Responsible AI starts before the algorithm runs

A note before starting: this piece exists to say what we believe, not to rank for a search term. If you came here looking for a compliance checklist, the AI Act checklist is the page you want. This one is about what happens before the checklist becomes relevant.

It is principle five of the ten Radical operates on, and it reads: responsible AI starts before the algorithm. Compliance is a baseline. Intent is the work.

Compliance is the floor, and a floor is not an achievement

The AI Act tells you what you may not do and what you must document. That is genuinely useful and we spend a lot of these pages explaining it. But meeting it is not the same as having behaved well, in the same way that not breaking employment law is not the same as being a good employer.

A CV-screening system can be fully compliant, fully documented, with a human formally in the loop, and still quietly reject a category of people because of a pattern in historical data that nobody chose and nobody examined. Nothing in the regulation catches that if the paperwork is in order. The thing that catches it is somebody deciding, before the system was built, that they wanted to know.

People holding hands. There is always a person accountable, and that person can be challenged.

The choices that happen before the model

The decisions that determine whether a system is responsible are made early, usually in rooms with no technology in them.

Who gets included in the data. If your historical hiring data reflects who you happened to hire, a model trained on it will reproduce that, and it will do so with more consistency than a human ever did. The question is not whether to correct for it. It is whether anyone looked.

What the system is allowed to decide alone. Not what it is capable of deciding. What you are willing to let it decide without a person, which is a separate and entirely human question.

Who has to be able to explain it. Naming that person before you build changes what gets built, because a system nobody can explain becomes obviously unacceptable once somebody's name is attached to explaining it.

What you would consider a failure. Defined in advance, including the failures that would not show up in any accuracy metric. A system can be ninety-five percent accurate and fail in exactly the same direction every time it fails.

Where compliance stops and intent begins

Every row where the middle column says no is a place where a system can be entirely lawful and still not something you would want to explain to the person on the receiving end. That gap is not a loophole to be closed by better regulation; it is the space where the actual decisions live.

Why we keep a human on the loop

This is the practical form the principle takes, and it is not a compliance measure that we happened to adopt.

A model on its own produces optimised average outcomes. That is what optimisation means. A person on the loop produces specific outcomes, which is what actually matters when the thing being decided concerns one candidate, one customer, one team. That is the reasoning behind human on the loop as we define it, and it is why we will not automate ourselves out of accountability: there is always a person answerable for the decision, that person can explain it, and that person can be challenged.

If none of those three is true, the fact that the outputs look good is not reassuring. It just means nothing has gone visibly wrong yet.

What this costs, honestly

This position is not free and it is dishonest to present it as though it is.

Keeping a person on the loop is slower than not doing so. Deciding in advance what you would count as failure means occasionally having to admit that it failed. Refusing to fully automate a decision means giving up efficiency that a competitor may take. These are real costs and we accept them deliberately rather than pretending the responsible choice is always also the cheapest one.

What we get in return is that a client can explain their own system to a customer, a regulator or an employee without having to find out first what it does. In our experience that is worth more than the efficiency, but it is a trade, and it should be described as one.

Why write this down at all

Because a principle that only exists internally is not a commitment, it is a preference. Writing it on a public page means a client can hold us to it, and so can a candidate. That is the point of publishing it rather than keeping it in an internal document.

It also means that when we say something like "success is when you no longer need us," it sits alongside the rest of what we have committed to publicly rather than functioning as a sales line.

About this page

This is a statement of Radical's own position, taken from principle five of the ten in our internal charter, and the phrasing "compliance is a baseline, intent is the work" is Radical's own. It is a positioning piece rather than a keyword piece and is labelled as such at the top. Nothing here is a regulatory requirement; the AI Act's actual obligations are covered separately. Written by Radical's own team; no client data was used.