Manifesto for AI that makes room for learning

Our technical, pedagogical and ethical design choices, and why they differ from what many vendors do.

In a few words

Today, adding AI to a learning platform has become simple. An API subscription, a few lines of code, a chatbot in the interface: it can be done in a day.

The result often looks like this: an assistant that answers everything, whose data flows nobody precisely knows, and that may be incompatible with the next LMS update. Useful in appearance. Not always useful for learning.

We made different choices. They lengthened development time. They sometimes make explanations harder. We think they were worth it.

1. Build in Moodle, not alongside it

Since version 4.5, Moodle has had an official architecture for AI integration: the Provider/Placement subsystem. Its principle is to separate the user interface (the placement — where and how AI appears) from the AI model itself (the provider), with a central manager that orchestrates both and applies access policies.

This architecture is not a technical detail. It lets Moodle manage in one place user acceptance of AI use policies, usage reports, course- and activity-level access controls, and compliance with the platform's security standards.

Most Moodle AI plugins bypass this architecture. They bring their own API-key management, configuration settings and data logic, disconnected from Moodle. It is faster to develop. It also creates maintenance issues with each major update, security blind spots and fragmented governance for administrators.

PimenkoAI Suite is built with two plugins that follow Moodle's official architecture:

  • A provider: the connector to the AI model, developed to Moodle Provider standards — including Mistral, the model we selected.
  • A placement: the user interface, developed to Moodle Placement standards.

In practice, your Moodle administrator finds familiar native mechanisms: AI policy, usage reports and access control. PimenkoAI extends Moodle — it does not bypass it.

2. One tool, several modes — because the problems are connected

Most AI tools for learning address one problem at a time: a chatbot for learners, a content generator for educators, a dashboard for administrators. Separate tools, separate configurations, separate data policies.

We chose the opposite: one widget, several modes that share the same context, rules and data policy.

This choice is significant. A learner opening Tutor mode benefits from the same course context as the educator who used Creation mode to prepare activities. An administrator consulting data works in the same permissions system as the rest of the suite. The tool understands the framework in which it operates — because it works from within Moodle, not alongside it.

We add modes as new needs are confirmed. The logic stays the same: each mode shares the same infrastructure, data principles and contextual consistency.

3. Support that makes learners search, rather than searching for them

Learning-science research converges on one point: what anchors learning in long-term memory is the learner's own cognitive effort, not the quality of the explanation they receive.

The evidence is documented. A 2025 study in the Proceedings of the National Academy of Sciences followed secondary-school students using an AI tutor without pedagogical constraints. On an exam without AI, they scored 17% lower than learners who had never had access to it, despite higher scores in practice exercises where AI was available.

Another study, conducted in 2026 with more than 1,200 participants in a controlled experimental protocol, found that 10 to 15 minutes of unrestricted AI-tool use was sufficient to reduce persistence and independent performance.

The same work documents the reverse: when an AI tutor is designed with progressive support levels — leading learners to formulate their own answer before receiving more direct help — progress reached +127% in the same experimental protocol (Bastani et al., 2025, PNAS, 122(26)). An independent Harvard study published in Scientific Reports (Kestin et al., 2025, vol. 15, art. 97652) found that an AI tutor designed according to pedagogical best practice outperformed in-person active-learning courses, with effect sizes of 0.73 to 1.3 standard deviations.

These results align with the work of researcher Manu Kapur (ETH Zurich) on productive failure (2024): learners who first struggle independently before receiving an explanation retain more, and for longer, than those given the solution directly.

Our Tutor mode is built on these principles. The assistant asks questions before giving answers. It escalates to more direct help only when a learner remains stuck. It withdraws automatically during assessment. Teaching teams can configure its behaviour for each activity.

It is not a chatbot. It is a tool designed so learners do the cognitive work, not one that does it for them.

4. Humans remain in the loop — always

Whatever the feature, the tool proposes. The decision remains human.

In activity creation mode, AI generates a complete proposal; the educator reviews it, edits it when needed, then decides whether to add it to the course. Nothing is added without explicit confirmation.

In Tutor mode, the educator configures the level of support available for each activity. AI operates within that framework; it does not redefine it.

In Administration mode, every action requires confirmation before it runs. The tool shows what it will do, then the administrator approves it.

This principle also applies to content. When the tool does not have a validated source to answer a question, it says so explicitly. It does not fill gaps with speculative generation.

5. Your data — what is configurable, and what is not

We prefer transparency to vague wording.

What you control: your learners' Tutor-mode conversations are stored in your Moodle. Your administrator can configure the retention period — 90 days by default. Export and deletion are available on request, in accordance with the Moodle Privacy API. User identity is pseudonymised before any transfer to an external model: their Moodle identifier never leaves the LMS.

What is sent to our infrastructure: to operate, PimenkoAI sends our servers the elements strictly required to process each request: the user's message, active context (course, activity, mode) and a pseudonymised identifier. This data is sent to our infrastructure to ensure the quality and operation of the suite.

What belongs to our infrastructure: data used for contextual search in course content (RAG) is indexed and maintained on our infrastructure. The retention policy for these indexes is tied to our update processes. Your administrator cannot configure it, and we document its terms in our privacy policy.

Our model and infrastructure choices: Model: Mistral — developed in France, compliant with European law. Infrastructure: Scaleway — French hosting provider, SecNumCloud certified.

The AI Act entered into force in August 2026. We anticipated its requirements from the first day of development, not because it was mandatory when we launched, but because we believe it is the right way to build a tool for institutions responsible for their learners.

What we do not promise

We do not promise to generate your course for you, manage your platform for you, or guarantee spectacular gains for your learners. A course is the sum of pedagogical decisions that only teaching and technical teams who know their audience can make. AI does not know your constraints, organisational culture or actual audience.

Nor do we promise that our tool will replace a good educator or a well-administered platform.

We promise a carefully designed tool, built to Moodle standards, grounded in learning-science research and transparent about its data.

Research references

  • Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122.
    Online: https://www.pnas.org/doi/10.1073/pnas.2422633122

  • Kestin, G., Miller, S. T., McCarty, L. S., Callaghan, K., & Deslauriers, L. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, article 97652.
    Online: https://doi.org/10.1038/s41598-025-97652-6

  • Liu, X., Christian, J., et al. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. ArXiv preprint.
    Online: https://arxiv.org/abs/2604.04721

  • Kapur, M. (2024). Productive Failure: Unlocking Deeper Learning Through the Science of Failing. Jossey-Bass / Wiley.

PimenkoAI Suite is developed by Pimenko, a Certified Moodle Partner specialising in online learning solutions and open technologies.