Safety, trust, and clear limits

Responsible AI safeguards are part of the product, not an appendix

Safety and trust

The public site makes the project’s limits visible so AMSA-AI is understood as a support layer for civic learning and preparation.

AMSA-AI is designed to support users, not replace institutions, lawyers, elected bodies, or public officials.

Privacy protection

No sensitive personal data collection unless necessary.

Human oversight

Escalation rules for situations requiring expert or institutional verification.

Source verification

Grounded responses and user prompts to check weak or uncertain outputs.

Misinformation awareness

Training content helps users identify unreliable AI-generated answers.

Clear limits

Transparent language about what the tool can and cannot do.

Inclusive language

Arabic, Darija, and low-resource language considerations are built into the model.

Bias awareness

Users learn to question outputs and avoid automated over-reliance.

Accessible interaction

User-friendly flows for non-technical local governance actors.

Enhancing administrative capacity through ethical, human-guided artificial intelligence

AMSA is a co-pilot, not a replacement. Every answer is a starting point for a human decision - never the decision itself. Elected representatives, civil society leaders, and citizens stay in control of what gets sent, signed, or said in their name.

AI prepares

Drafts questions, structures memos, simplifies procedures.

Human verifies

Checks citations, reviews context, validates against local realities.

Human decides

Owns the final intervention, document, or institutional action.

Where the lines are - and how we hold them

Ethics & data privacy pledge

AI in civic life only works if the safeguards are explicit. These four pillars are what AMSA-AI commits to, in writing, before the technical implementation.

Data sovereignty

Corpus, conversation history, and embeddings stay within our managed cloud infrastructure under strict access controls. The only outbound traffic is the language model provider call required to generate the answer text, and an optional notification webhook for user-reported issues.

Bias mitigation

Special attention to gender equity (the corpus deliberately surfaces IEECAG materials, gender-sensitive planning, and Article 6 of the Constitution), and to marginalised groups: rural communities, low-literacy users, Amazigh-speakers, persons with disabilities. Our 100-case Darija evaluation suite tracks where the model is still falling short.

Privacy compliance

Aligned with Moroccan Law 09-08 on personal data protection. No real names, ID numbers, phone numbers, or emails are required to use AMSA. Optional contact data submitted via the proposal form is used only to follow up with the sender.

No retraining on user data

Stored conversations are used only for evaluation, debugging, and quality review by the project team. They are never sent back to the model provider as fine-tuning data.