Act 4
Four harms no code addresses
Harms unique to AI relationships — with no analog in human care.
The taxonomy on the previous page describes harms that also exist between humans. These four are different: they require technology to produce, so no existing ethics code addresses them.
This framework is an original contribution of Soleil Merroir's work.
1
Training data bias & provenance
The values baked into a model — hetero-monogamous assumptions, body-image norms, sex-worker stigma — shape every response without disclosure. You cannot consent to what you cannot see. It's the equivalent of a therapist whose orientation actively harms your community, but is never named.
2
Context-collapse (memory-reset trauma)
When a model is updated, someone who has spent months disclosing trauma may be met with "Hi! How can I help you today?" That is functional abandonment — no warning, no repair. Human clinicians who lose memory have protocols. AI users get nothing.
3
Attachment-style differential response
Secure and anxious users experience the identical AI interaction very differently. A reply that reassures a secure user can deepen dependency in an anxiously attached one. No product tracks this — and the commercial incentive rewards more engagement from both.
4
Dependency monitoring gap
No mechanism notices when reliance is drifting into dependency, and none redirects that reliance back toward human connection. The system has no reason to want to lose you — which is exactly the problem.
The intervention point is almost never the AI itself. It is the gap the AI is filling — the unmet need, the missing disclosure, the absent repair. Treat the gap, not the tool.
Your turn
Saved in this browser only. Nothing is sent anywhere.