How to Cite AI Use in Academic Writing (APA, MLA, Chicago, 2026)
Citing AI use in academic work is no longer optional in most institutions, but the rules are still settling. The three major style guides, APA, MLA and Chicago, all now say to disclose AI assistance, and a growing list of universities, including Oxford from Trinity Term 2026, require a formal AI statement. The catch is that your university's policy overrides any style guide, so the first step is never the manual, it is your course profile or thesis regulations. Below is the 2026 state of play and the format each guide expects.
Rule zero: your institution's policy beats any style guide
Before you format a single citation, check what your school actually requires. Oxford's Trinity Term 2026 policy mandates a formal AI-use declaration placed immediately after the abstract, naming the tool and version, the publisher, a brief description of use, and a compliance statement, plus the ability to defend that use in your viva. The University at Buffalo requires all graduate programs to publish discipline-specific AI policies by fall 2026.
These binding documents sit above APA, MLA and Chicago. A citation formatted perfectly to a style guide can still fail if it omits what your institution specifically demands, so read the local policy first and treat the style guide as the formatting layer on top. The detector most campuses already run, Turnitin, is also the system whose AI indicator shapes how policies are written, which is why its output shows up in integrity discussions more than any other tool.
APA 7th: credit the company, never the chatbot
APA's current guidance is clear that a generative AI tool cannot be named as an author, because it cannot take responsibility for the work. Instead, the company that developed the tool is treated as the author and the specific tool as the source. A reference for a specific chat reads, for example, OpenAI (2026, January 10). Methods of providing alternative text for accessibility on websites. [Generative AI chat]. ChatGPT-5. https://chatgpt.com/share/... .
APA also asks you to describe in the methods section how, when and to what extent AI was used, and to upload the full AI output as supplemental material for publications. Student-facing tools such as Scribbr, which frames its detector around responsible and transparent use rather than evasion, are a useful companion here: they help you understand what a checker will see before you submit, which is different from and complementary to proper citation.
MLA and Chicago: the prompt is the title
MLA grants no author credit to the AI. You use a short description of the prompt as the title, the tool name with version as the container, the developing company as publisher, and the generation date as the date, with a public URL as the location when available. Chicago treats the tool with its version as author and the company as publisher, handled in numbered notes or author-date parentheticals, adding a bibliography entry only when a public URL exists.
One warning that applies to every style, and especially to international and multilingual writers: detectors systematically misflag careful, conventional prose, and the false-positive rate climbs on non-native English text. If your cited draft is in a second language or a formal register, prefer a multilingual-aware checker such as Copyleaks for any self-review, and keep the citation honest regardless of what a score says.
Don't let the AI invent your references
The fastest way to fail an integrity check is not hiding AI use, it is trusting AI-generated citations. A Columbia University audit of more than 2.5 million peer-reviewed papers found AI-generated references rose roughly twelvefold in three years, reaching one in 277 papers containing at least one hallucinated citation in early 2026. These phantom sources look plausible and pass casual reading.
The fix is mechanical, not moral: verify every reference an AI suggests against the primary source before it enters your bibliography. Originality.ai and similar combined detectors pair AI scoring with plagiarism checking, which is the layer that catches fabricated or misattributed references, but no tool removes the need to open the source yourself. Cite the underlying study directly rather than citing the chatbot's summary of it.
Run a self-check before you submit
Disclosure and detectability are separate questions. A draft you have cited correctly can still read as machine-written to a statistical checker, and a hidden draft can still pass. Before submission, run a free, low-false-positive self-check so you are not surprised by what a reviewer sees, and so you can tighten any passages that read uniformly artificial.
GPTZero is the educator-and-student default for this: its sentence-level highlighting shows exactly which lines a model would flag, which is more useful than a single percentage when you are polishing a near-final draft. Use the result to edit, not to decide whether to disclose, because the disclosure decision is governed by your institution's policy, not by the score.
Oxford's mandatory statement and the formal-documentation trend
The direction of travel is toward explicit, auditable disclosure. Oxford's Trinity Term 2026 requirement is the strictest public example: a declaration after the abstract naming tool, version and publisher, plus the ability to give a critically reflective account of your use in the viva. Expect more institutions to adopt similarly structured statements through 2026 and 2027.
For thesis and formal-submission writers, the habit to build is documentation. Tools such as Winston AI that produce dated, defensible reports of any AI check are useful here not for evasion but for the paper trail, letting you show exactly what was scanned, when, and by which engine. Pair that with your citation and your transparency statement, and the submission withstands both a detector and a human reviewer.
Bottom line
Start with your institution's policy, because it overrides APA, MLA and Chicago. APA 7th names the company as author and the tool as source, discloses use in the methods section, and uploads the full output; MLA and Chicago treat the prompt as the title and footnote the rest. Verify every AI-suggested reference against the primary source, since hallucinated citations now appear in roughly one in 277 papers. Run a low-false-positive self-check such as GPTZero before submitting, and treat disclosure as a separate duty from detectability.