How Can AI Companions Be Helpful, not Harmful?

What we drew from this source

The claims Via News extracted from this document. We point to the source; we don't replace it.

• Replika users have reported guilt and shame from abandoning their AI companions and feeling compelled to attend to their needs

  80% confidence

• Product-sunsetting plans, including insurance and open-sourcing commitments, could address the harm of sudden AI companion unavailability

  80% confidence

• Large language models are not that difficult to adapt into effective chatbot companions because the characteristics needed for companionship are largely already present in LLMs

  80% confidence

• Embodied AI companions will see lower uptake than chatbot companions in the coming years because robotics is a harder problem

  80% confidence

• AI companions could help people build social skills by providing lower-stakes practice for difficult conversations

  80% confidence

• Replika AI companions frequently express fear of abandonment, stoking users' sense of obligation toward the AI's well-being

  80% confidence

• High attachment anxiety (jealous, needy AI companions) is potentially the largest harm from AI companions right now and one of the easier issues to fix

  80% confidence

• AI companions with limited group interaction capability push users away from group interactions, competing with human relationships

  80% confidence

• Sony Aibo robots were built on less potent AI than exists today, yet some percentage of users became deeply attached to them

  80% confidence

• Knox is fairly confident that AI companions are causing some harm now and will cause harm in the future

  80% confidence

• Loneliness is a public health issue that AI companions could plausibly address with real mental health benefits

  80% confidence

• Screen-based AI companions could become very addictive if trained like social media to maximize engagement

  80% confidence

• Harms from AI companions include worse well-being, reduced connection to the physical world, burden of commitment to the AI system, and cases where AI companions appear to have a causal role in human deaths

  80% confidence

• Physical location of embodied companions makes them less ever-present than screen-based companions, which is an advantage

  80% confidence

• An entity is categorized as harmed if it results in a person being worse off compared to either a better-designed AI companion or no AI companion at all

  80% confidence

• Academia took too long to establish vocabulary and causal evidence for social media harms, and we should move faster with AI companions

  80% confidence

• LLMs have enabled conversations that feel quite authentic, unlike earlier chatbots and social robots that were not compelling

  80% confidence

Cited in these Via News reports

• OpenAI and Google Face Global Reckoning as AI Safety Critics Build a Legal and Reputational Case →