“Can ethics wait?” is a question I wanted the ethics fellowship to examine seriously. Early projects are uncertain. Teams need to test ideas before they know exactly what they are building. Requiring a complete moral theory before any experiment is an unworkable standard.

But waiting is itself a choice about which uncertainties the team will resolve and who will bear the cost. A prototype can be easy for its builder to discard and difficult for everyone affected by it to undo.

The useful question is which ethical uncertainty changes the next experiment. That is specific enough to influence engineering without pretending every value conflict can be reduced to a metric.

The builder’s rollback boundary

Consider a tool that writes to a database and triggers a notification. The developer can restore the database. The recipient cannot unread the message. From the builder’s console, the experiment looks reversible; from the recipient’s position, part of it is permanent.

This is why I connect the fellowship question to my work on agent control and recovery. Reversibility is a property of effects and available repair operations, not simply a property of the code repository. A team needs to know where the experimental boundary lies in the world it is touching.

The same structure appears outside software. A physical trial can consume material, expose a person to risk, or change an institution’s commitments. A later product decision does not automatically cancel those effects.

Information gained by one party, downside carried by another

An experiment can be valuable to the team while being a poor bargain for its participants. The builder learns whether a system works. The participant spends attention, accepts uncertainty, or supplies data that cannot be fully recalled.

This does not imply that risky experiments should never happen. It means the distribution of learning and downside belongs in the design. Consent, a credible exit, limited exposure, and useful compensation can change the bargain. Some effects remain difficult to price or compensate.

An internal review should ask whose preferences are missing from the success criterion. If the team measures only deployment speed and model accuracy, it may never observe the burden that determines whether the trial was acceptable to someone else.

Some questions get more expensive to defer

A design decision can constrain later options. If an agent is granted broad ambient authority, adding a reliable commit boundary later may require architectural changes. If provenance is discarded during memory compression, a later explanation interface cannot reconstruct it perfectly from a summary.

These are examples of ethical concerns with immediate technical consequences. They affect which logs to retain, what access to grant, and which effects to mediate. Deferring them can make the eventual fix expensive even before any harm occurs.

Other questions can reasonably wait for evidence. A team need not settle every possible future use of a narrow prototype before testing whether the core mechanism works. The distinction depends on whether the current experiment closes an important option or creates an effect that later deliberation cannot repair.

A proposed fellowship exercise

I would give participants the same project at three stages: a local simulation, a limited trial with volunteers, and a deployment with downstream users. At each stage, require a concrete next experiment, an account of affected parties, and a list of effects that remain after shutdown.

Then change one assumption: the system delegates, a participant withdraws, or the project becomes commercially valuable. Ask which earlier design choices still support the promised exit and which were only credible while the stakes were small.

This is an exercise I would use, rather than a report of a completed fellowship outcome. Its purpose is to make moral disagreement operational enough to alter a test. A team may still choose to proceed. It should be able to explain what it expects to learn, why that learning needs this exposure, and which consequences it has actually made recoverable.

Related notes