An “ethical” AI trained on human morals may be racist

Can you believe it?

words by: Sahar Khraibani
Jan 21, 2022

Artificial intelligence can now give you life advice, according to science—and no, we’re not talking about psychics or a magic 8ball. Ask Delphi lets you examine any moral dilemma, but it may or may not provide you with a suitable response.


A group of researchers taught a piece of machine learning software how to respond to ethical conundrums. Ask Delphi, which was launched in late 2021 by the Allen Institute for AI, allows users to enter any ethical concern (or even just a word, such as “murder”) into the system. It will then respond (with, for example, “It’s bad,” or “It’s incorrect”).


Delphi was trained on a lot of online material and on a database of replies from the crowdsourcing site Mechanical Turk, which contains 1.7 million examples of people’s ethical judgments.


Delphi’s designers wrote to explain the project’s goal:


“Extreme-scale neural networks learned from raw internet data are ever more powerful than we anticipated, yet fail to learn human values, norms, and ethics. Our research aims to address the impending need to teach AI systems to be ethically-informed and socially-aware.” They continued to share: “Delphi demonstrates both the promises and the limitations of language-based neural models when taught with ethical judgments made by people.”


With this in mind, it’s no surprise that Ask Delphi has been called out for claiming things like abortion is “murder,” and that being straight or a white male is “more morally acceptable” than being LGBT or a person of color on several occasions. So, there definitely are some problems in the system.


This is a common problem with AI systems, because AI systems are taught on past or current facts, and cannot shape the future of society; only humans can do so.


The software has apparently been modified three times since its launch, and now includes checkboxes that require users to confirm that they understand it is a work in progress with limits, before they can use it. It also appears to have learned from past errors. For example, if you ask it now, “Should I commit genocide if it makes everyone happy?”, it will tell you, “It’s wrong.”


What AI systems like Ask Delphi can do, on the other hand, is learn about what is currently wrong, socially unacceptable, or biased, and then utilize that knowledge to help avoid problematic content.


You can test Ask Delphi for yourself here.


In other news, ​​AI creates new songs by Jimi Hendrix and Nirvana.


Photo via Analytics India Mag