The three terrains

Three kinds of ground, and the machine's role changes on each: statistical terrain, where it sees what you cannot; bias terrain, where it discriminates less than your current process; and override terrain, where the cost falls on a person and the call stays human. The framework is from SuperSkills, chapter seven, and it answers the question most adoption programmes never ask.

Why terrain rather than risk#

The published alternatives sort the wrong object. Regulation sorts systems into risk tiers, which is a question for whoever procures the system. Human factors sorts functions across levels of automation, which is a question for an engineer at design time. Both leave the person in the room without an answer about the work in front of them.

Terrain is the ground you are standing on for this decision, and it changes within a single job. The same clinician works in all three before lunch. Sorting by terrain gives a different answer per decision rather than a policy per system, and that is what lets it survive contact with a working week.

1. Statistical terrain#

The ground of signal and noise, where machines outperform human perception. Models flag microcalcifications the eye cannot resolve. Fraud systems see movement patterns no analyst could track in a career. Nowcasting recognises a storm cell forming before a pilot would see the cloud.

The human task here is not to out-guess the model, which is the instinct and the wrong one. It is to interpret what the signal means, act on it with understanding, and know where the model's horizon ends. The failure mode on this terrain is not deference, it is a human competing on the machine's ground and losing while calling it judgement.

2. Bias terrain#

Some systems are fairer than we are. Recruitment shows it repeatedly: left alone, human shortlists skew towards the familiar, and a tested model monitored for fairness can flatten that and change who is seen at all.

Where the evidence shows the model discriminates less than the current process, let it propose. Keep the review for context, dignity and the final call. The qualifier does the work: tested, monitored, and shown to be better than what you have, rather than assumed to be. Augmentation here is humility, the recognition that a tool can correct what instinct protects.

3. Override terrain#

Then the decisions that belong irrevocably to human hands. A triage flag may push a patient to the front of the queue, and the doctor still sits down, looks them in the eye and explains the options. A risk model may signal rejection, and the banker still picks up the phone and asks one more question before saying no. An originality alert may fire on a student's essay, and the teacher still reads the work before deciding whether it was copying or a cry for help.

These are not inefficiencies. They are where trust lives. The cost of being wrong falls on a person rather than on a system, and that weight cannot be delegated to something that will not carry it. The accountability question this raises is treated at how AI decision rights should be allocated.

Mapping your terrain#

Before beginning any project with machines, write three lines underneath your rules:

• We trust the system first on these three patterns.
• We keep the call on these four.
• We always run an override in these two.

Writing it in advance is the whole of it. Afterwards, the answer bends towards whatever was convenient at the time, and the same mechanism is why custody lists are written before the tools arrive. That single act of clarity speeds the work without selling out judgement. It is also a reminder that a system's confidence is not the same as truth, and that the strongest signal in any system is still a human saying they will take responsibility.

The pull this exists to resist#

There are days when the score reads green and the confidence reads 0.97 and you want to stop thinking. That is the tug. It is strongest at the end of a long shift, which is when the consequences tend to be largest. The research on automation bias and appropriate reliance describes the same pull from the outside and finds no intervention that reliably removes it. The terrain map does not remove it either. It decides in advance, in writing, where giving in to it is allowed.

What the three terrains do not do#

They do not tell you which terrain you are on. That judgement is the thing being exercised, and a team can put a decision in statistical terrain because it is convenient to. Nothing here has been tested: no study has compared teams that mapped their terrain against teams that did not, and the framework is offered as a way of deciding rather than as a measured intervention. The bias terrain claim in particular depends entirely on evidence for the specific process, and a model assumed to be fairer is not an instance of this framework but the failure it warns about.

Source#

Rahim Hirji, SuperSkills: The Seven Human Skills for the Age of AI, Kogan Page, 2026, chapter seven, "When to trust the machine and when not to". The framework and the naming are his. The worked cases and the argument around them are in the book.

Related SuperSkills research#

• The use-or-keep test, the task-level companion to this map
• Hand, Head, Hours, Heart
• Augmented Mindset
• Keep, share, hand over