On July 16, 2026, the American Library Association’s governing Council adopted a new Guidance on the Use of Artificial Intelligence in Libraries. Most coverage will file it under library policy, although much of it works as a procurement specification.
The guidance’s privacy section contains a ten-item vendor review checklist. Before adopting any AI-enabled system, libraries are asked to determine whether prompts and searches are logged, whether that data trains the vendor’s models, where it is stored, how long it is kept, and what audit or exit rights the library actually has. These questions are familiar as concepts; the change lies in who is now expected to answer them, and in what form.
In short
ALA’s guidance turns AI vendor accountability from a values statement into a ten-question vendor review covering logging, training use, storage, retention, deletion and audit or exit rights. For AI vendors, discovery platforms and the publishers whose content flows through them, most of these questions can currently be answered with a policy sentence, and very few with a record a library could audit. That gap is the decision point for anyone selling AI-enabled products to libraries this year.
A procurement checklist rather than a values statement
Library AI ethics documents already exist. ALA’s guidance differs in where it sits: inside the vendor review section, next to the procedural questions a library or consortium works through before signing a contract.
The guidance applies existing ALA vendor privacy standards to AI systems and keeps patron data, institutional data and usage logs under library control. It states that libraries should not have to give up control of their data to gain access to AI functionality. It then lists what a vendor review should establish: default settings, what data is collected, whether prompts and account activity are logged, whether data trains the model, where it is stored, which subprocessors handle it, how long it is retained, how it is deleted, and what audit or exit rights apply. A procurement officer can copy that list directly into a request for proposals (RFP).
The guidance also covers what libraries should do when they cannot determine how a system collects, retains or deletes data: delay, limit, refuse or discontinue that use. It presents non-adoption as a legitimate professional decision rather than a failure to modernize.
From ten questions to one gap
Most vendors selling to libraries can already answer these ten questions in prose, through a privacy policy, a data processing addendum or a sales call. Those answers are declarations.
An audit means that the library, or a third party acting for it, can check the answer against a record: which prompt touched which content, under which right, how long it was retained and on what date it was deleted. A declaration states what should be true, while an audit shows what was true for a specific interaction on a specific date. The table below maps each item in ALA’s checklist to that distinction as it typically applies today.
ALA vendor review item | Typical vendor answer today | What an audit actually requires |
|---|---|---|
AI features enabled by default | Stated in onboarding docs | Configuration log, timestamped |
Can the library disable them | Stated in admin settings | Change log showing the toggle was used |
What data the system collects | Privacy policy summary | Data schema tied to actual collection events |
Are prompts, searches, activity logged | General statement of practice | Per-interaction log a library can sample |
Used for model training | Opt-out clause in the contract | Record showing the opt-out was applied, not just offered |
Where data is stored | Region named in the DPA | Verifiable storage location per record |
What subprocessors are used | List in a subprocessor page | List kept current and tied to actual data flows |
Retention length | Stated retention window | Proof that deletion actually occurs on schedule |
How data can be deleted | Deletion process described | Confirmation log per deletion request |
Audit or exit rights | Clause granting the right | A right the library has actually exercised or could exercise today |
Seven of the ten items are usually covered by policy language a vendor already has on file. The remaining three require an operational record rather than a statement of intent: what was logged, whether that log fed model training, and what a library could inspect if it exercised its audit right.
A procurement committee familiar with the ALA guidance is most likely to press on those three, because the guidance names them explicitly instead of leaving them as standard clauses.
The objection: this is guidance, not a mandate
ALA guidance has no legal force. No library is required to adopt it and no vendor is required to comply.
The objection is correct and changes little in practice. Non-binding guidance becomes a de facto specification once it is copied into RFP language, consortium contract templates or school-district procurement policy, the route earlier ALA privacy and accessibility guidance has typically followed. The guidance anticipates this by committing ALA to advocacy with vendors, since individual libraries cannot create that pressure alone. A library board that declines a renewal because a vendor could not answer the retention question does not need a law to do so.
What changes for vendors and publishers
Two groups face this checklist directly. The first is AI vendors selling discovery, chatbot or workflow tools to libraries and library consortia. The second is publishers whose licensed content is retrieved, summarized or surfaced through those AI-enabled systems, often without control over how the vendor logs that use.
For both, the ten-item list shows what procurement discussions will ask for next: a usage record the library’s counsel can compare with the contract, in addition to a summary of data practices. Citations has documented the same shift on the publisher-licensing side, where a disclosure statement and an audit trail answer different questions. Here it arrives through procurement rather than regulation, and from a different buyer, the library rather than the AI platform. The same gap appears when reference and STM catalogues are consumed by AI systems inside library and discovery infrastructure rather than read directly.
The resulting decision is narrow and concrete. Before the next AI vendor renewal or RFP cycle, an organization should know which of the ten items it can answer with a record and which only with a sentence. AI usage evidence is designed to close that gap by turning each of those questions into a record.
The gap also starts earlier, inside the organization, with the question of who may upload which content to an AI system in the first place. When the system being procured is a tutor rather than a search tool, the checklist narrows to six acceptance tests for an AI tutor built on licensed content. When it is a catalog connector that can place holds, the tests move to each circulation event an AI assistant creates.
This guidance is one association’s position rather than a universal library standard, and consortia and library systems will apply it differently depending on their governance and risk tolerance. The next signal to watch is whether the ten-item checklist starts appearing verbatim in library RFP language during the coming renewal cycle.
Continue the evidence chain
AI Disclosure Is Not an Audit Trail for STM Publishers
AI Usage Evidence for Publishers
Book an AI usage evidence assessment
Sources
American Library Association — "ALA Council adopts Guidance on the Use of Artificial Intelligence in Libraries," July 16, 2026
ALA Council adopts Guidance on the Use of Artificial Intelligence in Libraries
American Library Association — "Guidance on the Use of Artificial Intelligence in Libraries" (CD#44.2, full text)
Guidance on the Use of Artificial Intelligence in Libraries (CD#44.2, full text)