September 29, 2026
Most ethics review boards are created in good faith, often after something has already gone wrong. They meet. They ask thoughtful questions. They write minutes. Then the project moves ahead unchanged, and everyone quietly agrees the process is a formality to be endured rather than a decision point that matters.
The failures are rarely dramatic. They are structural, repeated, and surprisingly easy to fix once you name them. Here are the seven that show up most often, and what to do instead.
1. Treating the Review as a Formality
An ethics review that cannot change an outcome is not a review. It is a signature.
The tell is easy to spot: the meeting is booked after the design is frozen, the agenda says "for awareness", and nobody can remember a project ever being paused. If a team can schedule a review without risking anything, they will.
Fix it with a written scope of authority. State plainly which decisions the group can block, which it can attach conditions to, and which it only advises on. If the answer is "advise", route high-consequence cases to a named executive who can say no. Decide who may convene an emergency review, and on what grounds. Ambiguity here gets resolved in favour of the release schedule every time.
2. Asking Vague Questions
"Is this ethical?" invites a philosophy seminar and produces a shrug. Specific questions produce specific answers.
Replace the general prompt with a small set that forces detail:
- Who is affected if the system works exactly as intended?
- Who is affected when it fails, and how does that failure look from where they stand?
- Which group carries the cost but has no vote in this decision?
- What is the harm mechanism — a service denied, a record flagged, hours lost to an appeal?
- What evidence would change our minds?
The last question matters most. A review that cannot name the evidence it would accept is a conversation, not an evaluation.
3. Reviewing the Model but Not the System
A model is one component of a system, and often not the component that hurts people. The system includes how data was gathered and by whom, who labels it and under what working conditions, how outputs reach a human decision, what the interface nudges people toward, whether an appeal route exists, and how long a wrong decision stays on someone's file.
Ask for a one-page system map before the meeting: inputs, decision points, the humans in the loop, the known failure modes, and the route a person takes to contest an outcome. If nobody can draw it, the team does not yet understand what it is deploying. Review the map, not only the metrics.
4. Building a Board Without Authority or Range
Two mirror-image problems. The first: the board has no standing, so its conclusions get filed and ignored. The second: the board is twelve people from the same three teams, who share the same blind spots and reach consensus in ten minutes.
Fix both. Give the group a reporting line, a modest budget for outside expertise, and a standing slot in the release process. Then widen it: bring in legal, security, support and operations, add at least one person who does not work on the project, and include someone who will have to handle complaints after launch. Rotate members so the group does not harden into a clique.
5. Arriving Late and Leaving Early
The week before launch is the worst possible time for a first review. Options are expensive, timelines are public, and everyone is invested. Reviews held at that moment tend to approve with conditions nobody tracks.
Split the process into gates: concept, data acquisition, build, pre-release and post-release. Keep the early gates short — half an hour will do — and reserve depth for the points where change is still cheap. Then schedule a return visit. A review date a few months after launch, with real incident data on the table, catches what no pre-launch discussion can. Put it in the calendar rather than relying on good intentions.
6. Leaving Escalation to Improvisation
Sooner or later a reviewer and a team lead will disagree. If the process does not say what happens next, the outcome will be settled by seniority, volume, or whoever happens to be in the room that day.
Write the ladder down: reviewer, then the accountable owner for that risk area, then the executive sponsor, with the final decision recorded and a time limit at each rung. Name what can pause a release, and for how long. Separately, give staff a confidential route to raise a concern that bypasses their manager, because the person closest to the problem is often the one with the least power to describe it.
7. Letting the Paper Trail Disappear
Six months on, nobody remembers why a condition was accepted or who agreed to monitor it. That is how the same debate happens three times with three different answers.
Keep one page per review: what was assessed, what evidence was used, who was consulted, what was decided, the conditions attached, the named owner, and the next review date. Version it and store it somewhere searchable, not in a private folder or a chat thread. When an auditor or a new hire asks how a decision was reached, that page is the answer — and writing it forces the group to be honest about how thin the reasoning really was.
A Practical Reset for the Next Review
You do not need to rebuild everything. Pick one entry point and change it this month.
- Write the authority statement: what the group can block, condition, or only advise on.
- Add the five questions above to every agenda, and require written answers before the meeting.
- Ask for the one-page system map alongside the performance metrics.
- Name the escalation ladder and its time limits in that same document.
- Fix a post-launch review date before approving anything.
Then check the small things. Are meetings long enough to be uncomfortable? Is anyone in the room accountable for what happens to the people affected? Is there a record, and could a stranger follow it?
None of this guarantees good decisions. It does make bad ones harder to hide, which is most of what a review process is for. And where a review touches legal or regulatory obligations, bring in qualified counsel rather than settling the question around the table.
Photo: RDNE Stock project / Pexels
