What the regulation actually says
The UGC (Promotion of Academic Integrity and Prevention of Plagiarism in Higher Educational Institutions) Regulations, 2018 were approved at the Commission’s 530th meeting on 20 March 2018 and notified in the Official Gazette of India on 31 July 2018. On 6 August 2018 the Secretary wrote to the Vice-Chancellor of every university asking that they be put in force and “accorded Top Most Priority”.
It grades similarity into bands. In the regulation’s own words:
- “Level 1: Similarities above 10% to 40%”
- “Level 2: Similarities above 40% to 60%”
- “Level 3: Similarities above 60%”
Below the 10% floor at which Level 1 begins, similarity is treated as minor. Penalties escalate by level; read them in the regulation itself rather than from anyone’s summary, including ours.
The gap this leaves
Every threshold above is a similarity percentage. Similarity software works by matching your sentences against a corpus of existing text. A language model does not copy from that corpus — it generates sentences that have never been written before.
The consequence is uncomfortable for everyone involved: AI-written work can return a clean similarity report, while an honest, heavily-cited literature review can breach Level 1 on quotations and standard phrasing alone. The 2018 framework was built for copying, and AI writing is not copying.
This is the same structural point as SafeAssign, which does not detect AI either — except here it is written into a national regulation rather than one vendor’s product.
Is there a UGC limit for AI content?
We could not find one. The 2018 regulations set no AI percentage and describe no AI detection procedure. If a department is applying a “20% AI” ceiling, that is an institutional rule — which it is entitled to make — and it is entirely fair to ask which document it comes from and when it was adopted.
A warning about sources. Searching for this turns up several pages confidently describing “UGC AI Plagiarism Guidelines 2026” with specific percentages. They are content marketing for thesis-writing services, and we could not trace their claims to any UGC notification. We are not going to repeat a number we cannot find in a gazette, and neither should a panel deciding your case.
What the government actually provides
The Ministry of Education runs ShodhShuddhi through the INFLIBNET Centre, supplying plagiarism detection to 1100+ institutions since 1 September 2019. It provided Ouriginal — formerly Urkund — until 30 September 2023. In INFLIBNET’s own words: “Starting from October 1, 2023, the INFLIBNET Centre now offers the ‘DrillBit-Extreme Plagiarism Detection Software’”, with full service from 1 November 2023.
Note what that programme is: a plagiarism detection service. If an AI score is being quoted at you in an Indian university, it is worth establishing which tool produced it, because it may not be the one the government supplied.
Who decides your case
Two panels, and the composition matters. The Departmental Academic Integrity Panel is chaired by the Head of the Department and includes a senior academician from outside the department and, in the regulation’s phrase, “a person well versed with anti plagiarism tools”. The DAIP “shall investigate the matter and submit its recommendations to the Institutional Academic Integrity Panel (IAIP)”.
That third DAIP member is the person to address a technical objection to. The regulation anticipated that reading these tools requires expertise; a panel treating a percentage as self-evident is not using the seat the regulation gave it.
Where we stand
We sell a detector and a humanizer, so weigh this page accordingly, and none of it is legal advice. Every figure above is quoted from UGC’s or INFLIBNET’s own published documents and linked below. Where we could not verify something — a UGC AI threshold — this page says we could not find it rather than filling the gap.
If you have been accused, the general guidance applies: what to do next and how to prove you wrote it. Draft history is the strongest evidence there is, and it is worth keeping before you need it.
Related
- The UK equivalent — where no regulator requires detection at all
- Are AI detectors accurate?
- SafeAssign — the same gap, in a product