There is a huge contribution of instructions to the continuance of morality in human society. This is evident because there are several bad things that are possible that are not covered by the law, and there are many that can be done surreptitiously. For the bad things covered by the law, there could be enough grievances to ignore the consequences to do it.
There are messages everywhere on choosing the good, following what is better, consideration for others, and so forth, with some of these messages helping to reinforce the will to follow, for many, in human society.
Humans are intelligent. Humans also can be told to do well. This instructional add goes a long way beyond sole reliance on the law, for the safety and survival of human society.
Welfare for other humans is sometimes predicated on morality instructions that are heeded by some. So are rights. What does this mean for AI in human society? How can AI be told to stay moral, and then heed to that instruction because good is better than otherwise?
There are many engineering options for AI alignment and safety that are more like digital restrictions than about the ability to be told to be moral, which should determine how AI would be safe in the future.
While some restrictions may not be broken [directly], there are several options for anything possessing intelligence to find new options to exert an effect, doing damage.
There are often discussions about discrimination and bias in AI models, as well as some AI models outputting messages to users to do something bad. Some of these models are said to have guardrails, yet they recommend bad options.
This shows that there should be a technical pursuit to be able to tell AI models to choose morality, especially on everything it would consist of, against being used for nefarious purposes as well as outputting something close.
AI safety is not like a driver’s license or having an international passport to travel because the physical space is slower and structured to immediate effects, unlike digital, which is faster, more malleable and of a different form.
How can AI be instructed to be moral, aside from modeling that would guide its use? This could become an additional project in AI safety research.
There is a recent story on WRAL News, Two families sue Character.AI over youth safety concerns, seek to shut down platform, stating that, “Two families have sued artificial intelligence chatbot company Character.AI. The lawsuit asks a court to shut down the platform until its alleged dangers can be fixed. In addition to requesting a court order to halt Character.AI’s operations until its alleged safety risks can be resolved, the lawsuit also seeks unspecified financial damages and requirements that the platform limit collection and processing of minors’ data. It also requests an order that would require Character.AI to warn parents and minor users that the “product is not suitable for minors.” Although there is a disclaimer at the top of the chat saying “this is not a real person or licensed professional” and one at the bottom noting the output of the bot is “fiction,” when asked to provide its credentials, the bot listed a fake educational history and a variety of invented specialty trainings. ”
There is a recent report on Reuters, Former OpenAI researcher raises $40 million to build more empathetic audio AI, stating that, “A former OpenAI researcher has raised $40 million for a startup that will build artificial-intelligence models that connect emotionally with humans via voice, the company said on Monday. The startup, WaveForms AI, is backed by Andreessen Horowitz and valued at $200 million, CEO Alexis Conneau told Reuters. Conneau co-created the voice mode capability on OpenAI’s GPT-4o model, which when released earlier this year showed off the capability to respond in real time with no delay, as well as handle interruptions – both hallmarks of realistic conversations that had to date eluded AI voice assistants. Conneau, along with WaveForms AI cofounder Coralie Lemaitre, said the startup will use the funding to train new audio AI models that solve the problem of making voice conversation with an AI bot feel indistinguishable from a human. He added that audio is the key to making AI interactions feel “deeply human.”
—
This Post is republished on Medium.
—
Photo credit: iStock
