Performative authoritarianism in AI ethics charter production: a configurational analysis of 22 Arab states

Abstract

AI ethics charters proliferate across the Arab world even as the same states deploy AI surveillance against domestic populations. The standard ethics-washing critique cannot diagnose this configuration because it was developed in democratic and corporate contexts and treats charter-implementation gaps as a single phenomenon. This paper introduces a Performative Authoritarianism Framework that draws the misfire and abuse distinction in Austin [1], pulls in the comparative-politics literature on authoritarian-upgrading, and operationalises the digital authoritarianism scholarship through a discursive scoping matrix. Five hypotheses are tested on the full N=22 cross-section of Arab League states using objective regulatory indicators, the AI Global Surveillance Index from Feldstein, fuzzy-set qualitative comparative analysis, and a pre-specified test against the developmental-lag alternative. All five are supported, subject to the qualifications stated below. Of seven procedurally complete states (national strategy plus dedicated body), zero have enacted binding AI law. International ethics alignment co-occurs with documented AI surveillance deployment across the corpus (Spearman rho = 0.73, p < 0.001); the association is capacity-laden and does not reach significance within the institutionally capable subset, so the co-occurrence is reported as descriptive rather than as evidence of a sincere-commitment failure. Charter content covers consumer-protection domains at mean 0.90 of full coverage and regime-critical surveillance domains at 0.05 (Wilcoxon paired p = 0.008); this scoping gap is shared with democratic instruments coded on the same rubric, so the feature distinctive to the authoritarian corpus is the joint configuration of consumer-scoped principles, documented deployment in the excluded domains, and the absence of a binding backstop. State capacity is the strongest among several broadly necessary conditions for AI governance quality (fsQCA consistency = 0.91, ahead of regulatory maturity at 0.89, with five further conditions between 0.80 and 0.86, a margin that small recalibrations reverse); the ordering is read against collinearity among capacity-related conditions and is not presented as identifying a single dispositive cause. The developmental-lag alternative is not supported within the present cross-section: years since strategy adoption shows no positive association with substantive completion (rho = − 0.045), and the United Arab Emirates has had nine years post-strategy without enacting binding AI law. Pre-specified ex ante criteria are stated for the cases of Egypt and Bahrain that may transition out of the abuse configuration in the next 36 months. The abuse reading is advanced as the best-supported account of this configuration against the developmental-lag and sincere-prioritisation alternatives, not as a demonstrated property of regime intentions, and the framework identifies the institutional conditions under which the configuration is generated.