AIGP cheat sheet

AI ⊃ ML ⊃ deep learning: AI is the broadest field, ML learns patterns from data, deep learning uses multi-layer neural networks, and generative AI produces new content from learned distributions.

Supervised learning maps labeled inputs to outputs (classification predicts discrete labels, regression predicts continuous values); unsupervised learning finds structure in unlabeled data (clustering, dimensionality reduction); reinforcement learning optimizes a policy via reward signals from an environment.

Training fits parameters to data and is a one-time, compute-heavy cost; inference applies the fitted model to new inputs and its cost scales with request volume.

Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 = harmonic mean of precision and recall. Moving the decision threshold trades one for the other, so metric choice must match the cost of each error type.

Overfitting = low training error with high generalization error; mitigations include more or better data, regularization, cross-validation, and early stopping.

Bias enters at data (sampling, labeling, historical), model (objective function, feature choice), and deployment (population shift, human use); mitigation must target the specific stage.

Narrow AI (ANI) is task-bounded; artificial general intelligence (AGI) is hypothetical; no deployed commercial system meets an AGI definition.

Common enterprise use cases: predictive scoring, recommender systems, computer-vision inspection, NLP extraction and generation, anomaly detection, and process automation - each with distinct accuracy and failure-mode profiles.

Impact classes: labor substitution vs. augmentation, allocative effects of scoring, accessibility gains, opacity that reduces contestability, and energy/water cost of training and inference.

EU AI Act (Regulation (EU) 2024/1689) entered into force 1 August 2024 and uses a four-tier risk pyramid: unacceptable (prohibited), high, limited-risk transparency duties, and minimal risk.

EU AI Act application dates: prohibitions 2 February 2025; general-purpose AI obligations 2 August 2025; high-risk Annex III systems 2 August 2026; Annex I product-embedded systems 2 August 2027.

EU AI Act prohibited practices include public-authority social scoring, subliminal or manipulative techniques, exploiting vulnerabilities, untargeted facial-image scraping to build databases, emotion recognition in workplaces and schools, and most real-time remote biometric identification in public spaces.

NIST AI RMF (AI 100-1, January 2023) is voluntary and organized around four functions - Govern, Map, Measure, Manage - and is paired with the Generative AI Profile (NIST AI 600-1).

OECD AI Principles (2019, updated 2024): inclusive growth, human rights and democratic values, transparency and explainability, robustness/safety/security, and accountability; they underpin the G20 and EU AI definitions.

ISO/IEC 42001 specifies a certifiable AI management system; ISO/IEC 23894 provides AI risk-management guidance; ISO/IEC TR 24028 addresses AI trustworthiness.

GDPR Article 22 gives a right not to be subject to solely automated decisions with legal or similarly significant effects; Article 35 requires a DPIA for high-risk processing.

Sub-national and state examples: NYC Local Law 144 requires annual independent bias audits and candidate notice for automated employment decision tools; Colorado's AI Act imposes reasonable-care duties on developers and deployers of high-risk systems.

Sector regulators still apply: FDA for software as a medical device, SEC for robo-advisory, EEOC/Title VII for hiring algorithms, and FTC Act §5 for unfair or deceptive AI claims.

Life-cycle stages: problem framing → data acquisition and preparation → model development → validation and evaluation → deployment → monitoring → retirement; governance controls attach at each stage.

Risk = likelihood × impact (consequence). Inherent risk is scored before controls and residual risk after; acceptance decisions compare residual risk against a defined risk appetite.

Data drift = input distribution changes; concept drift = the input-to-output relationship changes; model drift = observed performance decay. Each requires a different monitoring signal.

Bias mitigation timing: pre-processing (data rebalancing, relabeling), in-processing (fairness constraints in the objective), post-processing (group-specific threshold adjustment).

Testing layers: unit and functional tests, held-out and out-of-distribution evaluation, red teaming for adversarial and misuse scenarios, shadow or canary deployment, and A/B tests with guardrail metrics.

Required artifacts: model card (intended use, training data, subgroup metrics, limits), datasheet for datasets (provenance, collection, preprocessing), system card, and an AI impact assessment.

Human oversight modes: human-in-the-loop (approve each action), human-on-the-loop (monitor and intervene), and human-out-of-the-loop; oversight depth should scale with autonomy and harm severity.

Third-party controls: vendor due diligence, contractual audit and incident-notice rights, model-version disclosure, and exit or portability plans for foundation-model dependencies.

Deployment safeguards: rate limits, output filters, confidence thresholds, kill switch, and a documented rollback path; production incidents feed back into the Map and Measure functions.

The alignment problem: optimizing a specified objective can diverge from operator intent through specification gaming or reward hacking; RLHF and rule-based training are mitigations, not guarantees.

Interpretability split: inherently interpretable models (linear, tree, GAM) versus post-hoc explanation (SHAP, LIME, counterfactuals); post-hoc explanations are approximations and can be unfaithful.

Generative AI copyright debates: whether training on copyrighted works is fair use or fair dealing is jurisdiction-specific and largely unresolved; EU law grants text-and-data-mining opt-outs to rights holders.

Synthetic media: watermarking and provenance standards such as C2PA Content Credentials are voluntary and can be stripped; detection is probabilistic and degrades as generators improve.

Open-weight vs. closed models: open weights increase external scrutiny and local control but complicate revocation of harmful capabilities and allocation of liability.

Security debates: prompt injection and jailbreaks are unsolved at the model layer; data poisoning attacks training pipelines; agentic systems widen the attack surface through tool and system access.

Concentration-of-power concerns: compute, data, and talent bottlenecks favor a few firms; compute governance (chip export controls, compute thresholds, cloud reporting) is a proposed policy lever.

Accountability gaps: the many-hands problem in multi-party AI supply chains complicates liability; proposed fixes include deployer-centric duties, mandatory audit rights, and AI insurance.

Environmental and labor impacts: training and inference energy and water use, plus hidden data-labeling labor, are increasingly included in AI impact assessments.