Who makes DrillBit, and how it reached your department
The company is DrillBit SoftTech India Private Limited, of Bengaluru. It says it was established in 2016 and launched its AI detection engine in 2024 (its about page, read 26 September 2026). In a letter dated 4 June 2026, which INFLIBNET hosts on the ShodhShuddhi site, its chief executive announced a move from drillbitplagiarism.com to drillbitglobal.com and described the company as serving “2,500+ institutions and 400,000+ users across 20+ countries” (the letter). That count is DrillBit’s own.
Many researchers in India meet it through ShodhShuddhi, which INFLIBNET switched to DrillBit-Extreme from 1 October 2023, a trial month before full service began on 1 November 2023 (ShodhShuddhi’s about page). The story of that switch, and of the Ouriginal contract before it, is on our page on the UGC plagiarism rules; it is not repeated here. What has changed since is scale and uncertainty. ShodhShuddhi’s own counter, “as on 25.09.2026”, showed 1205 member institutions, 186315 users and 3016997 total documents submitted (ShodhShuddhi home page, read 26 September 2026). Beside a separate counter for documents submitted in August 2026, the page prints “1 Doc. = 20-A4 Pages”; INFLIBNET does not say whether the cumulative total uses the same unit.
The licence has been renewed in steps. Notices on the same page extend DrillBit-Extreme “for another six months” (30 April 2024), then “for one more year i.e. up to 31.10.2025” (1 November 2024), then “for another Six Months” (31 March 2026). That latest notice gives no end date, and INFLIBNET had published nothing about what follows when we read the page.
Who the national licence covers
Doctoral work and faculty, and no one else. INFLIBNET’s revised guidelines of 17 June 2025 restrict ShodhShuddhi checks to “PhD theses and research articles, Thesis chapters, and synopses”, grant access to PhD and post-doctoral students, JRF/SRF research scholars and faculty, and state that “Plagiarism check for Undergraduate (UG) or Postgraduate (PG) students is strictly prohibited” (the guidelines). For a doctoral candidate, what the AI report reads in a long chapter is the subject of our guide to AI checks on a thesis.
If you are a master’s student holding a DrillBit report, then, the national licence does not cover you. The report may still be genuine. It came through some other account: perhaps your institution’s own DrillBit subscription, where the vendor’s FAQ says it “supports administrator, instructor, and student roles”, or a personal account bought with page credits. Knowing which one is the first thing worth establishing.
Is AI detection part of the national licence?
Yes. INFLIBNET’s FAQ asks the question directly and answers: “Yes, there is a separate report for AI detection, and the report can be downloaded from the analysis page. It will generate an AI report for all submissions without requiring any settings” (ShodhShuddhi FAQs, read 26 September 2026).
So, by INFLIBNET’s account, uploads through ShodhShuddhi come with an AI report whether anyone asked for one or not. (DrillBit’s user guide makes an exception for regional-language files, covered below.) We could not find any document saying a faculty user can switch that report off. The FAQ describes no setting at all.
What DrillBit claims for its AI detector
The product page makes its case in words. It promises “Unparalleled accuracy” and “Low false positives”, the second explained as a model that “reduces false alarms so genuine human writing doesn’t get wrongly flagged” (DrillBit’s AI content detection page, read 26 September 2026). The percentages visible in that page’s graphics belong to a mock-up of the interface. They are not reported results, and we have not treated them as such.
The numbers sit in a white paper linked from the same page: DrillBit AI Detection: Accuracy, Methodology, and Performance Evaluation, Version 1.0, 2026. These are DrillBit’s figures about DrillBit, from its own study, and we found no independent reproduction of them (searched 26 September 2026).
| Measure | DrillBit’s own figure |
|---|---|
| Human-written documents classed as human | 93% |
| False positives (human documents that scored above 20% AI) | 7% |
| AI-generated documents identified | 83% |
| False negatives (AI writing missed) | 17% |
| Overall accuracy | 88% |
| Languages evaluated | English only |
The test set is described in some detail. Human texts were drawn from research papers published before 2018, student papers from before 2015 and “Historical news articles from archival databases”. AI texts came from ChatGPT, Gemini and Grok, and “Each AI-generated sample contained a minimum of 500 words”. The paper also says “The 500,000 hybrid samples were not separately reported in this evaluation”. Mixed authorship, where a person drafts and a tool touches some of it, is therefore the case the published figures leave out.
To its credit, the paper spells out what its own error rate means in practice: “In an academic institution processing 10,000 submissions per semester, a 7% false positive rate would generate approximately 700 incorrectly flagged documents”. It also records that “A subset of AI-generated samples received a 0% AI score”. One oddity: the document is stamped “CONFIDENTIAL -- FOR AUTHORIZED DISTRIBUTION ONLY”, although the product page links to it openly. We quote it because DrillBit published it.
For its evaluation (section 6 of the paper) a human document counted as a false positive once its AI score passed 20%, and an AI document counted as detected only at 60% or more. Nothing in the evaluation went to a person: a human document above 20% was simply a false positive, and an AI document below 60% a miss. Section 7.4 then describes that gap as “a deliberate 20–60% uncertainty zone that the system routes to human review”. That is the paper’s claim about the live system, and the report guide does not show it: the guide says “Detection thresholds are calibrated per language” but publishes none of them, and describes no band or review flag on the report.
Which languages? DrillBit’s own documents disagree
This matters more in India than almost anywhere, because theses are written in Hindi, Tamil, Bengali, Marathi and many other languages. Four DrillBit documents, read on 26 September 2026, do not agree.
| DrillBit source | AI languages stated |
|---|---|
| Home page | “Being the first platform to support both international and Indian regional languages for universities with similarity and AI generated content detection” |
| Languages page | AI detection “currently supports 5 languages”: English, French, Spanish, Indonesian and Vietnamese. Indian regional languages are listed for plagiarism detection only. |
| DrillBit Pro User Guide, the January 2026 file (dated in its name) that INFLIBNET hosts | AI reports for English, Spanish, French and Indonesian. “Grammar and AI Reports are not generated for regional uploads.” |
| White paper, 2026 | “The current evaluation covers English-language documents”. Hindi, Arabic, Spanish, French and Mandarin are listed as future development. |
We are not going to choose between them; that is DrillBit’s job. Whichever is right, two facts stand. The guide INFLIBNET gives its own users says a regional-language upload gets a plagiarism report and no AI report. And the only accuracy figure DrillBit publishes was measured on English, so for a thesis in any other language there is no published DrillBit accuracy figure to point to.
What the AI report shows you
DrillBit’s report guide (read 26 September 2026) describes a separate, multi-page PDF. Its headline is “AI Text Detected (%)”, given to two decimal places, followed by a count of matched words, a breakdown into Human, AI and Excluded text, a table of text blocks, and sentences highlighted in blue.
The guide is careful about what that headline means. “DrillBit’s AI Content Detection produces an estimate, not a verdict. Even with a high AI Text Detected percentage, the result is a hint that some sections may have been written or assisted by an AI tool - it is not proof.” The product page says the same in fewer words: results “should be reviewed alongside human judgment and context”.
It also names the writing most likely to push the number up: “formal, translated, or template-based text - can produce elevated signals in any system”. A thesis is formal by design, synopses follow university templates, and a scholar who drafts in one language and writes up in another is producing partly translated text. That combination sits close to the pattern the research on non-native English writers describes.
The guide flags a couple more things worth checking on any report you are shown.
- Excluded text leaves the calculation. Content marked as excluded drops out of the AI percentage, so the same chapter can score differently depending on what was excluded.
- A thin extract distorts the figure. In the guide’s words, if the text count “is far smaller than the document you submitted, DrillBit could only extract part of the text - usually because the file contains scanned images rather than selectable text. The percentage is then calculated on that smaller extract”. Our note on AI detection in PDFs covers why scanned pages cause this.
Every AI report prints a “Detection Methodology” link to drillbitglobal.com/ai-method. When we requested that address on 26 September 2026 it returned HTTP 404, page not found.
On length, DrillBit’s specifications give file limits of 100 MB and 800 pages, counting roughly 250 words to a page. At the other end, the user guide lists, among the reasons a file fails with a DocError, that it has “fewer than the minimum required 50 words”. A file that fails is not analysed, so it gets no AI report either. We found no length limit set for the AI report specifically. For scale, the white paper’s own AI samples were at least 500 words each.
What we could not verify
On 26 September 2026 we searched Google Scholar for independent evaluations of DrillBit’s AI detection, including searches for false-positive findings. We tried ShodhGanga, INFLIBNET’s thesis repository, the same day; its search returned a bot-check page titled “Making sure you’re not a bot!”, so it could not be searched. We also looked for a UGC or INFLIBNET circular on AI-generated text and found none.
One experimental test turned up. Dutta and Chakraborty, writing in College Libraries in December 2024, generated ten documents with Chatsonic (six in English, two in Bengali, two in Hindi) and ran them through three tools: DrillBit, Grammarly and iThenticate, Turnitin’s product for publishers. Only DrillBit and Grammarly returned AI results; iThenticate’s AI column reads “AI detection not been included yet”. All ten documents were entirely AI-generated.
The language split matters here. DrillBit returned 0% AI on six documents, and four of those six were the Bengali and Hindi ones. DrillBit’s current user guide says regional-language uploads get no AI report at all, and the paper does not say whether a 0% on those files was a reading or an absence of one. On the six English documents, DrillBit scored 0% twice and 56%, 64%, 70% and 75% on the other four (the paper’s Table 2). The study included no human-written text, so it says nothing about false positives, and the authors note it rested on only ten documents.
The other independent comment we found is a 2024 chapter from National Law University and Judicial Academy, Assam, which says that “it is too early to judge the reliability and accuracy of the AI detection mechanism of the software”. Nothing we found since has settled it.
The net position: the 7% false-positive rate is DrillBit’s own number, on English text, and we found no independent false-positive rate for DrillBit’s AI detection anywhere.
DrillBit and Turnitin, side by side
The two are often searched for together. The table holds only what each vendor, or INFLIBNET, documents, read on 26 September 2026. It makes no judgement about which is better, and the two false-positive figures are not comparable: each vendor measured its own tool, on its own test set, by its own definition. More on Turnitin’s side is on our Turnitin page.
| Documented fact | DrillBit | Turnitin |
|---|---|---|
| AI-detection languages | Sources conflict: 5 (languages page), 4 (user guide, file dated January 2026 in its name); accuracy tested in English only | English, Spanish, Japanese, Arabic (file requirements) |
| Minimum text for an AI result | No AI-specific minimum documented; any file under 50 words fails with a DocError and is not analysed (user guide, DocError section) | 300 words of prose (file requirements) |
| Maximum | None documented for AI; files up to 800 pages or 100 MB (specifications) | 30,000 words (file requirements) |
| Vendor’s false-positive figure | 7% of human-written documents scored above 20% AI, English (white paper, sections 6 and 7.4) | Under 1%, for documents with over 20% AI writing (FAQ) |
| Who sees the AI result | The report is downloaded from the analysis page (INFLIBNET FAQ); ShodhShuddhi accounts go to PhD scholars and faculty (guidelines) | Instructors and admins; “not visible to students” unless an instructor shares the PDF (FAQ) |
| Can it be switched off? | No setting documented; generated “for all submissions” (INFLIBNET FAQ) | Admins can switch it off (FAQ) |
| Supplied under ShodhShuddhi | Yes, DrillBit-Extreme from 1 October 2023, full service from 1 November 2023 (about page); latest notice, 31 March 2026, extends it “for another Six Months” with no end date given (ShodhShuddhi) | Not now; Turnitin’s Ouriginal was the programme’s software until 30 September 2023 (about page) |
What the UGC regulation does with a DrillBit report
The UGC’s 2018 regulation grades a report’s similarity percentage into Levels 0 to 3, starting from “Level 0: Similarities upto 10% - Minor similarities, no penalty”, and its text makes no reference to AI or machine-generated writing; the only “generated” in it is “computer generated microfiche”. A DrillBit AI percentage therefore sits outside every level the regulation defines. How the levels and panels work is on our UGC rules page, and why the two numbers measure different things is set out in is a plagiarism score the same as an AI score?
If a DrillBit AI score is quoted at you
Start with questions, not a defence. Each of these has an answer somebody can look up.
- Which report is the number from? DrillBit issues the similarity report and the AI report separately. The UGC levels apply to the first.
- Which account ran it, under which licence? ShodhShuddhi access is limited to doctoral work and faculty. A report on a postgraduate dissertation came from somewhere else.
- What language was the document in, and how much text was read? The INFLIBNET-hosted guide says regional-language uploads get no AI report. A text count far below your page count means the percentage was taken over a partial extract.
- Which rule attaches a consequence to an AI percentage? Not the 2018 regulation. If your university has its own AI policy, ask for the document and the date it was adopted.
- Is the panel’s anti-plagiarism tools expert involved? The regulation seats “A person well versed with anti plagiarism tools” on the Departmental Academic Integrity Panel (UGC compendium). That is the person to show DrillBit’s own caveats: “an estimate, not a verdict”, a 7% false-positive rate on its own English test, and a 20–60% band the paper says the system routes to human review.
- Bring the drafting record. Version history, dated chapter drafts, notes and supervisor emails show a process no percentage can. How to prove you wrote it covers what to gather, and what to do if you are accused covers the conversation itself.
Where we stand
HumanFlow sells an AI detector and an AI humanizer, which puts us in the same market as DrillBit, so read this page knowing that. We have not run DrillBit ourselves and publish no result for it. Every figure above belongs to DrillBit, INFLIBNET, the UGC or Turnitin, and each is linked to the page it came from.
HumanFlow works in English only. It cannot tell you anything about a Hindi, Tamil or Bengali chapter, and its reading of an English one is a different model’s opinion, not a preview of DrillBit’s. The Dutta and Chakraborty test shows how far two tools can differ on identical documents. We will revise this page when DrillBit’s methodology link resolves or an independent test with human-written text appears.
Related
- UGC rules on AI and plagiarism — the levels, the panels and the ShodhShuddhi history
- Compilatio, reviewed — another checker that dominates one country’s universities, with published figures of its own
- Are AI detectors accurate?