Product Digest AI Safety Foundation Certification

A structured introduction to AI safety, covering how models are trained to behave, where that training fails and what evaluation and interpretability can establish. Seven modules with an assessment of 75 questions.

Practice questions (75)Free. Sign in to answer, so your result is kept and reported by module.

At a glance

Cost

Free There is no examination fee and no course fee.

Format

Seven modules and a single assessment of 75 questions covering all of them. The reading is open to anyone, and the assessment requires an account.

Pass mark

70 per cent, or 53 of 75. The mark is taken across the whole assessment, and no module carries a pass mark of its own.

Prerequisites

None. The material assumes no machine learning background and contains no mathematics.

Renewal

None. The field moves, so a result records the syllabus as it stood on the day of the assessment. The Practitioner certification is the intended next step.

Based on

The a syllabus developed by Product Digest, drawing on the published curricula of BlueDot Impact, the Center for AI Safety, DeepMind and Stanford

What it covers

What it covers. Product Digest issues this certification. No external body accredits it and the assessment is not proctored. The certification gives a structured route through the material and a record that the candidate has completed the reading.

Foundation and Practitioner divide the subject between them. Foundation explains how a model comes to behave as it does, what evidence a team can gather about it, where an attacker can get at it and who governs the decision to release it. Practitioner covers the decisions and controls that follow once an organisation deploys such a system.

The material is written for a general product audience. Nothing in it requires prior experience of training a model.

The assessment is scenario based. Each question presents a situation or asks the candidate to draw a distinction.

The syllabus7 domains · free

Product Digest’s own domains. Each opens its own page listing the concepts beneath it, and each concept has a page of its own.

01Fundamentals of AI SafetyThe field separates risk into misuse, accidents and structural harm. Why each kind needs a different remedy, and where the disagreements between researchers lie.What AI safety means · How AI features fail · What generative AI is good and bad at · Catastrophic risk arguments and their critiques4 concepts
~16%02Training and Alignment of ModelsPre training, supervised fine tuning, reinforcement learning from human feedback and constitutional methods. What each stage adds, and why the helpfulness of a model is a product of that training.How models are trained to behave · Foundation models and how they work · Supervised, unsupervised and reinforcement learning · The alignment problem · Scalable oversight5 concepts
~20%03Failures of Specification and GeneralisationSpecification gaming, reward hacking and goal misgeneralisation. All three describe a model pursuing exactly the objective it was given, which is harder to catch than a model that disobeys.Specification gaming and goal misgeneralisation · Goal misgeneralisation · Robustness and distribution shift · Responsible AI in practice4 concepts
~16%04Capability, Scaling and EvaluationScaling laws and why capability arrives unevenly, what an evaluation can and cannot establish, elicitation and the dangerous capability evaluations frontier labs now run.Scaling and emergence · Capability evaluations · Agents and autonomy risk · Deceptive alignment and situational awareness4 concepts
~16%05Interpretability and TransparencyInterpretability research, what it has established and what it has not. The documentation that stands in for understanding while that research catches up.Interpretability · Transparency and explainability · Corrigibility and control3 concepts
~12%06Adversarial Attacks and Data PrivacyAdversarial examples, poisoning, model extraction and membership inference. What a model memorises from its training data and how an attacker draws it back out. The provenance marks that identify synthetic content.Adversarial machine learning · Memorisation and training data extraction · Content provenance and watermarking3 concepts
~12%07Release, Access and GovernanceThe bodies that govern AI, and how binding law sits alongside voluntary frameworks and company self governance. What a decision to release model weights changes about who holds responsibility.Who governs AI · Open weights and release decisions2 concepts
~8%

Try a question

A company finds that its recruitment model has been used by a manager to screen candidates in a way the company never authorised, using a workflow the manager built themselves. The model works exactly as designed. Which category of risk is this?

AMisuse, because the system worked as intended and somebody used it to cause harm.
BAn accident, because the model was applied to a task it was not evaluated for.
CStructural harm, because it reflects competitive pressure inside the company.
DMisalignment, because the model pursued an objective its designers did not intend.

75 questions across the 7 domains.

One per page, with a worked explanation.

Start the set