For M&E Professionals

AI tools are increasingly used for survey data collection, real-time analysis, beneficiary verification, and outcome tracking. But are beneficiaries truly consenting to how their data is used — or just clicking “agree” to receive services?

For International Development

From refugee registration to agricultural extension services, AI-powered digital tools are transforming service delivery. Without meaningful consent, we risk creating new forms of exclusion, surveillance, and harm — especially for those with the fewest alternatives.

1. Freely Given

Consent is not valid if users feel coerced or if refusing means losing access to essential services. In development contexts, power imbalances between service providers and beneficiaries are common and must be addressed.

2. Specific

Users must consent to each specific use of their data — not blanket consent for undefined future purposes. For M&E, this means separating consent for data collection, analysis, storage, and sharing with third parties (including government ministries).

3. Informed

Users must understand what they are agreeing to — including what data is collected, why, how it will be used, who can access it, how long it is kept, what risks exist (including re-identification risk), and that AI may infer sensitive attributes. This is the most challenging aspect across literacy, language, and cultural differences.

4. Unambiguous

Consent must be shown through a clear affirmative action — not pre-ticked boxes, silence, or inaction. In low-literacy contexts, verbal consent or gesture-based consent may be needed.

5. Revocable

Users must be able to withdraw consent as easily as they gave it, without penalty. In development programs, this requires clear, accessible withdrawal pathways (SMS, hotline, in-person).

Consent approach: Account creation required, terms accepted without reading for most users. Data collected includes conversations, IP addresses, device info, and usage patterns.

💡 What this teaches us: Even leading AI tools fail to meet basic informed consent standards for low-literacy, low-connectivity, or multilingual contexts.

Consent approach: Uses visual consent cards, verbal scripts in local languages, and step-by-step explanations before fingerprint capture. Users can refuse without losing services.

💡 What this teaches us: Visual aids, local language scripts, and facilitator-led consent can work effectively in development settings.

Consent approach: Enumerators read consent scripts aloud. Respondents give verbal or written consent before starting surveys. Consent is recorded and auditable.

💡 What this teaches us: Audio/verbal consent works well, but ongoing consent and withdrawal options need improvement.

Consent approach: Multi-language consent forms. Biometric data collection requires separate explicit consent. Data use is explained by registration staff.

💡 What this teaches us: Consent must be meaningful even when there is a power imbalance between aid providers and beneficiaries.

Accessible

Available in local languages, with audio and visual options for low-literacy users. Works offline or with low bandwidth.

Specific

Separate consent for different data uses (collection, storage, sharing with government, third-party access). Users can choose.

Transparent

Plain language explanation of AI’s role, potential errors, data retention, real risks (including re-identification), and inference risks — not just benefits.

Choice-Full

Meaningful alternatives for those who decline (human support, anonymous use, partial participation, time-limited consent).

📱 SMS Withdrawal

User texts a keyword (e.g., “STOP” or “WITHDRAW”) to a shortcode. Automated response confirms deletion request and provides a reference number. Example: “Text STOP to 12345 to stop using this tool and request data deletion.”

📞 Hotline Withdrawal

Dedicated phone number (toll-free where possible) staffed during program hours. Caller provides their ID number and verbal confirmation. Example: “Call 0800-123-456 to speak with someone who can delete your information.”

🏠 In-Person Withdrawal

Users can visit a local program office or community health worker to request withdrawal. Especially important for low-literacy or offline contexts. Example: “Visit [local office name] and ask for [staff name] to delete your data.”

📱 Prototype A: Minimalist Approach

Common in commercial apps

Screen displays:

“By using this AI tool, you agree to our Terms of Service and Privacy Policy.”

I Agree
Cancel

🔍 Critique for M&E & Development Contexts:

• ❌ Not accessible: Links to legal documents assume literacy and internet access
• ❌ Not specific: No explanation of what data is collected or how AI is used
• ❌ Power imbalance ignored: Users may feel they cannot refuse
• ❌ No choice granularity: All or nothing consent
• ⚠️ Result: Likely to fail meaningful consent standards in development settings

📱 Prototype B: Layered Visual Approach

Better for global development

♿ Accessibility note for digital implementation: Ensure buttons have proper aria-label attributes for screen readers. Use aria-live="polite" regions for dynamic content updates. For mobile/web: <button aria-label="Yes, I agree to data collection">. Test with screen reader software (NVDA, VoiceOver) before deployment.

🔍 Critique for M&E & Development Contexts:

• ✅ Accessible: Layered approach reduces cognitive load
• ✅ Specific: Clear data categories and purposes explained
• ✅ Choice-full: Multiple meaningful options including anonymous use and time-limited consent
• ✅ AI inference warning: Now includes risk of inferred data
• ⚠️ Needs visual support: Add icons for low-literacy users (🔒 = data retention, 🤖 = AI role, 🗑️ = deletion, 🙋 = human support)
• ⚠️ Audio needed: Should include spoken version for non-readers
• M&E relevance: This approach is promising for digital data collection in diverse settings

👥 Prototype C: Facilitator-Led Approach

For low-literacy or offline contexts

Facilitator Script (read aloud to user):

📋 Minimum Facilitator Training Requirements:

• Never read the script monotonously — use natural pauses and eye contact
• Pause after explaining each risk to allow questions
• Observe hesitation (looking away, silence, body language) and ask “Is there anything you’d like me to explain again?”
• Practice responding to common questions without pressure or rushing
• Role-play with community members before field deployment
• Document refusals and reasons to improve the process

🔍 Critique for M&E & Development Contexts:

• ✅ Most accessible: No literacy required
• ✅ Relationship-based: Facilitator can answer questions and observe understanding
• ✅ Power dynamic acknowledged: Explicitly states no penalty for refusal
• ⚠️ Scalability: Requires trained facilitators, more time per user
• ⚠️ Consistency: Quality varies by facilitator training — see training note above
• 📊 M&E relevance: Gold standard for sensitive data collection or low-literacy populations

Steps 1-5: Foundation

1. Define the AI tool and its users
2. Map what data the AI tool collects
3. Explain why each type of data is collected
4. Identify data risks (especially for vulnerable groups) — Include risk of re-identification (even from anonymized data) when using AI models that retain training data patterns.
5. Decide what choices users really have (including anonymous use and time-limited consent)

Steps 6-10: Design & Testing

6. Write the first consent message in plain language
7. Use layered consent (summary → details → full policy)
8. Explain the role of AI (it may make mistakes and may infer information)
9. Translate and adapt for local languages and literacy levels
10. Test with real users from your target population

Steps 11-15: Implementation & Monitoring

11. Add a consent checkpoint (comprehension check)
12. Make withdrawal easy and clearly explained (SMS, hotline, in-person)
13. Review consent for equity and safety (gender, displacement, children, emergency contexts)
14. Create the final consent flow with ongoing control options
15. Monitor and improve consent over time as the tool evolves

□ User knows they are interacting with AI

□ Consent message is in user’s local language

□ I have explained what data is collected and why

□ I have explained risks (including AI inference and re-identification)

□ User has seen the withdrawal methods (SMS/hotline/in-person)

□ I have explicitly said “No penalty if you refuse”

□ User has been offered private consent option (if relevant)

□ For children: Parent/guardian consent AND child assent obtained separately

□ For government data sharing: Separate consent obtained

□ User has passed a simple comprehension check

□ User’s consent decision has been recorded

□ I have answered all questions honestly and without pressure

□ Users know they are interacting with AI

□ Consent message is in plain, local language

□ Data collection is clearly explained

□ AI inference risk is disclosed

□ Re-identification risk is disclosed

□ Data risks (especially for vulnerable groups) are visible

□ Users have meaningful choices beyond yes/no

□ Low-literacy users are supported (audio/visual)

□ Users can ask questions before consenting

□ Withdrawal pathways are clear (SMS, hotline, in-person)

□ No penalty for refusal is explicitly stated

□ Process tested with real users from target population

□ Gender and equity safeguards are included

□ Children and adolescents have separate consent/assent processes where required

□ Humanitarian emergencies have a documented protocol for modified consent

□ Government data sharing has separate explicit consent

□ Consent process is reviewed and updated regularly