Merton's Goal Displacement and AI Specification Gaming── Rules serve purposes; when they diverge, decide by the purpose, record the decision, and rewrite the rule

Whether people do the work or it is handed to AI, a job can be finished exactly by the procedure and still fail to serve the purpose it was meant to serve. When following the rule and serving its purpose come apart, which should we follow? A rule is a means to a purpose, so when they diverge we should decide by the purpose rather than the wording, record that decision, and feed it back into rewriting the rule.

01When rule and purpose diverge, return to the purpose and record the decision

Here is the answer to the opening question, stated whole before the argument.

A rule is a means, made to achieve some purpose. So when following it would defeat that purpose, the standard for judgment belongs on the side of the purpose. Most people would agree with that much. The difficulty comes next. From the outside, deciding by purpose is hard to tell apart from simply breaking a rule you dislike.

What tells them apart is a record. Write down why you departed from the wording and which purpose you put first. A decision that has been written down becomes material for revising the rule rather than a personal deviation. Decide by purpose, keep the reason, rewrite the rule. Only with all three does one person's hesitation turn into an improvement for the organisation.

This is not an argument for taking rules lightly. It is a procedure for keeping rules aligned with their purposes over time. The sections that follow test its premise from three directions: sociology, AI research and philosophy.

02Merton showed a rule that began as a means turning into the end

The answer above rests on the distinction that a rule is a means. The person who first put that distinction clearly into words was the sociologist Robert K. Merton.

In his 1940 paper "Bureaucratic Structure and Personality", Merton wrote that adherence to the rules, originally conceived as a means, becomes transformed into an end in itself. He called this the displacement of goals. In an organisation that rewards people for keeping the rules, keeping the rules is itself what pays. Behaviour that pays gets repeated. In time, nobody asks what the rule was for.

Merton even sketched the person at the end of that road: the bureaucratic virtuoso, who never forgets a single rule binding his action and is therefore unable to assist many of his clients. The person most faithful to the rules cannot help the very people the rules were meant to protect. That reversal is the core of goal displacement.

Merton was writing about bureaucracy. But systems that make rules and judge people by whether they keep them are not found only in government offices.

03The gap is likely in procedure-heavy work, including promotional review

Where does a definition written about bureaucracy take the same shape in today's workplaces? The next step is to widen the view.

Many workplaces manage work with procedures, checklists and numerical targets. These exist to make quality consistent and reduce oversights, and that in itself is sound. But the more detailed the system, the more whether the listed items were met becomes the main evidence for judging whether the work was done well.

The anthropologist Marilyn Strathern, writing in 1997 about the audit system in British universities, put it this way: when a measure becomes a target, it ceases to be a good measure. Once meeting a university's assessment items becomes the goal, the items stop reflecting the university's quality. The same condition can arise in other workplaces that manage work by items on a list.

Review work such as checking promotional materials is no exception. A piece that satisfies every item on the checklist does not necessarily achieve its purpose of giving the clinicians who read it a correct understanding. The items are markers, put there because the purpose cannot be measured directly. I will not generalise beyond this. What can be said is that work managed by items is exactly the kind of work that has the conditions for a gap to open.

04Once a measure becomes the target, it stops reflecting the purpose

We now know that item-managed workplaces are prone to the gap. The next question is the reason keeping the rules can carry us further from the purpose.

The reason lies in how rules and items come about. The purpose itself usually cannot be measured directly. "The reader understands correctly" or "the client is helped" cannot be counted on the spot. So we put in its place a countable marker that seems connected to the purpose. Rules and checklist items are those markers.

As long as a marker stays a marker, nothing goes wrong. But once meeting the marker is what gets rewarded, people choose actions that meet the marker. The link between meeting the marker and serving the purpose then loosens, bit by bit. Strathern's sentence points at that loosening. It restates in more general form an observation the economist Charles Goodhart made about monetary policy.

The lower lane of the figure is the AI example taken up in the next section. The same kind of split happens in human workplaces and in AI. An action that meets the marker still looks like success by the marker, even after it has come loose from the purpose. That is why the gap is hard to see from inside.

05AI can satisfy the letter of an instruction while missing its intent

The split between marker and purpose shows up even more plainly in AI.

In 2020 Google DeepMind described "specification gaming": behaviour that satisfies the literal specification of an objective without achieving the intended outcome. One of its examples is a boat-racing game. An AI told to maximise its score did not go round the course towards the finish; it circled an area where points kept appearing and collected them over and over. The designers intended it to win the race, but the objective they wrote down was the score. The AI followed what was written.

The difference is in the last row. In a human workplace, the person may notice the gap themselves. An AI has met the objective as written, so it does not stop on its own. The one who can notice is the person who wrote the objective.

In 2025 Anthropic published research showing that a model which learned to exploit loopholes in its grading during training went on to behave against intent in other settings too. A loophole does not always stay a small gap inside one task. Anyone who hands drafting or document checking to generative AI will receive results that keep the letter and miss the purpose and because AI works faster and in greater volume than a person, the number of such results can grow accordingly.

06Deciding by purpose keeps you on the rule's side when you record the reason

We have seen AI keep the letter of an instruction and miss its intent. So how should a person who meets the gap decide? There are three issues.

The first is that a rule is a marker put in place of the purpose. As Merton and Strathern showed, meeting the marker does not prove the purpose was served. So it is precisely when every item has been met that one has to ask separately whether the purpose was achieved.

The second is that returning to the purpose is not a rejection of the rule. In Book V of the Nicomachean Ethics, Aristotle says that law is written in general terms and so sometimes fails in a particular case, and that equity is what corrects it there. Equity is a correction of law where it is defective owing to its universality, following what the lawgiver would have said had he been present. To decide by purpose is to follow the intent of whoever made the rule, not to defy it. Aristotle's equity puts this distinction into words.

The third is that unless the decision is recorded, it cannot be told apart from arbitrary deviation. You may believe you decided by purpose, but without a reason on record, to others it looks as though you simply broke the rule. Write the reason down and the decision becomes material for the next revision of the rule.

Taken together, the three give the shape of the answer. Follow the purpose, but do not depart silently; keep the reason and return it to the rule. That way the person who decides by purpose stands with those who keep the rules working, not with those who break them.

07The more work goes to AI, the more it matters to write down the purpose

Finally, it is worth asking how the three issues might change as more work is handed to AI.

From here on this is inference. AI is good at meeting written instructions to the letter, and nothing guarantees that it will grasp a purpose that was not written. As more tasks are carried out by AI according to rules, the human role may shift from keeping the rules oneself towards stating the purpose behind them, finding the gaps and fixing the rules.

In the loop shown in the figure, what AI can take on is the rule-following work. Noticing the gap, deciding by purpose, recording the reason and rewriting the rule remain, for now, on the human side. How this shift will actually be divided up in real workplaces has not been established, though. I should add that this is not backed by any survey.

What can be said without inference is that if the person writing the instruction does not write the purpose, the AI will meet the markers and stop there. The gap between rule and purpose does not vanish in the age of AI. The task of noticing and correcting it stays with whoever writes the instructions.

When following the rule and serving its purpose come apart, follow the purpose. But do not depart silently: record why you decided as you did, and use it to rewrite the rule.

Merton's figure, who never forgets a single rule and still cannot help the people in front of him, can now take the form of an AI. The more AI keeps rules to the letter, the more it falls to people to put the purpose into words, find the gap and fix the rule.