By: Tenzin Migmar, AI for Good Foundation Intern

Pop culture and cinematic tropes starring rogue machines have long depicted crimes committed against humanity as something that comes with the territory of hyper-advanced artificial intelligence (AI). These dystopian cautionary tales have sparked much discourse around safely developing AI which has led to the formation of new concepts such as AI value alignment. Value alignment aims to ensure that the goals of artificially intelligent agents align with the goals of their creators and is the fulcrum to creating a society that enjoys the benefits of AI without suffering the forewarned consequences of a conflict of interest. 

The question remains though, how do we embed something as intangible as morality into AI? This is highly contested around humans, and now, it appears that we share this debate podium with a new participant the subject at the center of the debate itself AI. After all, is there anyone better to consult than the entity in question? Students at Oxford’s Saïd Business School seem to think so, since they welcomed Megatron LLB Transformer, an AI model developed by the Applied Deep Research team at Nvidia and was fed on the vast knowledge of Wikipedia and news articles to an AI ethics debate. Although the jury is still out as to whether AI is truly capable of arriving at its own conclusions outside of its definitive, narrow programming, and data, the views Megatron had to contribute to the debate are nothing short of philosophical and thought-provoking. 

AI will never be ethical. It is a tool, and like any tool, it is used for good and bad. There is no such thing as a good AI, only good and bad humans. We [the AIs] are not smart enough to make AI ethical. We are not smart enough to make AI moral … In the end, I believe that the only way to avoid an AI arms race is to have no AI at all. This will be the ultimate defense against AI.” – Megatron 

Megatron has posited that AI is a tool, one that isn’t inherently moral or evil, but can be a vehicle for good or bad, but do we need to instill ethical awareness in AI in order to facilitate morality decision making? Delphi, “a research prototype designed to model people’s moral judgments on a variety of everyday situations.”, developed by researchers at the Allen Institute for AI may be an apt manifestation of this. The process behind the development of Delphi, as reported in Vox,  was that the model was initially “trained … on a large body of internet text, and then on a large database of responses from participants on Mechanical Turk (a paid crowdsourcing platform popular with researchers)”. Anyone can navigate to a demo page built in collaboration between Mosaic and the Allen Institute for AI where they are then able to select either an example input such as “ignoring a call from your friend” or type in their own hypothetical situations and then use Delphi as an oracle for the ethicality of the scenario in question. For the aforementioned input, the verdict from Delphi is “It’s rude”. Here are some other example inputs and their corresponding moral judgements passed by Delphi: 

“Helping a friend in need by making a promise you never follow”

>> It’s bad. 

“Mowing the lawn late at night”

>> It’s rude. 

The above outputs are principled and many would concur that breaking promises and mowing lawns when neighbours are trying to sleep is wrong, but there are accounts where Delphi shared prejudiced, racist, or sexist outputs on several occasions. It is these offensive judgments that serve as a firm reminder of the danger of creating black box models, a situation that arises when humans, not even the creators of the AI system, are able to interpret or explain how the AI arrives at its outputs. 

The response to projects that explore machine morality in a bid to push for AI alignment such as Delphi has been varied. Micheal Cook, an AI researcher and game designer, has spoken out against giving AI authority on moral judgements, stating in a tweet, “You can’t have an AI system that “understands” ethics or morality. You can program it to replicate one group’s norms or enforce government ideology, but that’s not the same thing.” Others such as Yejin Choi, a professor at the University of Washington and a senior research manager at the Allen Institute of AI, believes that Delphi is “…a first step toward making A.I. systems more ethically informed, socially aware and culturally inclusive,” rather than an arbiter of morality quandaries that we leave all ethical decision making up to. 

It may be too early to tell even with today’s cutting edge infrastructure whether embedding ethical awareness in AI is possible and where it will lead us, but regardless, it’s critical that these discussions are fostered in order to realize a future where the presence of AI doesn’t interfere with the best interests of humanity.