Drupal AI Content Review Alpha Can Overwrite Entire Fields
Drupal editors can use the contributed AI Content Review module to score pages against their site’s own editorial rules. The module also offers an Apply button for suggested rewrites, but tests by The DropTimes found that the latest tagged alpha can replace an entire text field instead of the passage selected for correction.
In the current tagged release, 1.0.0-alpha3, clicking Apply replaces the entire target field rather than only the quoted passage. In testing by The DropTimes on 10 and 14 October 2026, with suggestions that named the field's machine name, whether the earlier text survived depended on content moderation. On a content type without it, Apply overwrote the page's current revision without creating a new one, so the earlier text was not preserved in a separate revision. On a moderated content type, the original text was kept in an earlier revision.
A change on both of the module's development branches, 1.x and 1.0.x, narrows Apply to the quoted passage and forces a new revision, though no tagged release includes it yet. On both branches, Apply still saves the page as soon as it is clicked.
In a roadmap issue opened on 7 October 2026, David Galeano of Factorial.io, one of the module's maintainers, describes the behaviour as one where applying a suggestion "auto-saves the entity (creating unwanted revisions and bypassing moderation)".
The Site Writes the Rubric and Sets the Pass Mark
The module is organised around review rules that apply to one entity type and to selected bundles, such as standard pages or news articles. Each rule holds one or more criteria, and each criterion checks a single aspect of quality, such as tone of voice or inclusive language.
A criterion combines an AI agent with a written rubric, entered under Guidelines & rules. The module's documentation advises writing it "the way you'd brief a new teammate: what to look for, and what to flag". Several criteria can share one agent.
Each criterion also has a passing score and an optional warning score. The agent returns a score from zero to 100, and the module compares it with these thresholds to show a pass, a warning or a fail. According to the documentation, the thresholds "are not sent to the AI", so changing the pass mark regrades stored reviews without calling the AI again.
Site builders can also teach the agent with scored examples grouped by aspect, with at least two per group. The documentation recommends examples that span the scale: one at 90 or above, one at 30 or below, and one near the passing score.
A review ends as complete when it produces a score, skipped when the content gives the agent too little to judge, or failed when the agent cannot do the job or its reply cannot be read. Whether a page with too little content counts as skipped is the agent's decision, not a check in the module. In testing by The DropTimes on 10 October 2026, on 1.0.0-alpha3 with OpenAI's gpt-4o-mini, a simple agent scored a page with an empty main text field at 50, with three suggestions. That run checked the mechanics, not the quality of the review.
An Agent's Instructions Can Send Any Field to the AI Provider
What the AI provider receives depends on how the site sets up its agent. The documentation's example agent instructions insert the page's title and body through Drupal tokens. During a review run, the module adds the criterion's rubric and examples, plus the page's type, bundle, title and non-computed field machine names. Because the module provides the page as a token context, the agent's instructions can insert additional fields, not only the title and body, subject to the available tokens and access permissions.
Site owners should check what their own agent sends and where their chosen provider processes it.
Recommendations Become a Checklist, but Apply Saves at Once
Editors with the Run AI content reviews permission get an AI Review tab on each page a rule covers. Each criterion runs separately and shows the score, a severity and the agent's explanation, with a numbered list of recommendations.
Each recommendation can be accepted or ignored. That decision only records what the editor intends to do and, in the documentation's words, "never touches the reviewed content". Accepted recommendations become an improvement plan, a checklist in the sidebar of the page's edit form, which the editor works through by hand.
Recommendations that propose replacement text also carry an Apply button. Apply writes the suggestion into the field and saves the page at once, without going through the edit form.
In the Released Alpha, Apply Overwrites the Whole Field
Each suggestion stores the original passage it is meant to replace, but the 1.0.0-alpha3 release does not search the field for that passage and does not explicitly request a new revision before saving. Apply also goes ahead when the quoted passage is not on the page: it reports the suggestion as applied and replaces the whole field.
In testing by The DropTimes, the result depended on whether the suggestion named a valid field machine name and whether content moderation was enabled:
| Test setup | What Apply did | Was the earlier text kept? |
|---|---|---|
No content moderation; the suggestion named the body field machine name; the content type was set to create a new revision on every save | Replaced the whole field in the page's current revision and cleared the field's text format. A second Apply of the same suggestion was refused with HTTP 409. | No. No new revision was created, so the earlier field text was not preserved in a separate revision. |
Content moderation; the suggestion named the body field machine name | Created a new revision, though the new revision's whole field was still replaced and its text format changed. | Yes, in the earlier revision. |
With or without content moderation; the suggestion named the human-readable label Body rather than the machine name body | Rolled back the save and returned HTTP 500, because the node had no Body field. The browser showed no error message. | The page's revision was left unchanged. |
The tests used AI Content Review 1.0.0-alpha3 on Drupal 11.4.9, with and without content moderation, and on Drupal CMS 2.2.3 with content moderation and OpenAI's gpt-4o-mini.
The revision difference between moderated and unmoderated types comes from how Drupal handles saves made in code. A content type's Create new revision setting does not, by itself, force every programmatic save to create a revision, whereas content moderation caused new revisions to be created in the tested configuration. The project's own issue on Apply, described below, says the same of the site it was found on: "a revision was only created because content moderation forces one".
On a launched copy of the Northmoor University Drupal AI Demo, Apply also returned a server error, showed no message and left the page's single revision unchanged. The cause in the demo copy was not established.
The Fix Is on Both Development Branches but Not Yet Released
Eric Homanchuk of Foster Interactive reported the whole-field behaviour in the project's GitLab work-item tracker on 18 September 2026 (issue #3585854). Ahmad Khader of Vardot, one of the module's maintainers, committed a fix to the module's 1.x development branch on 5 October 2026 as commit 9a4d89b0, and the issue was closed as fixed the same day.
Artem Dmitriiev of 1xINTERNET, the project's creator, added the same change to the 1.0.x branch, from which 1.0.0-alpha3 was tagged, as commit 32c692f3 on 15 October 2026.
The code on both branches makes Apply replace only the quoted passage, and only when it appears exactly once in the field; the rest of the field and its text format are left as they were. In testing by The DropTimes on the 1.x branch at commit 9a4d89b0, every successful Apply created a new revision, even on a content type set not to create new revisions, with the log message "Applied AI Content Review suggestion #" followed by the suggestion's number.
A suggestion whose quoted text is missing, or appears more than once, is refused and the page is left unchanged, while its button changes to Not applied and is disabled. When the field is empty and the suggestion quotes nothing, the suggested text fills it.
With content moderation, the author's tests found that Apply on a draft saved a new draft, while on a published page it saved a new published revision that went live immediately. When a published page also had a newer pending draft, Apply worked on the published text. A passage found only in the draft was refused. When a passage appeared in both versions, Apply changed the published text and created another revision, after which the draft's edit was no longer part of the page's latest revision. This does not establish that the earlier draft revision was deleted.
Issue #3585877, David's roadmap, sets out the steps towards a stable 1.0.0. They include removing the automatic save after a suggestion is applied, storing suggestions as revisions and adding a review history with rollback. A separate issue, #3585856, reports that Apply, run from the edit form, leaves that form out of date, so the editor's next save is rejected.
Feature request #3585879 was filed on 9 October 2026 by Levente Besenyei of Brainsum, a Silver partner on the Drupal AI partners page. It asks for a setting that applies a replacement as soon as an editor approves it, citing a custom module built for the VisitEurope site.
The request defines approval narrowly: "Here, approved means accepted by a human editor, not merely generated or validated by AI." It adds that "Approval accepts an edit to a draft" and that "Configuration does not bypass permissions or moderation transitions." These are requirements proposed in the feature request, not functionality included in the released alpha.
Every Run Is Recorded, and Bulk Reviews Can Spend Provider Budget
A review belongs to one revision of a page, and the alpha3 code offers no way to force a fresh review of an unchanged revision. A criterion runs again when the page is saved as a new revision, when its rubric, examples or agent settings change, or when the earlier review failed. The module identifies a review by the SHA-256 hash of the criterion's configuration and the page's revision ID, not by hashing the page's text. Thresholds are kept out of that check.
According to the documentation, every run of a criterion creates an append-only review record that stores the criterion, who started the review, which revision was reviewed, the AI provider and model and the full result.
Reviews also run in bulk, behind a restricted permission whose description warns that "one click here can spend a large amount of AI provider budget". A separate permission lets a role watch the queue without initiating reviews and incurring provider costs. The module adds an AI Review column to Drupal's content list, a filterable listing, bulk queuing and optional daily cron re-reviews. It also adds a summary to the top bar of each page, listing each criterion's run and pinning each suggestion to the sentence it quotes.
Drupal AI Partners Are Among the Project's Supporting Organisations
The project was created on Drupal.org on 26 March 2026, and its page lists four supporting organisations: 1xINTERNET, Factorial.io, Foster Interactive and Vardot. Of the four, 1xINTERNET is one of the founding partners of the Drupal AI Initiative, according to the initiative's partners page, where Vardot and Foster Interactive are listed as Gold and Silver partners.
The Alpha Package Declares AI Agents as a Required Dependency
The module works with Drupal 10.5 or later in the Drupal 10 series, and with Drupal 11.2 or later. The 1.0.0-alpha3 package metadata and Composer file declare the AI module, AI Agents and the Entity API module as required dependencies. The Drupal.org project page, by contrast, states that AI Agents is needed only for the default agent-based criterion.
The default criterion also requires a configured AI provider that supports tool-calling chat, and only agents configured with the module's AI Content Review Result tool can be selected.
No Release Is Covered by Security Advisories Yet
AI Content Review is in development. All three of its tagged releases are alphas: alpha1 on 17 September 2026, alpha2 on 23 September and alpha3 on 25 September. A 1.x-dev snapshot sits alongside them.
The project has opted in to Drupal's security advisory coverage, but the security advisory policy excludes development, alpha, beta and release candidate versions, so no release is covered yet.
According to the documentation, the module can be extended with custom criterion and suggestion plugins without changing its code, and issues and contributions go through its issue tracker on git.drupalcode.org.
Until a tagged release ships the passage-level replacement and revision changes now on both development branches, editorial teams testing 1.0.0-alpha3 can use the review scores and the sidebar improvement checklist while reserving the Apply button for non-production sandboxes. The project's release history on Drupal.org shows when a tagged release includes that change.
Disclosure: The DropTimes was a media sponsor of DrupalCon Rotterdam 2026, where 1xINTERNET, Vardot and Factorial.io were also sponsors.


