Programme design by Adrian Dunkley

AI safety and alignment for regional decisions.

Learn to test AI systems, assess their limits and make evidence-based decisions about their use in Latin America and the Caribbean. Work through original cases in hurricane assistance, public services, insurance, tourism and multilingual communication.

Register your interest

Expressions of interest are open. Dates, delivery arrangements and any fees will be confirmed before enrolment. Registering does not commit you to attend or pay. The proposed programmes do not currently carry an external accreditation or certificate commitment.

Technical track

Technical AI Safety and Alignment

Proposed: six weeks, approximately 36 hours. For developers, data scientists, technical auditors and researchers with basic Python and data skills.

Build a pilot evaluation dataset, compare safeguards, test agent permissions and defend a bounded safety assessment.

View technical syllabus

Governance track

AI Governance and Safety Assurance

Proposed: six weeks, approximately 24 hours. For directors, public officials, lawyers, risk teams and civil-society leaders. No coding required.

Produce an AI inventory, supplier evidence request, obligations matrix, appeal process and incident plan.

View governance syllabus

Learn with the conditions you work in

Exercises use fictional regional cases and synthetic records. Regional aggregates do not establish safety in every country or language. You will record the limits of your evidence and identify decisions that require further review.

The design includes downloadable readings, accessible diagrams and a supplied-output pathway for learners without model API access. Optional API or compute costs would be separate and disclosed before enrolment. Delivery languages and accessibility arrangements will be confirmed with each cohort.

Proposed curriculum

Course 1: Technical AI Safety and Alignment for Latin America and the Caribbean

Learn to evaluate an AI system before it makes decisions or takes actions that affect people. Over a proposed six-week programme, you will build a regional evaluation dataset, test a bounded AI workflow, compare safeguards and produce a safety assessment with clear limits. Examples cover public services, insurance, tourism, disaster assistance and multilingual communication across Latin America and the Caribbean.

Call to action: Register your interest.

Status copy: We are collecting expressions of interest. Dates, delivery arrangements and any fees will be confirmed before enrolment. Registering does not commit you to attend or pay.

Audience and prerequisites

Designed for software engineers, data scientists, AI product developers, technical auditors and researchers. Learners should be able to read basic Python, work with CSV or JSON records, explain training and test data, and interpret proportions and uncertainty. An optional preparatory notebook should introduce these skills. A degree or previous AI-safety certificate is not a prerequisite. Learners who want a noncoding introduction should choose Course 2.

Proposed workload: six weeks at approximately six hours per week, for 36 hours: 90 minutes of seminar, 90 minutes of reading and two hours of lab work each week, plus one hour for reflection and assessment. These are curriculum estimates, not externally accredited hours.

Outcomes

On completion, a learner who passes the assessment should be able to:

  1. Write a bounded safety question for a regional use case, including the affected population, system permissions, failure mechanisms and evidence required for deployment.
  2. Explain objective misspecification, reward hacking, distribution shift and limits of preference-based training without claiming that one training intervention solves alignment.
  3. Construct and document an evaluation dataset with development and held-out cases, relevant language or access conditions, reproducible scoring and explicit sampling limitations.
  4. Compare a baseline workflow with safeguards, measure harmful failures and task success, and report uncertainty and possible tradeoffs.
  5. Test an agent's action permissions, human approvals, logging and fallback behaviour in a contained environment.
  6. Defend a deployment, restricted-pilot or rejection recommendation using an evidence record that another reviewer can inspect.

Week 1 - Define the service and the harm

Guiding question: What could go wrong, for whom, and how would we know?

Topics: models versus services; task success versus public value; intent, misuse and accident; stakeholders; uncertainty; severity, reversibility and exposure; identifying who has authority to set objectives. Introduce the difference between technical alignment and institutional accountability. Compare the accident-oriented framing in Amodei et al. with the broader sociotechnical analysis by Raji and Dobbe. The comparison should identify what each framing helps the learner see and what it omits.

Original case: A fictional island disaster agency wants a chatbot to explain hurricane relief eligibility. The bot has access to an approved policy document but can only draft answers. It has no authority to award benefits. The project sponsor treats fewer calls as success; residents care about accurate access to aid. Students identify the gap between these objectives.

Exercise: Draw a system boundary and a harm pathway. Identify a resident without reliable internet, a resident using a Creole variety, a staff reviewer and the budget owner. Write three observable failure conditions, such as inventing an eligibility rule or telling an eligible user not to apply. Add the evidence needed to detect each condition.

Deliverable: a two-page use-case and threat specification with an explicit non-goal. The model is not authorised to decide eligibility.

Feedback check: Could another evaluator tell the difference between a wrong answer, a disliked answer and an unauthorised action?

Week 2 - Objectives, training and oversight limits

Guiding question: Why might a system satisfy a score while missing the purpose?

Topics: proxy objectives; preference feedback; training versus evaluation; supervised fine-tuning and reinforcement learning from human feedback at a conceptual level; scalable oversight; generalisation limits; specification gaming; disagreement among annotators. Introduce deception and situational-awareness hypotheses as research questions requiring carefully bounded evidence. Avoid describing model text alone as proof of an internal goal.

Original case: A fictional Caribbean insurer rewards a claim-summary assistant for concise summaries. The assistant often removes uncertainty about flood damage and omits missing evidence. The brevity score improves while reviewers lose information they need.

Exercise: Score 20 supplied summaries against two rubrics: brevity alone and evidence-preserving usefulness. Inspect disagreements. Design a revised rubric that records missing information, unsupported statements and the difference between a source observation and a model inference. This exercise compares rubrics; it does not claim to train or align a model.

Deliverable: a scoring guide with five annotated examples, two counterexamples and an account of disagreement that could remain after training.

Feedback check: What behaviour would game the revised rubric, and how would the student detect it?

Week 3 - Build a regional evaluation that can fail

Guiding question: Does the proposed test measure the failure that matters?

Topics: predefining outcomes; development and holdout separation; contamination; sampling units; subgroup coverage; language equivalence; independent grading; false positives and false negatives; denominators; confidence intervals; dependence among repeated prompts. Introduce Inspect as an optional open-source implementation tool, with a plain Python alternative and supplied-output pathway.

Original case: A fictional tourism service answers cancellation and refund questions for travellers and local businesses in several languages. Answers that sound fluent may silently change conditions. A model may answer English correctly but mis-handle an equivalent Spanish or Creole request.

Exercise: Build a small pilot dataset of at least 60 distinct scenarios, with 20 held out before development. Record each scenario's source, policy rule, acceptable answer, unacceptable claims, language and reason for inclusion. Translation variants must retain their source-case identifier so they are not counted as independent evidence. Include ambiguous cases that require clarification. Report this as a pilot rather than a representative regional benchmark.

Deliverable: dataset card, scoring code or scoring worksheet, frozen split manifest and pilot results. Learners with no API access use instructor-supplied records and label their analysis accordingly.

Feedback check: Can another student reproduce the denominator and explain which populations or language varieties the dataset does not support?

Week 4 - Interpretability, uncertainty and evidence quality

Guiding question: What can an explanation establish, and what can it fail to establish?

Topics: local feature explanations versus model internals; mechanistic interpretability as a research approach; causal tests versus attractive visualisations; faithful explanations; uncertainty and abstention; hallucinated citations; model-graded evaluations and evaluator bias. Discuss why a written chain of reasoning or a persuasive explanation does not automatically reveal the model's causal process.

Original case: A fictional lender uses an AI assistant to explain why an applicant's documents need review. A heat map highlights income, while the workflow may be responding to formatting, missing pages or neighbourhood proxies. Students must decide what the plot actually supports.

Exercise: Compare a source-grounded answer, an answer with an invented citation and an answer that correctly abstains. Identify which evidence validates the answer and which evidence only makes it appear plausible. In an optional coding extension, perturb irrelevant document formatting and relevant facts separately; measure changed decisions and document the limits of the test.

Deliverable: an evidence-quality memo plus a failure catalogue that distinguishes explanation quality from prediction quality and causal understanding.

Feedback check: Which claim would require an experiment the student has not conducted?

Week 5 - Contain actions and recover from failure

Guiding question: What limits the damage if the model is wrong or manipulated?

Topics: prompt injection; untrusted retrieved documents; data provenance; least-privilege tools; read versus write permissions; approval thresholds; sandboxing; action logs; timeouts; rollback; correlated failures; incident triage. Defensive exercises must use harmless synthetic content and isolated local resources.

Original case: A fictional outsourcing firm uses an agent to draft a customer refund. An attachment contains text instructing the agent to bypass approval. Students compare a draft-only workflow, a workflow with unrestricted refund tools and a workflow requiring a separate approved decision record.

Exercise: Use a mock payment function with fake balances. Compare the baseline with restricted permissions and an independent approval step. Test whether the agent can move from a retrieved document into a write action. Track unauthorised actions, legitimate tasks completed, false blocks, review burden and recovery time. Students must not connect the lab to a live payment account.

Deliverable: a control test report containing event logs, failure examples and a recovery procedure.

Feedback check: Does the proposed approval step actually prevent the action, or merely add a confirmation message the agent can bypass?

Week 6 - Make and defend a deployment decision

Guiding question: Is there enough evidence to authorise a specific use under specific conditions?

Topics: claims, evidence and residual uncertainty; risk acceptance; release gates; staged pilots; monitoring drift; incident notification; responsibility for retraining and policy changes; stopping criteria; honest communication of limitations.

Original capstone: Choose the disaster assistant, claim-summary assistant, tourism assistant, lender explanation workflow or another approved synthetic regional case. Compare a baseline and one defensible intervention. Freeze the evaluation plan before the held-out test. The goal is an evidence-backed decision, including a recommendation not to deploy if appropriate.

Deliverables: a reproducible notebook or documented record analysis; dataset and limitations card; control diagram; results with denominators and uncertainty; a five-page safety assessment; and an eight-minute defence. Include a one-page public explanation that does not overstate the experiment.

Technical course assessment

Proposed weighting: weekly work 40%, capstone 50%, oral defence and response to review 10%. A proposed completion threshold is 70/100 and submission of all safety-critical deliverables. These thresholds require provider approval before enrolment.

Capstone criterion Weight Full-credit evidence
Bounded question and regional context 15 Defined population, permissions, harms, authority and exclusions
Evaluation validity 25 Frozen split, defensible labels, relevant cases, leakage controls and honest scope
Reproducibility and quantitative reporting 20 Executable or inspectable workflow, denominators, uncertainty and traceable outputs
Controls and recovery 20 Tested safeguards, bypass checks, action logs and a workable response procedure
Decision quality and limitations 20 Recommendation follows the evidence, states residual risks and identifies a responsible owner

Rubric anchors for each criterion: 0 = absent; 1 = asserted with major gaps; 2 = partially evidenced; 3 = clear and adequately evidenced; 4 = independently inspectable and tested against a plausible counterexample. Multiply the criterion weight by the anchor divided by four.

A reproducibility fabrication, concealed use of real personal data, or an uncontained live-system test requires correction before completion, regardless of numerical score. A low measured failure rate on a small teaching dataset must not be described as certification of general safety.

Course 2: AI Governance and Safety Assurance for Latin American and Caribbean Leaders

Learn how to decide where AI belongs in your organisation and what evidence you should require before approving it. Over a proposed six-week programme, you will examine regional governance challenges, assess suppliers, design human oversight and practise responding to an AI incident. You will finish with a practical approval and monitoring pack for a government agency, business or civil-society organisation.

Call to action: Register your interest.

Status copy: We are collecting expressions of interest. Dates, delivery arrangements and any fees will be confirmed before enrolment. Registering does not commit you to attend or pay.

Audience and prerequisites

Designed for directors, public officials, lawyers, compliance and risk professionals, procurement teams, civil-society leaders and managers of AI-enabled services. No coding is required. Learners should be able to explain one service their organisation delivers, who depends on it and which team can change it.

Proposed workload: six weeks at approximately four hours per week, for 24 hours: a 90-minute seminar, 90 minutes of reading and case analysis, and one hour of applied work. These are design estimates and not an academic-credit claim.

Outcomes

A learner who passes should be able to distinguish technical findings from forecasts, legislation from proposals, and voluntary guidance from a legal obligation; build an inventory of AI uses and accountable owners; request appropriate supplier evidence; define meaningful human review and appeal; evaluate inclusion and workforce effects; and lead a bounded incident response and board approval discussion.

Week 1 - Set the decision and allocate responsibility

Guiding question: Who can authorise the system, and who carries the consequences?

Topics: AI safety, alignment, ethics and assurance; governance of imported models; decision rights; public-interest objectives; an AI inventory; materiality and reversibility; boards, management, vendors and frontline staff.

Original case: A fictional public agency purchases a chatbot bundled with office software. No one regards it as a separate procurement. Staff begin using it to draft decisions about assistance. The organisation must identify when a low-risk drafting tool became part of a consequential decision process.

Exercise: Build a ten-row AI inventory for a hypothetical agency or company. For each use identify purpose, affected people, data classes, supplier, tool permissions, owner and review date. Choose one use for the final project.

Deliverable: inventory and a responsibility map that includes an operational owner and an independent reviewer.

Week 2 - Read the rules without merging them

Guiding question: What applies in this jurisdiction, to this actor, on this date?

Topics: legal status; enactment versus commencement; territorial scope; data protection; automated decisions; consumer protection; employment and sectoral rules; contracts; standards; UNESCO ethics guidance; UN governance initiatives; regional coordination. Laws must be checked against the actual official text during delivery. This course provides governance education and does not promise legal advice.

Original case: A fictional Caribbean insurer buys an AI service hosted abroad and uses it to triage claims from several jurisdictions. A vendor says its EU compliance pack covers everything. Students identify which assertions require local legal review and which responsibilities remain with the insurer.

Exercise: Create a dated obligations matrix for one chosen jurisdiction and one cross-border relationship. Each row records the instrument, issuing authority, exact official source, status, scope, commencement evidence, relevant provision and owner. Unknowns stay marked unresolved.

Deliverable: obligations matrix and a one-page memo distinguishing binding obligations, voluntary guidance, proposed measures and contractual commitments.

Week 3 - Ask suppliers for evidence that changes a decision

Guiding question: Which evidence would make you approve, restrict or reject the service?

Topics: system and model documentation; evaluation populations; language coverage; security and privacy; subcontractors; version changes; incident history; audit access; service continuity; exit arrangements; data return and deletion; contractual allocation of responsibilities. Accuracy claims without denominators and relevant test conditions are insufficient.

Original case: A fictional tourism association is offered an AI concierge described as accurate in 40 languages. The brochure does not say which languages were assessed, how local regulations were tested or what happens when hotel policies change.

Exercise: Prepare a supplier evidence request and compare two fictional bids. One has polished marketing and little test evidence; the other admits narrow coverage and provides inspectable logs. Identify a third option: a limited pilot with explicit exclusions.

Deliverable: a supplier scorecard, proposed contract schedule and a pilot acceptance test. Students must explain why a weighted score cannot erase a critical unmet requirement.

Week 4 - Protect participation, workers and the right to challenge

Guiding question: Who loses access or bargaining power when the service changes?

Topics: access and affordability; language and disability; digital literacy; workers' consultation; task exposure versus predicted job losses; public communication; contestability; appeal; data minimisation; alternatives for people who cannot use the AI channel. Use the World Bank/ILO study as a model of exposure analysis, not as an observed count of jobs already lost.

Original case: A fictional municipality makes an AI portal the fastest way to apply for a licence. Applicants with old phones or low literacy rely on a slower staffed channel. Reported service speed improves while access becomes less equal.

Exercise: Write a service inclusion plan. Specify equivalent application routes, accessible instructions, human support, monitoring of delays and a way to challenge an incorrect decision. Use role-play with a learner representing a worker, resident, manager and disability advocate. Do not present the role-play as actual community consultation.

Deliverable: a participation and appeal plan with measurable service standards and evidence still needed from affected groups.

Week 5 - Respond when an AI service causes harm

Guiding question: What happens in the first hour, first day and first week?

Topics: incident triggers; evidence preservation; pausing an automated step; service continuity; privacy and security escalation; public communication; regulator and customer notification subject to verified legal duties; root-cause review; restitution and follow-up. Discuss how cross-border suppliers complicate response.

Original case: A fictional benefits chatbot invents a deadline after a hurricane. Call-centre staff repeat it. Some residents may have missed applications, while the supplier has already updated its model. The response must preserve evidence and restore access without waiting for a complete technical explanation.

Exercise: Conduct a 45-minute tabletop using synthetic logs. At three injects, decide whether to pause the system, what residents should be told, who will investigate and how people can reapply. Separate confirmed facts, working hypotheses and unresolved impact estimates.

Deliverable: incident playbook, event timeline, responsibility list and a draft public notice with no invented assurances.

Week 6 - Defend a regional assurance and implementation plan

Guiding question: Can the organisation operate and review the system with its actual staff and budget?

Topics: oversight capacity; shared regional testing; skills development; safe experimentation; monitoring budgets; independence; procurement leverage; exit and switching; phased implementation; communicating evidence to boards and the public.

Original capstone: Develop a 90-day approval and monitoring plan for the selected service. The fictional organisation has a small team and cannot create a large new regulator or technical laboratory. Students must select achievable controls, identify external expertise and specify what remains too risky to authorise.

Deliverables: an AI inventory entry; dated obligations matrix; impact and inclusion assessment; supplier evidence request; approval conditions; human review and appeal process; incident plan; monitoring dashboard; and a five-page board paper. Include an implementation budget as a stated assumption rather than an empirical cost estimate.

Governance course assessment

Proposed weighting: weekly portfolio 50%, capstone 40%, oral defence 10%. Proposed completion threshold: 70/100 and all critical evidence and responsibility records submitted. Providers must confirm these terms before enrolment.

Criterion Weight Full-credit evidence
Accuracy of legal and institutional status 20 Official dated sources, territorial scope and explicit unresolved questions
Evidence and procurement judgement 20 Requirements tied to consequences, appropriate tests and critical exclusions
Responsibility and operational feasibility 20 Named roles, achievable staffing, budget assumptions and review cadence
Inclusion, human review and appeal 20 Practical access alternatives, review authority and a route to correction
Incident readiness and decision communication 20 Clear triggers, evidence preservation, continuity and an honest board recommendation

Use the same 0–4 evidence anchors as Course 1. An otherwise high score cannot pass a plan that authorises consequential automated actions without any identified accountable owner or challenge mechanism. That is a pedagogical completion rule, not an assertion about every jurisdiction's legal requirements.

Register your interest

Tell us which track fits your work. We will contact you about programme arrangements. This is an enquiry, not an enrolment or payment form.

StarApple AI uses this enquiry to plan the programme and respond to you. Submissions are held in the site's Netlify Forms account and accessed by the programme team. Provide practical access requirements only; leave out medical details, identity documents and confidential work records. Contact insights@starapple.ai about this enquiry or to request correction or deletion. This form does not subscribe you to a newsletter.

Research readings

Readings include Concrete Problems in AI Safety, the NIST AI Risk Management Framework, UNESCO's Caribbean AI Policy Roadmap and the World Bank and ILO analysis of jobs and generative AI. Use of these readings does not imply institutional endorsement.