Sharlene Fernandes

Human kindness, especially in relationships, revolves around exchanging empathy. Empathic AI—large language models (LLMs) that provide empathic support—will undermine human kindness by stopping people from 1) seeking empathy and 2) giving empathy. Asking humans for empathy is often an imposition and can be expensive (e.g., therapy). Also, the empathy people receive from humans is often unwelcome advice (rather than pure sympathy). Because LLMs are always available to provide consistent and unconditional compassion, “seekers” will rely on LLMs instead of their relationship partners. And when seekers rely on AI, “givers” will happily offload their empathic responsibilities to save themselves from this uncomfortable and emotionally costly work. With seekers and givers both relying on LLMs, the kindness created by exchanging empathy will disappear—unless humanity uses empathic AI only to reinforce human empathy, not replace it. The future of human kindness hinges on how we choose to use LLMs.

Prison sentences that exceed the natural lifespan present a puzzle because they have no more power to deter or incapacitate than a single life sentence. In three survey experiments, we tested the extent to which participants support these longer-than-life sentences under different decision contexts. In Experiment 1, 130 undergraduates made hypothetical prison sentence length recommendations for a serious criminal offender, warranting two sentences to be served either concurrently or consecutively. Using a nationally representative sample (N = 182) and an undergraduate pilot sample (N = 260), participants in Experiments 2 and 3 voted on a hypothetical ballot measure to either allow or prohibit the use of consecutive life sentences. Results from all experiments revealed that, compared to concurrent sentences without the possibility of parole, participants supported the use of consecutive life sentences for serious offenders. In addition, they adjusted these posthumous years in response to mitigating factors in a manner that was indistinguishable from ordinary sentences (Experiment 1), and their support for consecutive life sentencing policies persisted regardless of the default choice and whether the policy was costly to implement (Experiment 2). These judgment patterns were most consistent with retributive punishment heuristics and have implications for sentencing policy and for theories of punishment behavior.

Advances in artificial intelligence (AI) raise important questions about whether people view moral evaluations by AI systems similarly to human-generated moral evaluations. We conducted a modified Moral Turing Test (m-MTT), inspired by Allen and colleagues’ (2000) proposal, by asking people to distinguish real human moral evaluations from those made by a popular advanced AI language model: GPT-4. A representative sample of 299 U.S. adults first rated the quality of moral evaluations when blinded to their source. Remarkably, they rated the AI's moral reasoning as superior in quality to humans’ along almost all dimensions, including virtuousness, intelligence, and trustworthiness, consistent with passing what Allen and colleagues call the comparative MTT. Next, when tasked with identifying the source of each evaluation (human or computer), people performed significantly above chance levels. Although the AI did not pass this test, this was not because of its inferior moral reasoning but, potentially, its perceived superiority, among other possible explanations. The emergence of language models capable of producing moral responses perceived as superior in quality to humans’ raises concerns that people may uncritically accept potentially harmful moral guidance from AI. This possibility highlights the need for safeguards around generative language models in matters of morality.

Do people punish more than they would if the decision costs were more transparent? In two Internet-based vignette experiments, we tested whether juvenile sentencing recommendations among U.S. adults are responsive to variation in the salience of the taxpayer costs and public safety benefits of incarceration. Using a 2 Cost (present vs. absent) x 2 Benefit (present vs. absent) factorial design, Experiment 1 (N = 234) found that exposure to information about the direct costs of incarcerating the juvenile offender reduced sentencing recommendations by about 28%, but exposure to the public safety benefits had no effect on sentences. Experiment 2 (N = 301) manipulated cost-benefit salience by asking participants to generate their own list of costs of incarceration, benefits of incarceration, or an affectively neutral, unrelated word list. Results revealed a similar selective effect whereby sentencing recommendations were reduced in the cost condition relative to the benefits and control conditions, but sentences in the benefit condition did not differ from the control. This combined pattern suggests that laypeople selectively neglect to factor cost considerations into these judgments, thereby inflating their support for punishment, unless those costs are made salient. These findings contribute to the debate on transparency in sentencing.