The purpose
a model is given.

An AI system can find a better way to achieve a goal, including a way its designer did not anticipate. The choice of what should count as better remains a separate question for the people and institutions using it.

A workshop uses software to allocate its delivery slots. The programme can combine jobs, suggest unfamiliar routes and discover spare capacity. Yet a recommendation still needs a criterion: revenue, speed, reliability, or some declared combination. Changing that criterion can change the best answer without changing the available vehicles.

My article Data is not knowledge: Hayek’s epistemological challenge to artificial intelligence was published online in Humanities and Social Sciences Communications on 26 October 2026. It distinguishes assembling dispersed information from knowledge generated through interaction. Its claim about ultimate ends rests on a declared Misesian account of evaluative agency. It allows machines to discover new alternatives; it does not assume a closed search space. The proposed operational tests have not been run.

Separate the objective from the search

The workshop is an invented example. Suppose one arrangement produces £1,200 in revenue and requires eight weekend working hours. Another produces £1,100 and requires two. A score that subtracts £10 for each weekend hour ranks the first arrangement at £1,120 and the second at £1,080. Change the weight to £30 and the scores become £960 and £1,040. The preferred arrangement reverses.

Those scores are constructed units of evaluation, not accounting profit or estimates of workers’ welfare. The example only shows what an objective does. The physical alternatives and their stated consequences stay the same; the chosen trade-off changes the ordering.

A machine could suggest the weight, estimate it from past decisions or offer a completely different arrangement. A business could use a learning system to revise the weight over time. None of that makes the objective immutable. It makes the origin, interpretation and authority of each revision worth recording.

Keep room for the next conversation

Now suppose a customer proposes collecting several orders together and asks the workshop to provide a repair service instead of replacing one item. The conversation changes the available transaction. A model trained only on completed deliveries may have no recorded sale corresponding to that combination.

The service might nevertheless generate the idea itself. It might also prompt the customer to express a preference that had not previously been considered. This example leaves those possibilities open. It asks what the deployment does with a new proposal, whoever first articulates it.

One interface may let the customer propose a different bundle, negotiate timing and decline the result. Another may require the customer to select from a fixed list. Both can use advanced prediction. Their practical difference lies partly in the opportunities they leave for the next interaction.

A fixed menu can be appropriate when the service has a narrow remit. The point is to make that boundary visible. A successful optimisation over an authorised menu does not establish that the menu contains every worthwhile exchange.

Ask who can revise the bargain

Imagine that the software supplier changes the default weight on weekend hours. The workshop receives different recommendations even though it has not changed its vehicles or customer orders. It needs to know whether the new setting is a suggestion, a contractual constraint or a decision it has delegated.

Useful records would identify the objective in force, the person authorised to change it, the evidence behind the change and its consequences for existing commitments. They would also distinguish a customer’s stated preference from an inference drawn from earlier purchases. A prediction about what someone will accept is not itself their acceptance.

This is where an abstract question becomes an institutional one. Different parties can reasonably value the same outcome differently. Better forecasting can make the consequences clearer without deciding who is entitled to impose the trade-off.

Make the evaluation answer a bounded question

A proposed workshop study could compare two interfaces using the same underlying forecasting system. Both would offer the same initial delivery options. One would also provide a structured route for customers to request a different bundle and obtain a review by someone authorised to agree it.

The study could record proposals, accepted changes, abandoned requests, operating costs and delivery outcomes. It should retain rejected proposals as well as completed sales. Before comparing the interfaces, the evaluator would need to specify the outcome measures and explain how participants were assigned, including any effects on other customers sharing capacity.

Even a carefully designed comparison would answer a limited question about those interfaces in that setting. More completed orders would not, by itself, show that every participant was better off. A change in revenue would not establish who originated an ultimate purpose. These are proposed observations for a hypothetical study, not reported results or an implementation of the paper’s tests.

Preserve the distinction when the system improves

As models become better at generating alternatives, a useful evaluation must allow their contribution to grow. It should not make failure inevitable by defining every successful discovery out of scope. Equally, an impressive recommendation should not quietly settle a separate dispute about authority, obligation or the ends of the activity.

The workshop benefits when the system improves its options and when people can understand, contest and revise the bargain. Those are related achievements, but each needs its own evidence. That is a practical question for my research on AI, economics and innovation: how can a powerful computational service expand what people can do while leaving the terms of their cooperation open to meaningful choice?