Artificial intelligence can be used to help a company narrow down its pool of thousands of job applicants. AI can be applied to help a doctor make a recommendation for care or perform a procedure. AI might show up in daily life by helping a driver find a faster route home.

But what if the recommendation in the doctor’s office is wrong or the algorithm used to make hiring decisions systematically excludes certain types of candidates? This advanced field of computer science that’s intended to improve lives might end up doing more harm than good in some instances. Not to mention, companies can also suffer from reputational or legal damages if AI is used irresponsibly.

To ensure AI is being used in the most accurate, unbiased and moral manner, it is important for companies to put ethical AI into practice. That’s why companies like Microsoft and IBM have created comprehensive AI ethics guidelines — and smaller tech companies are creating standards around how to use AI responsibly, too.

“The precision needs to be much higher in healthcare than in other situations where we’re just going about our lives, and you’re getting a recommendation on Google Maps for a restaurant you might like,” said Sachin Patel, CEO at Apixio , a healthcare AI platform. “Worst case, you’re like, ‘Oh, I actually don’t want to eat that today,’ and you’re fine. But in this case, we want to make sure that you’re able to very specifically make a recommendation and feel like you’re 90 percent plus on that precision metric.”

Built In spoke with AI and ethics experts about best practices for tech companies to ensure they are exercising strong AI ethics. 

More on AI Innovation The Future of AI: How Artificial Intelligence Will Change the World

First, a company should articulate why they are planning to use AI and how it will benefit people.

“They have to say we want to be making technology that benefits the world, that’s not making the world a worse place, because we all have to live here together.” 

“Is it being used for bad things like autonomous weapons systems versus drug discovery for helping people in medicine?” said Brian Green , director of technology ethics at the Markkula Center for Applied Ethics at Santa Clara University . 

Even in less extreme cases, AI can cause harm to individuals by making people feel more isolated or addicted to their devices.

“There’s so many algorithms and apps out there that use machine learning or other kinds of tactics to try to keep you addicted to them, which kind of violates human freedom in some ways,” Green said. “You’re being manipulated, basically, by these things.”

For example, Flappy Bird was a mobile game where users navigated a digital bird around obstacles. After jumping to one of the top 10 most downloaded apps in the U.S. in 2014, the designer realized that the game was addictive, so he decided to pull the game from app stores.

“That probably made him lose money overall, but at the same time, at least he knew himself that he wasn’t going to be hurting the world by what he was doing because he had a bigger picture,” Green said. “There’s more important things than money.”

Companies should consider how the use of AI will affect the people that use the product or engage with the technology and aim to use AI only in ways that will benefit people’s lives. For example, AI can take a toll on the environment because of the significant amount of energy that machine learning models require for training, but AI can also be used to help solve climate and efficiency issues, Green said.

“I think the first thing to do for a company is the leadership has to fundamentally make a choice,” Green said. “They have to say we want to be making technology that benefits the world, that’s not making the world a worse place, because we all have to live here together.” 

When a company decides to proceed with using AI in its business model, then the next step should be to articulate the organization’s values and rules around how AI will be used.

“Just as long as they have a set of principles, that’s a good start, but then you have to figure out how to operationalize them and actually make them happen in the company,” Green said. “You need to get it into the product ultimately. That means you need to somehow engage the engineers, to engage the product managers. You need to engage people who are in the leadership in that part of the company. Get them on board. They need to become champions of ethics.”

The Markkula Center for Applied Ethics has a toolkit to help engineers and designers think about AI ethics in their work, such as conducting ethical risk sweeping or ethical pre- and post-mortems to repond to and adjust to any ethical failures. At Apixio, the company’s head of data science created an internal AI ethics oath for the whole company, but especially for the data scientists, that outlines best practices around topics like secure data transfer and data privacy, Patel said. 

HireVue , a hiring platform that uses AI for pre-hire assessments and customer engagement, has created an AI explainability statement that it shares publicly. The document outlines for its customers why and how the company uses AI.

“At my time with HireVue, I have seen us move more and more towards just being more transparent because what we’ve seen is that if we don’t tell people what we’re doing, they often assume the worse,” said Lindsey Zuloaga, chief data scientist at HireVue.

Startups using AI often find themselves rapidly testing. While it’s necessary, it can lead to forgetting how algorithms were initially created and why certain decisions were made at a given time, Patel said. Transparency around the creation of algorithms can help with understanding the traceability and reasoning behind decisions.

“It’s black box for the engineers who actually built it as well ... and that makes it even harder to figure out when the biases creeped in and how to fix the model.”

“We’ll train up a bunch of signals. They learn on their own. It’s the nature of machine learning, and then you’re like, ‘Do I know how to trace to make sure [I understand] what it learned on its own?’” Patel said. “Oftentimes, you go back a year later, and you’re like, ‘Oh, I’ve got to actually relearn that now.’”

Sometimes machine learning techniques can become so complex that humans can’t possibly understand them. Black box models in AI are created from data by an algorithm where there’s no explanation to humans as to why the decisions were made. “If we can’t understand the algorithm, that’s a problem. We want to try to protect the people who are being analyzed by the algorithm,” Green said. 

Transparency around algorithms is a way to help reduce potential biases in AI decision making, said Sameer Maskey , adjunct associate professor at Columbia University and CEO of Fusemachines , a machine learning company.

“These days, with deep learning systems with a hundred million parameters, it spits out a decision. It’s a black box for most people,” Maskey said. “It’s black box for the engineers who actually built it as well. A lot of engineers don’t have the needed transparency in figuring out why it made that decision, and that makes it even harder to figure out when the biases creeped in and how to fix the model.”

Bias can creep into algorithms when the data used in AI models is over representative, inaccurate or otherwise skewed by humans. “Bias is a big issue. I would say it’s the elephant in the room, and there’s no easy way to address it,” Maskey said, who further emphasized that lack of transparency in AI models is one of the major culprits. 

One way to potentially decrease biases is to have a checklist for engineers to think through with regard to the data they receive before building a model, he said. Those questions might be how was the data collected? What’s the history behind it? Who was involved in collecting it? What questions were asked? 

“I think one of the main things that a lot of enterprises should do is having very clear guidelines on how to do a data analysis to figure out if there is bias already seeped into the model and providing very clear guidelines on that when the engineers are designing the systems,” Maskey said.

Apixio uses data with wide geographic spread and representing a variety of lifestyle factors, Patel said. The company can then make healthcare recommendations based on what state a provider is in and the type of environment like a rural versus urban setting.

“We feel much more confident now eight years into our journey of training these algorithms, and having seen the 40 million charts, that we’re actually making predictions at a level that feels good. It takes time,” Patel said. 

HireVue has trained evaluators who analyze thousands of data samples for bias to ensure job candidates are assessed consistently and fairly. “Is the training data biased? Does it have groups that are not represented in the data? Does it represent the group of people that you want to apply the algorithm to?” Zuloaga said. 

Data scientists at HireVue will reoptimize algorithms if they find that decisions are not aligning with the way that a client wants to use the AI. “Often, if we do see any problems, it could be we have a customer that’s using an algorithm on a population that’s different from the population we trained on,” Zuloaga said. 

More on Trusting AI Are You Sure You Can Trust That AI?

Maintaining high quality data hygiene ensures accuracy and relevancy, and companies using AI should also make sure people’s personal information is safe and kept private, Patel said. 

HireVue adheres to the European Union’s General Data Protection Regulation, which is one of the toughest privacy laws in the world and regulates how companies must protect the personal data of EU citizens.

“We do business globally. We have to adhere to the strictest standards, so we’re seeing that Europe is really paving the way, and I think states are starting to follow,” Zuloaga said. 

In an ideal world, Maskey said opt-in for users deciding to share their personal data, rather than opt-out, would be the standard, and ideally, people would be able to easily access and research all data that’s collected about them.

“It’s counterproductive for a lot of companies because they are using the same data to make money,” Maskey said. “That’s where I think the government and organizations need to come together to come up with the right framework and write policy that is more balanced, taking user privacy into account, but allowing businesses at the same time to collect data, but with a lot of controls for the users.”

AI ethics evaluations can become part of a company’s regular risk assessment practice. Apixio has a team of four who regularly assess whether or not the company is abiding by its AI ethics oath, Patel said.

“All businesses do some sort of quarterly risk assessments, usually in the IT security realm, but what we’ve added to it a few years ago is actually this AI piece, so it’s more of a risk and ethics meeting,” Patel said. 

At HireVue, the company has conducted third-party audits to evaluate its AI practices, in addition to consulting with an expert advisory board that includes people with diverse backgrounds in law, industrial and organizational psychology and artificial intelligence. “There’s no standard of what an AI audit is at this point. Every audit we did was really different,” Zuloaga said. 

If you’re a minority candidate, you may be concerned that you’re gonna be treated differently, so how do we kind of address all of those concerns?”

HireVue conducted one AI audit with O’Neil Risk Consulting & Algorithmic Auditing, which is led by Cathy O’Neil, a data science consultant and author of Weapons of Math Destruction: How Big Data Increases Inequality And Threatens Democracy.

“She very much took a holistic approach of saying, who are all the different people that interact with this, and how do we represent all those groups?” Zuloaga said. “What are their concerns? Whether they’re legitimately true or not, the concern is real. If you’re a minority candidate, you may be concerned that you’re gonna be treated differently, so how do we kind of address all of those concerns?”

AI audits and consulting with third parties can point out potential risks of how a company’s AI is being used and ways to address these concerns.