Plagiarism and AI at Harvard: lockdown yes, detection no

The short answer

Harvard’s Faculty of Arts and Sciences buys software to stop AI being used in exams, and refuses to buy software that claims to detect it afterwards. That contrast tells you most of what you need to know about how a question would be handled.

There is also a rule in your favour that few universities have: your work cannot be put into an AI tool without your documented permission.

Which Harvard this is

A necessary caveat before anything else. Everything on this page comes from the Office of Undergraduate Education in the Faculty of Arts and Sciences — that is Harvard College and FAS courses.

Harvard Extension School, Harvard Law School, Harvard Business School and the other faculties set their own rules. If you are not in FAS, treat this as context rather than as your policy, and find your own school’s guidance.

The guidance page itself carries no effective or last-revised date. The events it links are dated August 2023 and its footer copyright reads 2026, which brackets it loosely but not precisely.

Prevention, not detection

The refusal is published as an FAQ answer, and it is worth reading in full because of how it reasons:

“There are a variety of tools that claim various degrees of success in finding instances when GAI was used, but this is something of an arms race. It would be inadvisable to count on automated methods for GAI detection. FAS does not plan to provide/license such a tool for use in courses.”

The argument is not that detectors are useless today; it is that the contest between generation and detection is not one detection can settle. FAS makes the same refusal about exam-blocking software of the invigilating kind: “While there are technologies for exam proctoring that aim to block students from using the internet, FAS does not plan to provide/license such a tool for use in courses.”

And yet, in the same guidance, it announces the opposite for exams:

“Starting in Fall 2025, faculty can find Respondus, a new browser lockdown tool, on their Canvas site(s) to use for in-person seated exams and quizzes to ensure that students do not use AI unless the course asks them to do so.”

Put together, the position is coherent: control the conditions of assessment in advance, rather than trying to classify the output afterwards. For a student that means the pressure point moves to the exam room, and away from a percentage attached to your coursework.

It is the same conclusion we reach about our own tool, from the other direction: our published error rates are measured on English and German corpora, and no detector — ours included — produces proof. See how accurate AI detectors really are. If a concern is ever raised about your work, what answers it is your process — see how to prove you wrote it yourself and what to do when you are accused.

If a detector score is what started this, our guide to an AI detector flagging your thesis sets out what that number does and does not establish on its own.

Your permission is required

This is the most under-reported thing on Harvard’s page, and it is a real protection:

“No confidential information can be loaded into GAI systems, since there is no expectation of privacy or confidentiality. Faculty must get documented permission from students before putting original student content into any generative AI tool, and students should be made aware of the risks of entering their own original work into such tools.”

So your essay cannot be pasted into a chatbot for marking, summarising or checking without you agreeing to it — and the permission has to be documented, not assumed.

The second half is addressed to you, and it is good advice anywhere: think before entering your own unpublished work into a service, because there is “no expectation of privacy or confidentiality”. Before you upload a draft to anything, find out whether it is kept. For what it is worth, our own check does not add your text to any third-party comparison database.

The three model course policies

FAS sets no single AI rule. It requires each course to have one and to say so: “All faculty are required to inform students of the policies governing generative AI use in class.”

It publishes three models instructors can adopt:

• Maximally restrictive: “We specifically forbid the use of ChatGPT or any other generative artificial intelligence (AI) tools at all stages of the work process, including preliminary ones. Violations of this policy will be considered academic misconduct.”
• Fully encouraging: “This course encourages students to explore the use of generative artificial intelligence (GAI) tools such as ChatGPT for all assignments and assessments. Any such use must be appropriately acknowledged and cited.”
• Mixed: “Certain assignments in this course will permit or even encourage the use of generative artificial intelligence (GAI) tools […] The default is that such use is disallowed unless otherwise stated. Any such use must be appropriately acknowledged and cited.”

Three things to take from that. The restrictive model reaches preliminarystages — brainstorming and outlining included, not just drafting. Both permissive models require acknowledgement, so permission never removes the citation duty. And the mixed model’s default is no, which is the safest assumption when an assignment is silent.

All three close with the same warning: “different classes at Harvard could implement different AI policies, and it is the student’s responsibilityto conform to expectations for each course.” One course’s permission tells you nothing about the next.

No university-wide citation format is specified. Where acknowledgement is required, our AI disclosure statement guide covers what a good one contains.

What we could not read

We want to be straight about the limits of this page, because they are larger here than elsewhere in this series.

Harvard’s Honor Council and FAS student handbook sites returned HTTP 403 to us — they blocked automated access rather than not existing. That means we could not read, and therefore will not describe:

• The current text of the Harvard College Honor Code, or its revision date.
• The sanctions scale.
• The Honor Council’s composition, its procedure, the appeal route, or the standard of proof.
• The Honor Council’s annual report — the address responds, so the report exists, but we cannot quote a figure from it.

Blocked is not the same as absent. All of this is very likely published and readable from inside Harvard, and if any of it matters to you, that is where to look. What we can say is where a course AI breach goes: violations of a course policy “will be considered academic misconduct”, which routes to the Honor Council.

No similarity vendor is named on the reachable pages, and no student self-check route or retention statement is documented.

Before you submit

• Find the AI policy for that course. FAS requires one, and they differ course to course.
• Treat silence as no. That is the published default in the mixed model.
• Acknowledge permitted use. Both permissive models require citation.
• Remember preliminary stages count under the restrictive model — outlining is use.
• You can decline to have your work put into an AI tool. Documented permission is required.

If you want a read on your own text before you hand it in, our AI check is free up to 1,500 words, with no account and no name.

Sources

• Harvard Faculty of Arts and Sciences, Office of Undergraduate Education: Generative AI guidance — the detection refusal, the Respondus announcement, the consent requirement and the three model policies. Page carries no effective or last-revised date.
• Harvard College Honor Council — honorcouncil.fas.harvard.edu, including its annual reports. Not publicly retrievable at the time of writing.

This page is orientation, not legal advice, and it covers the Faculty of Arts and Sciences only. What binds you is your course policy and your own school’s regulations.