Procedural justice can address generative AI’s trust/legitimacy problem

The much-touted arrival of generative AI has reignited a familiar debate about trust and safety: Can tech executives be trusted to keep society’s greatest interests at heart?

Because its training data is created by humans, AI is inherently prone to bias and therefore subject to our own imperfect, emotionally-driven ways of seeing the world. We know too well the risks, from reinforcing discrimination and racial inequities to promoting polarization.

OpenAI CEO Sam Altman has requested our “patience and good faith” as they work to “get it right.”

For decades, we’ve patiently placed our faith with tech execs at our peril: They created it, so we believed them when they said they could fix it. Trust in tech businesses continues to plummet, and according to the 2023 Edelman Trust Barometer, globally 65% worry tech will make it impossible to know if what people are seeing or hearing is real.

It is time for Silicon Valley to embrace a different approach to earning our trust — one that has been proven effective in the nation’s legal system.

A procedural justice approach to trust and legitimacy

Grounded in social psychology, procedural justice is based on research showing that people believe institutions and actors are more trustworthy and legitimate when they are listened to and experience neutral, unbiased and transparent decision-making.

Four key components of procedural justice are:

• Neutrality: Decisions are unbiased and guided by transparent reasoning.
• Respect: All are treated with respect and dignity.
• Voice: Everyone has a chance to tell their side of the story.
• Trustworthiness: Decision-makers convey trustworthy motives about those impacted by their decisions.

Using this framework, police have improved trust and cooperation in their communities and some social media businesses are starting to use these ideas to shape governance and moderation approaches.

Here are a few ideas for how AI businesses can adapt this framework to build trust and legitimacy.

Build the right team to address the right questions

As UCLA Professor Safiya Noble argues, the questions surrounding algorithmic bias can’t be solved by engineers alone, because they are systemic social issues that require humanistic perspectives — outside of any one company — to ensure societal conversation, consensus and ultimately regulation — both self and governmental.

In “System Error: Where Big Tech Went Wrong and How We Can Reboot,” three Stanford professors critically discuss the shortcomings of computer science training and engineering culture for its obsession with optimization, often pushing aside values core to a democratic society.

In a blog post, Open AI says it values societal input: “Because the upside of AGI is so great, we do not believe it is possible or desirable for society to stop its development forever; instead, society and the developers of AGI have to figure out how to get it right.”

However, the company’s hiring page and founder Sam Altman’s tweets show the company is hiring droves of machine learning engineers and computer scientists because “ChatGPT has an ambitious roadmap and is bottlenecked by engineering.”

Are these computer scientists and engineers equipped to make decisions that, as OpenAI has said, “will require much more caution than society usually applies to new technologies”?

Tech businesses should hire multi-disciplinary teams that include social scientists who understand the human and societal impacts of technology. With a variety of perspectives regarding how to train AI applications and implement safety parameters, businesses can articulate transparent reasoning for their decisions. This can, in turn, boost the public’s perception of the technology as neutral and trustworthy.

Include outsider perspectives

Another element of procedural justice is giving people an opportunity to take part in a decision-making process. In a recent blog post about how OpenAI is addressing bias, the company said it seeks “external input on our technology” pointing to a recent red teaming exercise, a process of assessing risk through an adversarial approach.

While red teaming is an important process to evaluate risk, it must include outside input. In OpenAI’s red teaming exercise, 82 out of 103 participants were employees. Of the remaining 23 participants, the majority were computer science scholars from predominantly Western universities. To get diverse viewpoints, businesses need to look beyond their own employees, disciplines and geography.

They can also enable more direct feedback into AI products by providing users greater controls over how the AI performs. They might also consider providing opportunities for public comment on new policy or product changes.

Ensure transparency

Companies should ensure all rules and related safety processes are transparent and convey trustworthy motives about how decisions were made. For example, it is important to provide the public with information about how the applications are trained, where data is pulled from, what role humans have in the training process and what safety layers exist to minimize misuse.

Allowing researchers to audit and understand AI models is key to building trust.

Altman got it right in a recent ABC News interview when he said, “Society, I think, has a limited amount of time to figure out how to react to that, how to regulate that, how to handle it.”

Through a procedural justice approach, rather than the opacity and blind-faith of approach of technology predecessors, businesses building AI platforms can engage society in the process and earn — not demand — trust and legitimacy.