Is AI Allowed in Academic Writing? What 2026 University Policies Actually Say

Yes, AI is allowed in academic writing at most journals, most universities, and most funders in 2026 — as an assistant, with disclosure, and with you carrying full responsibility for every word. What is banned nearly everywhere is the other thing people mean by the question: AI as ghost-author, undisclosed, producing text you submit as your own.

That two-sentence answer hides a genuinely messy landscape, because “allowed” is decided at three separate layers — the venue you’re submitting to, the institution you belong to, and (for grant work) the agency funding you — and the layers don’t always agree. This article maps what the major policies actually said as of July 2026, with the standing caveat that governs everything below: policies in this area get rewritten mid-semester. This is a description of the current rules, not advice; check your institution’s and venue’s current policy before you rely on anything here.

How we got to 2026: from bans to disclosure regimes

The trajectory matters because most of what you’ll hear secondhand is a policy from an earlier phase. In early 2023, the reflex was prohibition — Science editor H. Holden Thorp’s January 2023 editorial (“ChatGPT is fun, but not an author”) banned AI-generated text outright. By late 2023 the bans were already collapsing under unenforceability, and Science itself reversed to a disclosure model in November 2023. The 2024–2025 phase was convergence: every major publisher landed on some version of “use it, disclose it, own it.” The 2025–2026 phase is enforcement — Springer Nature has retracted articles at scale for undisclosed AI use, and the US funding agencies moved from silence to explicit rules with teeth.

So the question in 2026 is rarely whether you may use AI. It’s which uses your specific venue requires you to disclose, and where the disclosure goes.

Publisher policies: the big four, verified

Elsevier. Generative AI may be used “to improve the language and readability” of your paper — that framing is Elsevier’s, and it’s deliberately narrow. Use must be disclosed in a dedicated AI declaration section placed before the references (a requirement tightened in a September 2025 update). AI cannot be listed as an author, and AI-generated images are prohibited in scientific content. Elsevier also bars reviewers from uploading manuscripts into AI tools — a confidentiality rule, not a quality rule, and one authors should remember cuts both ways when they’re the reviewer.

Springer Nature. No AI authorship — the stated reason, which recurs across every policy in this article, is that an LLM “cannot be held accountable for the work.” Substantive AI use must be documented in the Methods section (or nearest equivalent). The notable nuance: AI-assisted copy editing — grammar, spelling, punctuation — does not require disclosure. Springer Nature is the clearest of the big publishers in drawing that line, and it’s the line most researchers actually care about. Generative AI images remain prohibited apart from narrow exceptions (AI-focused research, properly licensed agency images).

IEEE. The Publication Services and Products Board Operations Manual (§8.2.1.B.10) requires that AI-generated content “shall be disclosed in the acknowledgments section,” naming the system and identifying which sections used it. Like Springer Nature, IEEE carves out routine assistance: AI for “editing and grammar enhancement” is “generally outside the intent” of the policy, with disclosure recommended rather than required. No AI authors, and no feeding manuscripts under review into public AI tools.

Science / AAAS. Having swung furthest toward prohibition, Science now runs the most demanding disclosure regime of the four: AI use must be flagged in the cover letter, noted in the acknowledgments, and detailed in the methods or supplementary materials — tool, version, and which parts of the manuscript were affected. AI-generated imagery is banned by default, with case-by-case editor exceptions. And Science states the stakes plainly: undisclosed AI use is scientific misconduct, the same category as fabrication.

Notice what varies and what doesn’t. The disclosure location differs by publisher — declaration section (Elsevier), methods (Springer Nature), acknowledgments (IEEE), all three (Science). The principles don’t vary at all.

University policies: three real patterns

Institutional rules govern coursework, theses, and dissertations, and in 2026 they have mostly settled into an instructor-discretion architecture with a default. Three verified examples spanning the range:

Stanford (following its Academic Integrity Working Group’s 2025 recommendations) applies a human-collaboration standard: absent explicit permission from the instructor, AI assistance on an assignment is treated like assistance from another person — which is to say, prohibited. That’s the strictest common default, and it makes the syllabus, not the university handbook, the operative document.

Cambridge distinguishes by assessment type: AI for personal study is acceptable, but unacknowledged AI-generated content in summative work is academic misconduct. Oxford runs the same structure — permitted for learning, prohibited in assessed work without explicit permission plus declaration.

Harvard delegates almost entirely: course policies legitimately range from full prohibition to permitted-with-disclosure, and the university-level rule is essentially “follow your instructor’s rule.” Columbia, after a more turbulent policy process, landed on prohibition absent instructor permission; Caltech, similarly, permits AI “only for tasks explicitly allowed by their instructors.”

The practical upshot for students is uniform across all of these: the syllabus outranks the handbook, and silence means no. If your instructor hasn’t stated a policy, the safe reading everywhere is that undisclosed AI assistance on graded work is a violation — ask before, not after.

Thesis and dissertation writing sits in an awkward middle: it’s institutional work governed by graduate-school rules, but many universities have added AI-use declarations to submission paperwork. If you’re mid-dissertation, this is the single policy most worth looking up this week.

Funders: NIH and NSF now have explicit rules

Grant writing got its own policy layer in 2025, and it’s stricter than the publishers’.

NIH issued NOT-OD-25-132, effective for applications received on or after September 25, 2025: applications “substantially developed by AI, or containing sections substantially developed by AI” are not eligible for review. The same notice caps investigators at six new/renewal/resubmission/revision applications per calendar year — a limit NIH tied explicitly to AI-enabled mass submission (only about 1.3% of 2024 applicants exceeded it). Note the operative word substantially: NIH is not banning AI-assisted editing; it is banning AI-drafted science. On the review side, NOT-OD-23-149 (2023) prohibits NIH peer reviewers from using generative AI to analyze or write critiques — uploading an application to a chatbot is a confidentiality breach.

NSF has been in a similar place since its December 2023 notice: reviewers may not upload proposal content into non-approved generative AI tools, and proposers are encouraged — not yet required, as of this writing — to indicate in the project description how generative AI was used. Grant text you submit is your intellectual claim on future funding; both agencies treat AI-fabricated or AI-padded proposals as an integrity problem, not a productivity innovation.

The three principles every policy shares

Strip away the venue-specific details and every policy above — publisher, university, funder — rests on the same three commitments:

1. AI cannot be an author. Authorship carries accountability, and a model cannot be accountable, sanctioned, or asked to defend a claim. No exceptions exist at any major venue.
2. Meaningful use must be disclosed. Where “meaningful” starts varies (Springer Nature and IEEE exempt grammar-level assistance; Science wants nearly everything), but concealment is the offense that triggers retractions, not use.
3. The human authors own every error. A hallucinated citation, a fabricated statistic, a subtly wrong paraphrase — if it’s in your manuscript, it is your misconduct, and “the AI wrote that part” is an aggravating admission, not a defense.

If you internalize those three, you can navigate any venue’s paperwork; the rest is formatting.

What disclosure actually looks like

Policies demand disclosure; few show you one. Two templates that satisfy the requirements above, to adapt rather than copy:

For an Elsevier-style declaration section or IEEE acknowledgment:

Declaration of generative AI use in the writing process. During the preparation of this work, the authors used Claude (Anthropic, Claude Sonnet 4.5) to improve the clarity and language of the manuscript and to suggest restructuring of Section 3. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

For a methods section where AI touched analysis or drafting more substantively:

Portions of the literature synthesis in Section 2 were drafted with assistance from GPT-5 (OpenAI, March 2026 release) from author-written outlines; all cited sources were located, read, and verified by the authors. No AI-generated text was included without author revision.

The pattern in both: name the tool and version, state what it did and where, state what you did afterward, claim responsibility. A disclosure missing that last element reads, to an editor, like an attempt to share blame with software.

A decision tree you can actually use

Before any AI-assisted writing session on academic work, four questions in order:

1. What layer governs this document? Coursework → syllabus, then university policy. Manuscript → journal’s author guidelines (this month’s version). Grant → agency notice. Thesis → graduate school regulations.
2. Does the policy exist and what’s its default? Explicit permission → follow its disclosure format exactly. Explicit prohibition → don’t, full stop. Silence → treat as prohibited for assessed/submitted work and ask; every misconduct case that makes the news starts with someone resolving ambiguity in their own favor.
3. Is this use above or below the disclosure line? Grammar, spelling, and phrasing polish is below the line at most venues (Springer Nature and IEEE say so outright). Generating text, summarizing literature, restructuring arguments, drafting from outlines — above the line everywhere. When unsure, disclose; no paper has ever been retracted for over-disclosure.
4. Can you personally verify every AI-touched claim and citation? If not, the policy question is moot — you have an accuracy problem that disclosure won’t fix. Every reference an AI produces gets checked against the actual source before submission, no exceptions; how we operationalize that checking is in our testing methodology.

For the tool-level question — what current chatbots are actually good for at each stage of a manuscript, and where they reliably fail — see our guide to ChatGPT for academic writing.

The detection-tool paragraph, honestly

One paragraph, because it deserves exactly that. AI detectors remain unreliable enough that policy built on them is policy built on sand: Vanderbilt disabled Turnitin’s AI detector back in 2023 over false-positive concerns and never re-enabled it, more institutions have quietly followed since, and the false positives are not randomly distributed — the best-known finding in this literature (Liang et al., Stanford, 2023) showed detectors disproportionately flagging non-native English speakers’ unassisted writing, a bias subsequent tools have narrowed but not eliminated. As of 2026, no detector’s output is proof of anything; responsible integrity offices treat scores as a prompt for conversation, not a verdict. For writers the implication is double-edged: a clean detector score doesn’t make undisclosed use safe, and an accusation based solely on a score is challengeable — process, drafts, and version history are your actual defense.

Where this site stands

Our own rule, applied to everything published here and stated on the About page: assistance yes, ghost-authorship no, disclosure always. AI tools are used the way this article describes using them — as editors, checkers, and drafting assistants whose every output gets verified — and never as the uncredited author of a verdict or a claim. We hold the tools we review to a standard of honest failure; it would be strange to hold ourselves to less.

The policies will keep moving. The three principles underneath them — no AI authors, disclose meaningful use, own your errors — have been stable through every revision since 2023 and are the safest long-term bet for how to work. Verify the specifics for your venue this week, not from a blog post dated last year. Including, in a year, this one.