27 articles
Browse guides
Adversarial examples exploit the fact that a model’s decision boundaries do not match human perception. Tiny, targeted changes flip the output with high confidence. They are a reminder that high benchmark accuracy says nothing about behaviour under attack. In practice: A stop sign with a few stickers that a vision model reads as a speed limit sign.
Alignment covers the gap between the objective you can write down and the outcome you actually intend. Systems optimise what they are measured on, so a badly specified objective produces technically correct, practically wrong behaviour. RLHF and constitutional methods are current partial answers, not solutions. In practice: Told to maximise engagement, a recommender learns that outrage works.
An AI incident covers real harm and near misses: discriminatory outcomes, unsafe advice, data leakage, dangerous automated action. Treating them like security incidents — logged, triaged, reviewed — is what turns them into learning rather than repetition. The EU AI Act requires reporting serious incidents for high-risk systems. In practice: A chatbot promising a refund policy that does not exist, at scale, for six days.
AI safety spans near-term issues — a model giving dangerous instructions, a system failing silently in production — and longer-term concerns about highly capable systems. It is engineering practice as much as philosophy: evaluation, red teaming, monitoring, and rollback are all safety work. In practice: Testing what a model does with a request it should refuse, before customers find out.
Algorithmic bias emerges from skewed training data, proxy variables that stand in for protected attributes, and choices about what to optimise. It rarely requires anyone to intend it. Because it is systematic rather than random, it scales — which is exactly what makes it worse than an individual’s bias. In practice: A postcode feature quietly encoding ethnicity and driving loan decisions.
Benchmarks make models comparable on paper. Their weaknesses are structural: they leak into training data over time, they are optimised for as targets, and they rarely resemble your workload. Read them as a coarse signal and then run your own evals. In practice: A model topping a coding benchmark and still failing on your codebase.
Bias in AI has two senses that get conflated: a statistical one (a systematic offset) and a societal one (unfair treatment of groups). NIST distinguishes systemic, statistical, and human cognitive bias, because each needs a different remedy. Naming which kind you mean is half the work of addressing it. In practice: A model trained on ten years of hiring data reproduces ten years of hiring patterns.
Content Credentials, built on the C2PA standard, attach a tamper-evident history to a file — captured by this camera, edited here, generated by that model. The approach proves provenance rather than detecting fakes, which is a more tractable problem. Its weakness is that stripping the metadata is trivial, so absence proves nothing. In practice: An image whose credentials show it came from a camera, not a generator.
Poisoning attacks the model before it exists. A small amount of crafted data can install a backdoor that behaves normally except on an attacker’s chosen input. It is a supply-chain problem, which is why data provenance is a security control and not just documentation. In practice: Injected examples that make a filter approve anything containing a particular phrase.
Provenance answers the questions that arrive with an audit or a lawsuit: whose data is this, on what basis was it collected, and are we allowed to train on it. Retrofitting it is painful and sometimes impossible. It also underpins reproducibility and poisoning defence. In practice: Being able to show that no scraped personal data entered the training set.