AI ethics governance in EU-funded research is normatively rich and enforcement-hollow: grant agreements require ethics compliance; evaluation panels
approve proposals; Ethics Advisors file reports. Yet substantive compliance — the actual governance of sensitive data, meaningful interdisciplinary oversight,
protection of research subjects — is structurally avoidable throughout. This paper
presents an empirical account of how. Drawing on analytical autoethnography and
process tracing across 20 causal process observations from an MSCA Horizon Europe
doctoral network on hybrid intelligence, it documents a three-stage compliance
management pattern: (1) Refusal — a governance reform (dedicated ethics oversight)
was rejected at kick-off as bureaucratically impossible under the grant agreement
structure; (2) Concession — the same reform was implemented eighteen months later
under identical institutional conditions, falsifying the impossibility claim and
proving the barrier was motivational, not structural; (3) Neutralization — once the
oversight role was conceded, its function was prevented through documentation delays, incomplete submissions, and denial of requested resources, blocking any substantive
compliance assessment before the critical governance checkpoint. The finding is
analytically significant because it shows that the institution moved from refusing
the oversight mechanism to conceding it to neutralizing it, at each stage maintaining the formal appearance of compliance while avoiding its substance. The case was
selected precisely because it satisfies the conditions most favorable to principled
compliance — public institution, EC funding, explicit grant obligations, insider
advocate with formal standing — making the documented pattern a hard test, not an
easy one. Implications for the AI Act's governance architecture and the intermediary
function of Ethics Advisors are discussed.