenstein 2024 tested 289 teachers and no software, so it counts for no card ([Fleckenstein 2024 record](/studies/fleckenstein-2024-teachers/)). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify. and Questions about AI detectors, answered from the index.