🌱 Seedling

When to Kill a Feature: Ethics as a Product Decision

· 2 min read
I have recommended killing 3 AI features in the past 18 months. Each was profitable. Each was ethically questionable. The hardest engineering decision is not building something difficult. It is deprecating something successful that should not exist.

When should an AI feature be killed for ethical reasons?

An AI feature should be killed when its continued operation causes measurable harm that cannot be mitigated without fundamentally changing the feature’s value proposition, regardless of its profitability or user adoption.

The 3 features I recommended killing: a behavioral prediction system that profiled users for manipulative notification targeting (profitable, increased engagement 14%), a sentiment scoring feature that rated customer service agents on emotional performance (reduced training costs, increased anxiety-related sick days 23%), and an automated content generation feature that produced misleading comparison articles (drove $340,000 in annual affiliate revenue).

Each recommendation was contested. The business cases were strong. The ethical problems were real but required deliberate attention to see. This is the core difficulty: profitable features generate organizational defenders. Ethical concerns generate organizational discomfort. The asymmetry of advocacy (profit has dedicated advocates; ethics often does not) means that ethical deprecation requires explicit organizational mechanisms, not just good intentions.

I do not have a formula for when to kill a feature. I have a diagnostic question: if the full impact of this feature on all affected people were publicly visible, would we still ship it? If the answer is no, the feature survives only because its harms are hidden. That is not a business strategy. It is a risk that compounds with time. The subtraction principle applies: sometimes the most valuable engineering decision is removing something.

The organizations that handled these decisions well shared a common trait: they had pre-established criteria for ethical deprecation in their decision records. The criteria were defined before a specific feature was at stake, which removed the emotional attachment and political pressure from the decision. The criteria were imperfect. But having any explicit framework for ethical feature deprecation was infinitely better than having none.