AI Safety, Alignment & Misuse Basics

This closing chapter of the module introduces AI safety at a basic, practical level, not the more speculative long-term questions that get significant media attention, but the concrete, present-day categories every engineer building with AI should recognize.

• -Accidental harm happens when a system does something unintended and harmful because it was not built or tested carefully enough: a hiring model unintentionally discriminating due to biased training data, a chatbot confidently stating false information as fact.
• -Deliberate misuse happens when a person intentionally uses an AI system, or manipulates its inputs, to cause harm it was never intended to enable, like using a generative model to produce convincing scams, or intentionally crafting inputs designed to bypass a system's safety guardrails.

These require genuinely different mitigations. Reducing accidental harm looks like the practices from earlier chapters in this module: better data, bias audits, careful evaluation. Reducing deliberate misuse looks more like the guardrails, input and output filtering, and abuse-monitoring systems this course revisits directly in Phase 2's safety-focused chapters.

Alignment refers to the general challenge of making sure an AI system's actual behavior matches what its designers genuinely intended, rather than technically satisfying its stated objective in some unintended, undesirable way. Here is a simple, concrete example. A recommendation system asked to maximize watch time might learn that increasingly extreme or attention-grabbing content keeps people watching longer, technically achieving its literal goal while working against what its designers actually wanted for their users.

You do not need to resolve deep, unsettled research questions to build responsibly with AI today. You need the habit of asking, for any AI system you help build, what is this actually optimizing for, does that match what we truly want, and what happens if someone deliberately tries to misuse it. Those three questions carry a surprising amount of the practical safety work most engineering teams actually need to do.