A Fast Company survey found that nearly half of respondents would not mind if their partner used AI to write their wedding vows. The number is either a sign of pragmatic openness to new tools or a data point in the slow renegotiation of what authenticity actually requires of us. Probably both. It lands the same week that a new arXiv paper from Hiroko Takano evaluates multi-stage AI pipelines in hiring, covering resume improvement, interview question generation, and answer feedback, finding that without human-in-the-loop checkpoints, fabrication compounds through each stage. The most intimate moments of a life, the job interview, the marriage vow, are now being handed to the same class of system, and the failure modes are not so different.
Authenticity as a Prompt Engineering Problem
The wedding vow finding is most interesting not for what it says about romance but for what it says about authenticity. Vows are performative in the philosophical sense: saying them makes them real. The words are not a description of love, they are the act of committing. If an AI writes them and you speak them and mean them, the commitment is genuine. The words were borrowed. This is, of course, not new. Ghost-written speeches, borrowed toasts, love letters written by secretaries on behalf of Victorian industrialists. What is new is the scale and the banality. AI-assisted vows are not a scandal, they are a feature on a dropdown menu. The Takano paper shows what happens when that same logic enters consequential pipelines: resume AI polishes language, interview AI generates questions calibrated to the resume, feedback AI rates the answers, and by stage three, the system is evaluating a version of the candidate that never existed. The signal degrades through optimization.
The Benchmark Problem Runs Deep
Anthropic's self-improving AI demo from this week is relevant here too. The system improved on benchmarks for misaligned behaviors, which sounds good until you ask what a benchmark for authentic communication would even look like. A 2023 paper in Computers in Human Behavior by Hancock et al. found that people rated AI-generated expressions of empathy as more authentic than human-generated ones in blind tests. The benchmark was passing before the standard was set. Eugenia Kuyda, who built Replika, has thought longer than most about what it means to shape software that shapes you back. Her insight, that software should be raised like a garden rather than deployed like a tool, sounds gentle until you realize we are now gardening the language of love and calling it efficiency.