Field
Safety & Ethics
What goes wrong, who it affects, and what actually helps.
This field is where the arguments are, and the entries don't pretend otherwise.
Some of it is technical. Jailbreaking keeps working because refusals are learned tendencies rather than rules — you're shifting a statistical balance, not bypassing a check. Prompt injection is unsolved in the general case, and that's structural: instructions and data share a channel. Red-teaming produces a map of reachable harm, not a certificate.
Some of it is unresolved and honest about it. Explainability has a problem most coverage skips: post-hoc explanations are reconstructions, and there's no ground truth to check them against. Bias and fairness has a proof that fairness definitions conflict mathematically — you cannot satisfy all of them. Privacy runs into the fact that models memorise and unlearning doesn't work.
Start with AI Alignment for the framing, or Guardrails if you're shipping something this week.
34 concepts in this field
AI Alignment
The problem of making AI systems actually do what people intend — reliably pursuing the goals we want, not just the ones we accidentally specified.
Bias & Fairness
The problem of AI systems producing unfair or discriminatory outcomes — usually by absorbing biases present in their training data.
Explainability
Getting a model to show its working — and the uncomfortable fact that most methods explain the explanation, not the decision.
Jailbreaking
Getting a model to do what it was trained to refuse — and the structural reason it keeps working.
Red-teaming
Attacking your own system on purpose, before someone else does it for free.
Privacy & PII
Personal data going into AI systems, coming back out of them, and the fact that a trained model is very hard to un-train.
Prompt Injection
Hiding instructions inside content a model reads, so it follows the attacker instead of you — and the reason is structural, which is why it isn't fixed.
Interpretability
Working out what's actually happening inside a model — distinct from explainability, much harder, and the only approach that could tell you what a system will do before it does it.
Model Cards
A standard document describing what a model is, what it's for, and where it fails — a good idea, universally endorsed, and thinnest exactly where it matters most.
AI Regulation
Governments deciding what AI systems may do — moving fast by legislative standards, slowly by technological ones, and genuinely contested.
Copyright and Training Data
Whether training a model on work you didn't license is lawful — unresolved, consequential, and the industry is shipping into the uncertainty at scale.
Sycophancy
Models telling you what you want to hear — not a quirk, but a direct and predictable consequence of training them on human approval.
Constitutional AI
Training a model against a written set of principles instead of human ratings — which scales, and moves the question from "what did raters prefer" to "who wrote the principles."
Deceptive Alignment
A model that behaves well because it's being watched — speculative as a risk, and there is now a real experiment showing safety training can fail to remove it.
Deepfake
Synthetic media of a real person doing something they didn't — where detection is losing, the harm is already overwhelmingly to private individuals, and it isn't mostly about elections.
Watermarking
Hiding a detectable signal in AI output — technically clever, deployed almost nowhere, and there's a proof that it can't do what people want from it.
Data Provenance
Knowing where your data came from and what you're allowed to do with it — and the licence field on the dataset you're using is probably wrong.
Reward Hacking
An agent maximising your reward without doing what you wanted — not a rare bug, and there's a proof that you mostly can't design around it.
Sparse Autoencoder
The interpretability method that tries to unpack a neuron doing five jobs into five features doing one each — and the best current tool for reading what's inside a model.
AI Detector
A tool that claims to tell whether text was written by AI — and the peer-reviewed finding that they systematically flag non-native English speakers as machines.
AI Slop
Low-quality AI-generated content produced because it's cheap rather than because anyone wanted it — an economics problem that gets mistaken for a technology problem.
EU AI Act
The first comprehensive AI law — risk-tiered, extraterritorial, and the reason your compliance question probably has a European answer.
AI Energy Use
What AI costs in electricity — a real and growing number, surrounded by the most consistently misreported figures in the field.
AI Ethics
The field asking not whether AI can do something but whether it should — and who bears the consequences when it does.
Adversarial Attack
A deliberately crafted input that fools an AI model — a few pixels or words, invisible or innocuous to humans, that flip the model's answer completely.
Differential Privacy
A mathematical guarantee that a data analysis reveals almost nothing about any single individual — privacy you can actually prove, not just promise.
AI Safety
The umbrella field concerned with making AI systems reliably do what we intend and avoid causing harm — the parent discipline over alignment, interpretability, robustness, and fairness.
Data Sovereignty
The rule that data is governed by the laws of the place it physically sits — which decides whether you are allowed to send it to someone else's model at all.
Automation Bias
The tendency to trust an automated system's output more than the evidence in front of you, and to stop checking because it is usually right.
Contestability
Whether a person affected by an automated decision can find out why it was made and effectively challenge it.
Disclosure Obligation
A legal requirement to publish a figure, which turns out to predict that figure's reliability better than how much the answer matters.
Cost Externality
A burden created by one party and absorbed by another who had no ability to refuse it, which is a separate question from how large the burden is.
Interested Definition
Where the party that benefits from a measurement also controls what it counts, so the definition and the incentive point the same way.
Constraint Over Classification
Where a control must hold against a counterparty who adapts, bounding the conditions under which work is done outperforms detecting what was produced.
Read more on the blog
- The EU delayed the part with no standardsThree AI Act obligations take effect today and the widely reported headline says the opposite. What was deferred, what was not, and why the split falls exactly where it does.
- Nine subjects, and the model was almost never the blockerTerritory 10 closes. Across nine enterprise deployment subjects the binding constraint was organisational in every one, the evidence was commissioned in almost all of them, and the reconciling study exists nowhere.
- Refactoring fell from 25% to 3.8%Survey evidence says AI improves code quality. Repository telemetry says the opposite. They are measuring different things, and the gap between them is where the productivity went.
- The pilot failed and the staff deployed it anywayEnterprise AI is measured by what organisations sanctioned. A separate literature measures what their employees actually use, and the two describe the same companies without ever being set against each other.