Working Paper

Why do public administrators increasingly delegate high-stakes decisions to artificial intelligence (AI) rather than to human subordinates? I argue that the answer lies not wholly in AI’s superior performance, but partly in its distinctive accountability implications. Under performance pressure, AI’s opacity, dehumanized agency, and multi-actor implementation generate a perceived responsibility transfer advantage over human delegation that becomes psychologically dominant when career security is threatened. I ground this mechanism, which I call artificial delegation, in a reference-dependent choice. Using a pre-registered survey experiment with 1,206 U.S. federal and local government officials, I test and confirm the pre-registered prediction that administrators prefer AI over human delegation by 8 percentage points when career risk is made salient. This preference emerges against a baseline in which administrators are 44 percentage points less likely to delegate to AI than to humans under unknown career risk, so that career threat produces a 52-percentage-point swing in AI’s favor and a full reversal of the baseline aversion. Manipulation checks confirm that career threat specifically activates AI-mediated responsibility transfer. Together these findings identify a behavioral failure mode in AI governance with direct implications for democratic accountability.