A QTI-Certified Authoring Platform, Now Powered by High-Quality AI-Generated Items

Assessible’s Content Editor is a QTI-3 certified exam authoring platform that is already being used to author high-stakes and practice exams for our existing customers. It’s built to handle scale with item banks that hold an unlimited amount of items, an intuitive UI paired with WYSIWYG text editors, and a rich media and shared stimulus bank that make it easy for authors to build items.

We support features such as previewing and testing items in real time, item version control, and membership-based access to item banks, so not everyone sees everything. There are 15+ interaction types available today, with more in active development. The platform also lets authors create assessments either automatically from templates or manually, with full support for the attributes defined by QTI, giving authors a wide range of options for building genuinely unique assessments.

We also believe no item exists in a vacuum. Every item needs properties that support psychometrics and exam construction, so we let authors define their own metadata and customize it as much as their assessment program needs.

With an already feature-rich platform in our hands, the next direction was integrating AI. The challenge is that we support high-stakes exams, so AI-generated items need to be high-quality enough to perform just as well as battle-tested items backed by rich psychometric data. That’s why we partnered with IBE, who are leading the way in bringing AI support to educators.

Getting It Right

We have a long working partnership with the Institut für Bildungsevaluation (IBE). Recently, we worked together on Automatic Essay Scoring, a tool that scores a student’s essay, typed or handwritten, against a range of criteria and delivers detailed feedback and suggestions. That same expertise carries into this new effort: IBE has worked extensively with machine learning experts at EPFL Lausanne and psychometricians at the University of Zurich to develop their AIG capability. This time, the goal was more ambitious: scaling high-quality item production for mathematics and reading comprehension, natively across English, German, and French

Under the hood, AIG is a generator that takes cognitive evaluation models, the structured blueprints that define what an item is meant to measure, and turns them into fully functional test items. It’s not without risks. LLMs can hallucinate facts or produce items with inconsistent difficulty levels. In high-stakes testing, those aren’t minor issues, they’re the kind of thing that can undermine an entire assessment. That’s why fact-checking is already built into the service, and on top of that, we add human review. This has been our guiding principle from the beginning: human review, transparency, and clinical validation come first. They’re the safeguards that make AIG usable in a testing context.

The deeper mechanics of the generator itself, the modeling approach, the architecture, the guardrails, deserve their own write-up. What I want to share here is what we built on our end: what it actually feels like for an author to work with AI-generated items inside The Content Editor.

Prompting and Generating Items Without Leaving the Editor

Authors don’t have to leave Content Editor to generate items in the first place. Choosing a cognitive model, setting parameters, and triggering generation all happen right there in the same interface they already author in. The generated items are only as useful as what an author can do with them, so that’s where we focused next. We didn’t want authors exporting the output into some separate tool just to look at it. We wanted it to land right inside the editor, alongside everything else they author.

Refine and Edit, Not Just Approve or Reject

An item coming out of a model is rarely an assessment ready item, and we didn’t want the workflow to treat it as binary: accept or discard. So during the generation process the authors can:

• Edit generated items in real time. Authors can review generated text and items, and even edit them mid-process: tightening wording, reworking a distractor, adjusting a stimulus’s reading level.
• Send edited items back to the model. This feedback lets the model make those changes mid-process instead of starting over. Human feedback becomes part of the process itself, not just a final sign-off.
• Package the result. Once an item, or a full set, is ready, Content Editor produces a standards-compliant QTI package. From there, we give authors two paths: add it straight into an item bank inside the platform, or download the package for an external delivery engine.

One Workflow, Not Two

The first path, adding generated items straight into an item bank, was important to us for a specific reason: we didn’t want AI-generated content to live in a separate, parallel pipeline that authors had to manage on top of everything else. Once an item is in the bank, it’s just an item. Authors keep editing it, tag and classify it, route it through review and approval, and pull it into assessments exactly as they would with anything written from scratch. Generated and manually authored content share the same home, the same tools, and the same workflow.

What We’re Building Next: Insights and Analytics

Generating and banking items well is the first step. The question we keep hearing, and asking ourselves, is: how are these items actually performing? We’re already working on bringing performance insights and analytics into the same environment, so that over time, authors won’t just see the items they’ve generated, edited, and published; they’ll see how those items behave in the field. We want to close the loop between generation and real-world performance, so authoring decisions, human or model-assisted, are increasingly backed by evidence, not just intuition.

More on the technical side of AIG, including the modelling approach, the architecture, and the guardrails, is coming soon.