How do AI detectors handle bullet points, tables and code? The ten vendors that say, the three records that tried, and what each excludes
Ten of the 72 AI detectors in the index state how they treat input that is not prose, by the input limits, vendor claim and caveat fields of every card read 8 and 9 November 2026 in the data file, and the clearest statement is an exclusion: Turnitin's FAQ says the model "eliminates text strings that it predicts are not actual sentences, such as lists, headers, references, data tables, and text that appears to be formulas or code" and "does not reliably detect AI-generated text in the form of non-prose, or code, nor does it detect short-form/unconventional writing such as bullet points" (guides.turnitin.com, read 8 November 2026 and re-read 9 November 2026). Three cards sell AI detection for source code, Codequiry, Originality.ai and Copyleaks Codeleaks, and only one of them publishes a human-code false positive figure on its card: Originality.ai's "no false positives were observed among the 150 human-written samples" (originality.ai, read 8 November 2026). Of the 35 study records, 3 touched code or non-prose at all, and none scored a card on bullet points or tables.
The question has two halves and the table keeps them apart. Bullet points, tables, headers, reference lists and formulas are text that a prose classifier was not trained on, and the vendors that address them say so by excluding them from the score or warning that they raise false positives. Source code is a different input with its own detectors, metered in tokens rather than words. Rows quote the field of the card that carries the statement, name the input type it covers and link the page with its read date. This page is one of the questions answered from the index; the code languages ranking orders the three code detectors by languages covered, the input limits ranking carries the size caps, the code view reads the same cards for engineering teams, and the AI detector register carries every card.
| Detector | Input type addressed | Vendor statement on non-prose input, as the card quotes it | Source | Read date |
|---|---|---|---|---|
| Turnitin | Lists, headers, references, tables, formulas, code, bullet points | "Lists, headers, references, tables, formulas and code are excluded from scoring"; "does not reliably detect AI-generated text in the form of non-prose, or code, nor ... bullet points (short non-sentence structures)" | guides.turnitin.com | 8 November 2026 |
| Pangram | Code, markup, reference lists, templated writing | Model card: code, markup, reference lists, templated writing and AI translations are out of scope | pangram.com | 9 November 2026 |
| GPTZero | Procedural and machine-generated text | "Our classifier can sometimes flag other machine-generated or highly procedural text as AI-generated" | gptzero.me | 8 November 2026 |
| JustDone | Formatted writing, lists | "Short fragments, heavily formatted writing, and lists are the hardest cases for any detector" | justdone.com | 8 November 2026 |
| Quetext | Short phrases versus structured text | "Perform better on longer, structured text but may struggle with short phrases" | quetext.com | 9 November 2026 |
| BrandWell | Templated and repetitive structures | "Human-written text can be flagged, especially when it is short, technical, templated, highly edited, or uses repeated sentence structures" | moxby.com | 9 November 2026 |
| ZeroGPT | Formulaic, low-entropy writing | "Highly polished, formulaic, or low-entropy writing can resemble AI" | zerogpt.com | 8 November 2026 |
| Originality.ai | Source code | "Code: no false positives were observed among the 150 human-written samples" (5.0 launch post); "Code detection still lacking compared to long-form text" (accuracy post) | originality.ai | 8 November 2026 |
| Codequiry | Source code, 65+ languages | "It catches 90% of AI/GPT-written code and falsely flags 1.3% of human-written code" (no sample size published) | codequiry.com | 9 November 2026 |
| Copyleaks Codeleaks | Source code, metered in tokens | "Every 250 tokens count as 1 credit"; since 16 December 2025 the API submit endpoints return no AI verdict for source code files | copyleaks.com | 9 November 2026 |
Statements as each card quotes them from the page linked, read on the date shown; the index has not run any detector on a list, a table or a program.
What happens to bullet points and tables in a detector score?
Turnitin removes them before scoring; the other six text vendors with a statement warn about them without saying what the model does. Turnitin's FAQ states that eliminated sentences "do not factor into the percentage", that the document score is "the percentage of evaluated sentences that are classified as AI", and that a document mixing writing types "would result in a disparity between the percentage and the highlights" (guides.turnitin.com, read 9 November 2026). The same vendor's release notes record two changes to the rule: on 9 August 2023 "long-form prose text in tables" became processable and bibliographies were excluded after a bug "occasionally highlighting AI writing within references", and the 300-word minimum counts only "prose text in a long-form writing format" (guides.turnitin.com, read 9 November 2026). The Turnitin claims record carries the same limit in the vendor's words. JustDone calls "heavily formatted writing, and lists" the hardest cases (justdone.com, read 8 November 2026); BrandWell says templated text and "repeated sentence structures" can be flagged (moxby.com, read 9 November 2026); ZeroGPT says "formulaic, or low-entropy writing can resemble AI" (zerogpt.com, read 8 November 2026); GPTZero says "highly procedural text" can be flagged (gptzero.me, read 8 November 2026). Pangram's model card puts "reference lists, templated writing" out of scope (pangram.com, read 9 November 2026). A reader who pastes slides, a table of results or a reference list into any of the ten is therefore testing input the vendor has either removed or named as a false positive risk.
Which AI detectors detect AI-written code?
Three cards sell it, and the evidence behind them is the vendors' own. Codequiry's AI code detection covers "65+ programming languages", takes code by paste, ZIP or LMS import and states that model v3e "catches 90% of AI/GPT-written code and falsely flags 1.3% of human-written code", with per-tool rates from 91% for GitHub Copilot to 80% for Amazon Q and Meta Llama and no sample size on the page (codequiry.com, read 9 November 2026). Originality.ai lists code among its modalities, reports zero false positives in 150 human-written code samples in its 5.0 launch post (originality.ai, read 8 November 2026) and adds in its accuracy post that "Code detection still lacking compared to long-form text" (originality.ai, read 8 November 2026). Copyleaks Codeleaks lists Ruby, JavaScript, CSS, Markdown, TypeScript, Python, C++, C, Java, C# and PHP, meters code at 250 tokens a credit, and carries a "99.8% (accuracy)" launch figure from 9 November 2023 with no sample; its live page's "0.2% false positive rate" is footnoted to English text datasets, the text detector's figure (copyleaks.com, read 9 November 2026). The same vendor's release notes removed the AI verdict for source code from the API on 16 December 2025, as the shut-down detectors question records, while the Copyleaks text card still lists code among its modalities. None of the three has a study record on code, so the code languages ranking orders them by languages covered, and nothing else. Turnitin's FAQ, read 9 November 2026, says the company is "not pursuing ChatGPT code detection at this time".
Has any study tested detectors on code or non-prose?
Three records touched the question, and all three ended with prose. Weber-Wulff 2023 had computer code available and left it out of the 54-document test, noting that the Writeful GPT Detector "would not accept computer code" because it read code as not English, and that Compilatio returned "NaN% reliability" on a ChatGPT text that included program code (arXiv 2306.15666, read 9 November 2026; Weber-Wulff 2023 record). Orenstrakh 2023 set out to test computing coursework, mined over 2,000 submissions from programming, data structures and compilers courses, and could not use them "because removing the code left too little context", so every one of its 164 tested texts is prose (arXiv 2307.07411, Section 3, read 9 November 2026; Orenstrakh 2023 record). RAID 2024 built RAID-extra, 2.3 million generations of Python code, Czech news and German news, but the paper's detector tables score the main prose set (DOI 10.18653/v1/2024.acl-long.674, read 8 November 2026; RAID record). No record in the index scored any card on bullet points or tables.
Study records that met code or non-prose input
| Record | Non-prose input | What the record reports | Cards scored on it | Source | Read date |
|---|---|---|---|---|---|
| Weber-Wulff 2023 | Computer code, available but untested | Code left out of the 54-document test; Writeful GPT Detector refused code as non-English; Compilatio returned "NaN% reliability" on a text containing code | 0 | arXiv 2306.15666 | 9 November 2026 |
| Orenstrakh 2023 | Over 2,000 programming-course submissions | Unusable once the code was removed; all 164 tested texts are prose | 0 | arXiv 2307.07411 | 9 November 2026 |
| RAID 2024 | RAID-extra: 2.3 million generations of Python code, Czech and German news | Dataset released; detector tables in the paper score the main set | 0 | DOI 10.18653/v1/2024.acl-long.674 | 8 November 2026 |
Figures as each record prints them from its paper, read on the date shown. The other 32 text, image, voice and video cards, read 9 November 2026 in the data file, state nothing about lists, tables or code in the fields read, and the 27 records not listed tested prose, images, speech or video only.
A detector trained on paragraphs scores paragraphs, and every vendor statement on bullet points, tables and code in the index is a statement about what the model leaves out or gets wrong.
Where a detector marks the text it scores, a bullet list shows up as a run of short flagged or unflagged lines: the sentence-level highlighting question, added 10 November 2026, lists the 17 cards whose pages state a per-sentence, per-passage or per-word mark, GLTR and IvyPanda colouring every word among them.
Non-prose noise has a cousin in non-text input: a scanned page is an image until something runs OCR, and the scanned PDF question records that no card in the index states that step for a PDF, read 10 November 2026.
Questions about AI detectors, bullet points and code
Does Turnitin check bullet points?
0 bullet points are scored: Turnitin's FAQ, read 9 November 2026, says the model "does not reliably detect AI-generated text in the form of non-prose, or code, nor does it detect short-form/unconventional writing such as bullet points", and that lists, headers, references, data tables, formulas and code are eliminated before the percentage is calculated (guides.turnitin.com).
Which AI detectors work on source code?
3 of 42 cards sell AI code detection, read 9 November 2026: Codequiry (65+ languages, "catches 90% ... falsely flags 1.3%", no sample size), Originality.ai (0 false positives in 150 human code samples) and Copyleaks Codeleaks (11 languages listed, AI verdict removed from the API on 16 December 2025) (codequiry.com; originality.ai; copyleaks.com).
Do tables make an AI detector score wrong?
1 vendor answers directly: Turnitin excludes data tables from scoring and has processed "long-form prose text in tables" since 9 August 2023, so a table changes which sentences are counted, not the verdict on them (guides.turnitin.com, read 9 November 2026). The other 9 vendors with a statement warn about formatted or templated text without naming tables.
Has any study tested AI detectors on code?
0 of 35 records scored a card on code, read 9 November 2026: Weber-Wulff 2023 left its code aside (arXiv 2306.15666), Orenstrakh 2023 found 2,000 programming submissions unusable once the code was removed (arXiv 2307.07411), and RAID 2024 released 2.3 million code and non-English generations in RAID-extra without scoring the cards on them (DOI 10.18653/v1/2024.acl-long.674).
Sources
All read 9 November 2026 unless an earlier date is shown; the source of every row is linked in the tables.
- 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 for the card, re-read 9 November 2026)
- Turnitin AI writing detection model release notes: https://guides.turnitin.com/hc/en-us/articles/28294949544717-AI-writing-detection-model
- Pangram 4.1 model card: https://www.pangram.com/research/model-card/pangram-4-1
- GPTZero FAQ: https://gptzero.me/faq (read 8 November 2026)
- JustDone AI detector: https://justdone.com/ai-detector (read 8 November 2026)
- Quetext AI content detector: https://www.quetext.com/ai-content-detector
- Moxby AI Content Detector Mod (BrandWell): https://moxby.com/marketplace/ai-content-detector/
- ZeroGPT terms of use: https://www.zerogpt.com/terms-of-use (read 8 November 2026)
- Originality.ai, Introducing Originality 5.0 and Max: https://originality.ai/blog/introducing-originality-5-0-and-max (read 8 November 2026)
- Originality.ai, AI accuracy post: https://originality.ai/blog/ai-accuracy (read 8 November 2026)
- Codequiry, detect AI-written code: https://codequiry.com/detect-ai-written-code
- Copyleaks Codeleaks: https://copyleaks.com/codeleaks
- Weber-Wulff et al. 2023, preprint: https://arxiv.org/pdf/2306.15666
- Orenstrakh et al. 2023: https://arxiv.org/pdf/2307.07411
- Dugan et al. 2024, RAID: https://doi.org/10.18653/v1/2024.acl-long.674 (read 8 November 2026)
Revision log: 2026-11-09: page created from the input limits, vendor claim and caveat fields of 42 cards and three study records at index version 2026.11.8; 2026-11-10: sentence-level highlighting question linked (eleventh wave); 2026-11-10: twelfth-wave pages linked
Related pages: Which AI detectors were in the Weber-Wulff 2023 study? The 14 tools, their result on 54 documents, and what each one is today, Which AI detectors detect AI voice and video in one product? The nine cards that list both, and the zero records that tested either, AI code detectors ranked by the programming languages the vendor names and Questions about AI detectors, answered from the index.