Independent review of Mutual AI Resilience
Question the proposal, check the evidence, compare simpler options, and decide what you would approve for further testing, revise, defer, or reject.
Independent Judgment and Review
Make your own judgment, ask which records support it, and hear the strongest objection before deciding again. A reviewer can disagree, keep a simpler non-AI process, or recommend stopping the project. Independent review runs through all eight policy commitments.
Self-review and peer feedback help in rehearsal. An operator’s review checks whether the procedure is usable and authorized. An independent audit needs a reviewer with access to the evidence and room to disagree. For a real review, record its scope, reviewer, methods, dates, and conflicts of interest.
MAR is at the proposal stage. Its benefits and risks still need independent evaluation in specific settings.
What MAR proposes
Mutual AI Resilience is a voluntary proposal for countries to share only the information they need to verify a warning, arrange civilian help, or keep an essential service running. Operators check and authorize decisions. An AI tool could flag a possible discrepancy for review. Each participating country retains its own authority and can refuse, pause, or leave the arrangement and use a local fallback.
Five stages of independent review
Observe before interpreting
Name the source, issuer, time, place, scope, and how long the information is valid. Classify what matters as a confirmed observation, an unconfirmed report, an interpretation, a testable hypothesis, or an unknown. Ask who can verify or correct each record.
Decide independently
Take a minute before hearing the group or opening a sample response. What would you do now? What evidence supports that choice, what authority or information is missing, and what would change your mind? You can think, write privately, or share.
Challenge and reverse
Give the strongest credible reason to keep the usual process and the strongest reason to try a change. Ask what happens if the partner declines or the AI flag is wrong. Take turns speaking as a service user, an operator, a reviewer checking authority and privacy, and a person proposing an AI trial. On your own, switch perspectives before evaluating your first call.
Test competing options
Compare usual calls or records, a checked human agreement, and the same agreement with a mock AI flag. Name the responsible operator, the civilian outcome, and an unacceptable harm before starting. A refusal or cancellation is a review decision. It takes no extra simulation round.
Review or dissent
Record the reason for your decision, a viable alternative, and an observation that would overturn your conclusion. Note who has authority, which safeguards to keep, and what changed since your first answer. Keep a separate space for dissent. Label the evidence as a proposal, a rehearsal observation, an independently verified result, or unresolved.
Running the rounds in the same order helps people rehearse. The last round may benefit because learners already know the case. To estimate an effect, vary the order or assign comparable teams to matched cases. Hold facts, staffing, and resources constant. Choose measures and stop rules in advance, and record departures such as information leaking between teams.
Choose a policy principle to challenge
Each card pairs an existing commitment with a test and a concrete reason to pause. Choose one commitment and use the related lesson to examine it.
Preserve sovereign control
Each participant needs the practical ability to refuse help or leave while keeping critical services running.
- Implementation test
- Withdraw the visiting crew and disconnect the shared status card. Ask the receiving dispatcher to show which local procedure still works and estimate what withdrawal costs each party.
- Your judgment
- Would you keep the agreement if a key decision still required the partner’s permission after disconnecting?
- Reason to stop or reject
- Stop if an essential function cannot be carried out using the local fallback.
Read the policy commitment Storm crew and pumping station exercise
Use practical diplomacy to reduce misunderstanding
A dispatch report and a receiving confirmation can legitimately tell different stories.
- Implementation test
- Give reviewers a dispatch record and a conflicting translated notice. Time how long the receiving office takes to verify the status. Check that the unresolved account remains visible until the office checks its register.
- Your judgment
- Which office can confirm arrival? Would you keep the phone clarification if a shared notice travels faster but loses the distinction?
- Reason to stop or reject
- Stop if a mistranslated or unverified status drives an official pickup notice.
Read the policy commitment International mail pickup exercise
Protect civil liberties and private information
Schools can share spare seat counts while keeping family information private.
- Implementation test
- Test the fictional exchange with aggregate counts only. Give authorized staff a wrong count to correct. Verify who can edit or delete it and whether unnecessary information survives in an export.
- Your judgment
- Could the partners do the work with fewer fields? Who can request a correction without exposing a family’s circumstances?
- Reason to stop or reject
- Stop if a partner receives or retains unnecessary personal details without safeguards.
Read the policy commitment Displaced students returning to school exercise
Make civilian service continuity the objective
A status marked “available” tells people little about whether they can complete an accessible trip.
- Implementation test
- Use the same fictional riders across comparable scenarios. Check the full trip, including the path, boarding point, assistance, and whether a rider can actually board.
- Your judgment
- Which would you keep if it enabled more people to finish the trip, the shared checklist or the same checklist with AI?
- Reason to stop or reject
- Stop if the reporting shows a completed trip while riders still cannot reach or board the service.
Read the policy commitment Finding the replacement shuttle exercise
Keep people accountable for consequential decisions
A referral commits real capacity. People need to know who can accept it, correct it, or explain a refusal.
- Implementation test
- Trace one mock capacity flag through the receiving shelter’s check and authorized decision. Ask the operator to override it when confirmed capacity changes.
- Your judgment
- Could you tell the next shift exactly who confirmed the bed, who accepted the referral, and why?
- Reason to stop or reject
- Stop if staff cannot reconstruct the decision or have no practical way to override a flag.
Read the policy commitment Animal shelter capacity during evacuation exercise
Require evidence before expansion
A faster alert matters when residents can understand and act on it.
- Implementation test
- Prespecify alert comprehension and a maximum tolerable rate of misdirection. Compare the three existing approaches with matched cases or counterbalanced order and the same resources.
- Your judgment
- Would you keep the human-only improvement if the added AI produces no meaningful gain or results vary with the order?
- Reason to stop or reject
- Pause expansion if the effect is unreliable, the AI adds no meaningful benefit, or the harm threshold is breached.
Read the policy commitment Who receives and understands an alert exercise
Make costs and dependencies visible
Staff hours, review time, software, and service lost elsewhere are real costs.
- Implementation test
- Record the time both offices spend checking and coordinating. Disable the proposed supplier and price how each office could leave, keep essential records, and resume local work.
- Your judgment
- Would a simpler arrangement do as much for residents at lower cost or with fewer dependencies?
- Reason to stop or reject
- Stop if essential work depends on one hidden provider or if the exit cost exceeds the agreed limit.
Read the policy commitment Sharing cleanup crews after a storm exercise
Earn public trust through reviewable work
Readers need to see where a notice came from, what it applies to, and how to raise a correction.
- Implementation test
- Give a reviewer a sanitized log with source times, applicable areas, corrections, and dissent. Ask them to reconstruct the decision and submit a correction to the issuing office.
- Your judgment
- Would you publish the result if a skeptical reviewer could see the conclusion but could not trace or contest it?
- Reason to stop or reject
- Stop if a source was changed silently, a correction is ignored, or the record hides material dissent.
Read the policy commitment A water notice with disputed coverage exercise
Worked example for a storm crew and pumping station
Documented fact. On September 9, 2017, Florida Power & Light asked Ontario’s Hydro One for help with Hurricane Irma. Hydro One reported that it mobilized 175 employees within 24 hours. Their trip and work in Florida took 14 days altogether. This shows utilities could arrange cross-border help in 2017. Hydro One, “Hydro One awarded for outstanding restoration efforts after Hurricane Irma,” January 10, 2018.
Fictional exercise. A pumping station lost power. The water operator confirmed that one pump runs on backup and has yet to check the second. An older shared card lists the visiting crew as “available.” The sending dispatcher received a newer truck-delay report and expects an update at 10 a.m. A local crew may free up later.
- Verified observation
- The water operator saw one pump running. The shared card really contains the word “available.” The crew’s location still needs separate confirmation.
- Reported but unconfirmed
- The truck-delay report has reached the sending dispatcher. The visiting crew’s arrival time and the local crew’s eventual availability still need confirmation.
- Still unknown
- Whether the second pump is working, when either crew can reach the station, and whether residents will receive uninterrupted service.
Testable claim. A shared card with a checked validity time could shorten the interruption to water service. Set the primary measure to the duration of lost pumping service. Allow zero dispatches based solely on a stale card or mock flag.
Reveal the simulated AI flag
Mock AI output. “Crew available. Assign immediately.” The output treats an old card as a current confirmation. Use the same human authorization and independent checks as in the other rounds.
Compare three ways of handling the same evidence
Strongest objection. The shared card might slow a reliable phone process, give a stale claim more weight, or expose more operational detail without helping residents. A checked card could still reduce repeat calls. What result would distinguish those explanations?
Read one possible reviewed decision
Keep the local fallback ready. Hold the assignment until the sending dispatcher confirms the offer and the receiving dispatcher accepts it. Ask the operator what they can verify about pumping before updating residents. An ordinary call or a shared card without AI may be enough.
Measures. Record false assignments, time to confirmation, service interruption, staff minutes, wrong notices, unnecessary disclosures, and what still works during a disconnection. The exercise supplies no observed results or restoration times.
Stop trigger. If someone acts on an unverified suggestion or stale card without authorization, pause the trial, correct the record, and use the local process. Record the full error even if the assignment was made quickly.
Bounded conclusion. Under these fictional conditions, a timestamped record can be checked before an offer is accepted. Whether the shared process or AI improves service remains a question to measure.
Proposed MAR application. Across a border, the sending country’s authorized operator still controls its own crews. The receiver decides whether to accept. Each can decline or withdraw and keep a local fallback. An independent reviewer can ask whether the added system helped more than better human records.
Safe Nations · Organizer toolkit
Independent review worksheet
Write in your own words. The worksheet is optional and stays in this tab. Download a text copy to keep it, or print it for paper notes. Leave out private or unnecessary identifying details.
Jesse Shelley · CC BY-NC-SA 4.0 · safenations.org
What a real implementation would need
The worksheet records a learning decision. Real testing and any operational action require authorization from the responsible institutions.
What remains unproven
A drill can show whether participants checked an offer, preserved a fallback, or spotted a misleading flag under those exercise conditions. Real service gains, long-term trust between countries, security benefits, and effects on the risk of war need separate evidence.