Most thesis chapters, business memos, legal contracts, and board papers still live in Microsoft Word in 2026. TextSight reads the .docx natively via our officeparser v7 file-extract endpoint, walks paragraphs and headings in reading order, and flags the sentences that read like ChatGPT, Copilot, Claude, or Gemini wrote them. Sentence-level colour map, an Authenticity Score with five named bands, and a four-step upload workflow you can run before sending. Free to try, no card.
Contracts, theses, board memos, journal submissions, white papers; the .docx is the unit of work for most professional writing, and the moment to audit ChatGPT residue is before the export to PDF.
Google Docs has eaten a lot of the casual web-document market, but inside corporations, law firms, publishing houses, and universities, Microsoft Word is still the default. Contracts get redlined in Word with Track Changes; reports get built from Word templates with custom styles; publishers send manuscript guidelines as .docx; most graduate programs require a Word submission for the final thesis draft before the PDF export.
A clear majority of corporate documents pass through Word at some stage. That dominance is exactly why a Word-specific ChatGPT check matters. If your finished artefact is a .docx, scanning a PDF export is one step too late: edits are easier in Word, comments and Track Changes are still live, the formatting is intact, and a writer can revise a flagged paragraph in place without rebuilding the layout.
By the time the document is rendered to PDF for distribution, the cheapest moment to fix AI-flavoured prose has already passed. The Word phase is the audit window. Drop the .docx onto TextSight, read the sentence-level highlights, fix what reads like Copilot, then export the PDF and ship. The classifier is the same one that runs on a paste; the difference is that uploading preserves the reading order across paragraphs, headings, and footnotes far better than copy-paste from a forty-page document.
The same upload route the Chrome extension and the WordPress plugin use, exposed in the web dashboard for Word writers.
Save the Word document, click the file-upload icon next to the text input on app.textsight.ai (or drag the file directly onto the input). Pro accepts .docx, .pdf, and .txt through the same icon. A typical thesis chapter or business memo uploads in under two seconds. The free tier stays paste-only because file processing has a real compute cost beyond the classifier itself.
The file is sent to our file-extract endpoint, which uses officeparser v7 to walk the Word XML package. The reader reads document.xml plus the headers and footers, walks the paragraph tree in reading order, and reconstructs the same plain text a colleague would see in the rendered document with all changes accepted. Paragraph breaks, Word heading styles (Heading 1, 2, 3), bulleted and numbered lists, footnote text, and table cell text all carry through. Tracked deletions, comment threads, embedded images, and revision metadata are stripped by design.
The extracted text fills the textarea so you can verify what the classifier will read before clicking Scan. The classifier returns sentence-level scores in about five to fifteen seconds for a 2,000-word brief. The result panel is identical to a paste: overall Authenticity Score on a 0 to 100 scale with five named bands from Likely AI to Likely Human, sentence-level colour map, Plagiarism Risk warnings, the top AI tells the classifier flagged.
Toggle back to Word, rewrite the flagged sentences in place using Track Changes if you want the diff visible to a reviewer, then re-upload for a second scan. Most pieces go from red to green in two or three iteration cycles. If a sentence flags on every pass, run it through the integrated AI rewriter (Light, Balanced, or Maximum) and drop the rewritten line back into the Word document.
Free covers casual paste scanning. Pro unlocks direct .docx upload for regular Word writers. Business covers legal, compliance, and editorial teams. Full details on the pricing page.
Billed $89.88/year — Save $30
Billed $179.88/year — Save $60
Billed $359.88/year — Save $120
Yearly billing saves 25%. View full pricing →
Copilot for Microsoft 365 in Word is a GPT-4o class generator wired into the ribbon. The Draft With button, the Rewrite pass, and the Editor Copilot suggestions all carry the same statistical fingerprints as a ChatGPT paste.
Microsoft is open about the underlying model class. The classifier is calibrated on Copilot Draft With outputs, Rewrite passes, and the Editor Copilot suggestions that appear inline in Word, alongside raw GPT-4o, Claude, and Gemini drafts. Detection accuracy on raw Copilot prose is comparable to ChatGPT. The classifier reads prose, not the source app, so a paragraph drafted via Copilot triggers the same sentence-level highlights and Authenticity Score drop as a ChatGPT paste.
A lot of writers do not think of Copilot output as "AI-written" the way they would a ChatGPT paste. The user-facing copy frames Copilot as an assistant rather than a generator, but at the model level the output is the same kind of text. For any policy that requires disclosure of AI assistance (academic, legal, editorial), a writer who used Copilot to draft a section has not actually used a different system from a writer who pasted from ChatGPT.
Most real Word documents mix human prose with one or two Copilot Rewrite sections. A single overall score would hide that. TextSight returns sentence-level highlights so a lead editor can target the two paragraphs that account for most of the AI signal rather than rewriting the whole draft.
A modern Word file is a zipped bundle of XML parts. The body text lives in document.xml as a sequence of paragraphs, runs, and text nodes. officeparser v7 walks that tree in reading order so the highlights line up with the document's section structure.
Heading hierarchy (Heading 1, 2, 3), body paragraphs in their original order, list items, table cell text, footnotes, and endnotes. Bulleted and numbered lists are scored as separate sentence units. Footnotes and endnotes are read separately so citation-style language does not pollute the body score. Bold, italics, fonts, and inline formatting do not affect the score because the classifier reads prose, not styling.
Comment threads, deleted text, formatting markup, embedded images, equations, and Track Changes metadata. The output is the prose a reader would read, in the order they would read it, with all markup hidden.
TextSight reads the current accepted text of the document, the same view a reader would see if they opened the file with all markup hidden. Tracked insertions and deletions in the source .docx are folded into the final body before scoring. The right pattern is to scan in parallel with the Track Changes review pass and treat any red sentences as another revision target alongside the reviewer's comments. To audit the original draft against AI residue specifically, accept or reject all changes first or scan a separate copy.
Thesis writers, business memo drafters, legal teams, and contract reviewers all share the same upload-extract-score-review loop. The differences are in cadence and tier.
Most graduate programs still require a Word submission for the final thesis draft because the formatting templates and citation styles target .docx. Students draft in Word, upload each chapter to TextSight the day before submission, and rewrite any red sentences in place. Long thesis chapters move to Pro for the .docx upload and unlimited scans. The Pro tier at 19.99 dollars a month (14.99 on yearly) covers the typical dissertation cycle.
Consultants, analysts, and internal communicators ship board memos, white papers, and client decks as .docx. The audience is small and senior, and a flat, AI-flavoured tone signals that the author did not invest the time. Scanning the .docx before circulating gives writers a chance to revise without anyone outside the loop noticing a Copilot pass. Per-paragraph scoring isolates the one or two Copilot Rewrite sections that usually account for most of an AI flag on an otherwise human document.
Drafting in Word with tracked redlines is the default workflow in law. Contracts, briefs, and filings are drafted, redlined, and exchanged as .docx because Track Changes is still the standard redlining surface. AI-flavoured boilerplate in a contract is rarely a legal problem on its own, but in pleadings, opinions, and briefs, courts have begun pushing back on filings that read as machine-generated. A scan on the final .docx before filing flags the paragraphs most likely to draw attention. The annotated .docx export route on Pro hands the marked-up file back to the drafter as a follow-up edit list.
Publishers ask for .docx submissions with comments and tracked edits enabled. Editors increasingly want a declaration on whether generative AI was used. A scan gives an author a clear-eyed look at which chapters read as their own voice and which drift into model flavour, before the manuscript leaves the desk.
Honest scope: there is no Word task pane today. The .docx upload and paste flows return the same score the future add-in will surface inline.
The planned Office Add-in will dock a task pane inside Word, score the active document on demand, and render flagged sentences using the native Word comments and Track Changes surfaces so the writer can accept or reject AI rewriter rewrites inline. The underlying scan and the Authenticity Score model are identical to the web app. The Office Add-in submission process for the Microsoft AppSource store takes around six to eight weeks of validation after the technical build is ready.
Until that ships, the .docx upload flow is faster than you might expect. Drop a 2,000-word brief onto app.textsight.ai and the scan returns in five to fifteen seconds. Heading hierarchy, paragraph breaks, footnotes, and endnotes all carry through. Long documents above 50,000 characters work cleanly on Business, which raises the per-scan ceiling for full-chapter and long-contract workflows.
If a native add-in would meaningfully change your workflow (for example because a corporate IT policy blocks file upload to third-party SaaS and you need an in-Word task pane to clear procurement), tell us on the contact page. Demand signal directly affects roadmap priority. Excel and PowerPoint support is also tracked there.
Same upload route, .pdf format. For documents already exported for distribution.
For PDF →Multi-model detection on the same .docx upload. ChatGPT, Claude, Gemini, Copilot.
All models →Paste-only spot-checks on the free tier. Three scans a day, no card required.
Free tier →Full tier breakdown for Free, Starter, Pro, and Business. Annual billing saves 25%.
See pricing →Free to try. No card. Paragraph and heading structure preserved.