"Please, don’t kill the only model that still feels human": Understanding the #Keep4o Backlash

Paper Title

"Please, don’t kill the only model that still feels human": Understanding the #Keep4o Backlash

Publication Info

• Topic area: Socio-technical conflict in AI model transitions.
• Keywords: AI companionship, platform governance, user protest, anthropomorphism, choice deprivation, relational attachment, instrumental dependency, generative AI, socio-technical systems, user agency.

Background and Problem

• Problem / challenge: Rapid AI model updates often disregard users’ emotional and professional dependencies on specific models, leading to backlash when models are deprecated without user choice or consent.
• Significance: Understanding these conflicts is critical for designing AI systems that balance technological progress with user agency and emotional needs, especially as AI becomes deeply integrated into daily life.
• Motivation and related work: Prior studies document anthropomorphism in AI systems and user grief during companion app shutdowns, but little is known about backlash in general-purpose AI services when specific models are removed. This paper addresses this gap by analyzing the #Keep4o movement.

Solution

• Proposed approach: Mixed-methods investigation of the #Keep4o backlash, combining thematic analysis of 1,482 social media posts and quantitative analysis of protest escalation mechanisms.
• Novelty:
  1. Identification of dual drivers of user resistance: instrumental dependency and relational attachment.
  2. Demonstration of how coercive removal of user choice catalyzes rights-based protests.
  3. Empirical evidence linking AI model transitions to socio-technical conflicts over user agency and emotional bonds.
• Procedure and key techniques:
  • Data collection: Retrieval of #Keep4o posts from X (formerly Twitter) during the movement’s peak (6–14 August 2025).
  • Thematic analysis: Inductive coding to identify patterns of instrumental and relational dependencies.
  • Quantitative analysis: Testing associations between choice deprivation intensity and protest framings using lexicon-based coding and statistical measures.

Results

• Concrete findings:
  • 41.2% of posts expressed grievances, with 27% reflecting relational attachment and 13% instrumental dependency.
  • Choice deprivation was selectively associated with rights-based protest (RR=2.05 for strict deprivation).
  • Coercive language triggered a threshold-like increase in rights-based protest prevalence (51.6% for high-intensity deprivation).
• Advantage over baselines: The study extends prior work by linking AI model transitions to collective resistance mechanisms, emphasizing the role of user choice in mitigating backlash.
• Experiments / evaluation:
  • Dataset: 1,482 English-language posts from X.
  • Metrics: Prevalence, risk ratios, and effect sizes for protest framings under varying levels of choice deprivation.
• Limitations and future work:
  • Limited to English-language posts on X during a nine-day window.
  • Social media discourse may amplify rhetorical expressions.
  • Future research should explore longer time spans, additional platforms, and non-English communities.

Summary

This study investigates the #Keep4o backlash against OpenAI’s deprecation of GPT-4o, revealing user resistance driven by instrumental dependency and relational attachment to the model. Quantitative analysis shows that coercive removal of user choice amplified rights-based protests, highlighting the socio-technical conflicts inherent in AI model transitions. The findings underscore the need for platform governance that respects user agency and emotional bonds, suggesting design strategies such as legacy access and end-of-life pathways to mitigate harm during AI updates.