Which AI Model Version Do You Put in a Citation?

A student cites ChatGPT for a paragraph drafted in March. In June, a supervisor reruns what looks like the same prompt and gets a noticeably different answer. Nothing in the citation says whether that difference is normal variation or a sign the paragraph was reworded after submission, because the entry never named which system produced the March text in the first place.

Which model version do I cite? Whatever the interface actually displayed at the moment it produced the text, copied exactly rather than summarized from memory. A generative AI system is not one fixed, stable thing. Providers update, retire and quietly swap the model behind a product name on a schedule no reader can be expected to track, and the version string is the one detail that pins your citation to a specific system rather than a moving target.

How to cite ChatGPT covers the full reasoning for why that matters in one paragraph, alongside the rest of what a ChatGPT citation needs. This piece stays on the version element by itself: where to actually find it, what to do when an interface will not give you a clean one, and how the major styles differ on where it goes.

Which Model Version Do I Cite? Start by Checking the Interface

The most common way for chat interfaces to name the model is to display its name somewhere on-screen. This might be a picker above the message box, a menu in settings, or a little label near the response itself. Open that control before you start writing your paper's method section, not after. The label can default to a different model in a new conversation than the one you used last week. Copy the label exactly as shown, including any qualifier attached to it, rather than shortening it to the product's general name.

Precision matters because these labels distinguish real, functionally different systems, not cosmetic variants of one another. Writing down the product name alone, ChatGPT, Claude, Gemini, when the interface actually showed something more specific is the same mistake as citing a book by its series title and leaving out which volume you read. A reader trying to interpret your method cannot tell a fast, lightweight response setting from a slower, more thorough one if your citation only ever gives the product name.

When the Interface Does Not Show a Clean Version Number

Some interfaces do not show a version number at all. Instead of a string built from a model name and a release number, the control shows a speed or effort setting, closer to a dial than a label: a quick response mode sitting next to a slower, more thorough one, with the underlying model name tucked away in a settings page or not exposed to the user at all. Record whatever label the tool does show, exactly as written, even if it reads like a setting rather than a version.

When nothing on screen names a system at all, a date does the same job a version number would have done. Citing a specific day, or the day and time for a fast-moving product, tells a reader which window of that product's history your text came from, even without a model name attached to it. That is the same logic every current style guide already applies to a shared chat link or a generation date, carried one step further for the case where a label was never on offer.

How the Major Styles Handle the Version Element

The style bodies that have addressed generative AI at all do not agree on whether the version gets its own slot in the citation, gets folded into the model name, or gets left out in favor of a date doing the same job. The table below covers only that one element, not a full worked citation for any of them.

Style	Dedicated version element?	What actually goes there
APA	No	The model name itself, with no separate version number attached
MLA	Yes	A dedicated Version core element, naming the specific model as precisely as the interface allows
Chicago	No	No version field in the template at all; the date of the exchange anchors it instead
IEEE	No format published	A disclosure statement naming the AI system, with no citation template to place a version in
Harvard (Cite Them Right)	No	No version slot in the template; the query and the AI's name identify the exchange instead
Vancouver / ICMJE	No format published	A disclosure statement describing how the tool was used, not a reference-list version field

How to cite ChatGPT in APA and How to cite ChatGPT in MLA both cover their full reference format in detail. What matters here is narrower: APA folds the version into the model name rather than giving it a separate slot, while MLA keeps Version as its own core element. Both are complete, deliberate answers to the same design question.

Recording the Version Once You Have It

Write the version string exactly as the interface showed it, in the same citation entry where the date and the tool name go, and keep a screenshot or a copy of the exact text alongside your notes if your course wants evidence beyond the citation itself. An AI citation generator that presets multiple model families turns that string into the right field automatically, once you already know which style you are writing in and which label you actually saw.

Search results raise the same versioning problem in a harder form, and citing a Google AI Overview is treated separately. Whether citing ChatGPT counts as plagiarism comes earlier than any formatting decision, and it is answered on its own.

None of this changes once a reference is confirmed accurate rather than invented. A citation generator built from real publisher records still handles every ordinary source in the same reference list without needing a version field at all, since a journal article does not update itself overnight the way a chat model does. The version string is what an AI citation is asking for instead, in the one field a book reference never needed.