MENU

How do we keep bias out of AI?

How do we keep bias out of AI?

Feature articles |
By Peter Clarke



We’ve seen the films where machines take over the world and mortals are obliterated. While they’re entertaining fare, the consensus is that this is, thankfully, a pretty far-fetched scenario and that it’s not going to happen. There is, however, a more realistic issue with which we should be concerned: algorithmic bias.

“Algorithmic bias” is when seemingly harmless programming takes on the prejudices of its creators or the data it is fed. The results are problems such as, for example, warped Google searches, qualified candidates barred from medical school, and a chatbot that posts racist and sexist messages on Twitter.

One of the trickiest things about algorithmic bias is that the engineers doing the programming don’t have to be actively racist, sexist, or ageist for these issues to raise their heads. By its very nature, Artificial Intelligence (AI) is designed to learn by itself, and sometimes it just makes mistakes. Of course, we can make adjustments after the fact, but the preferred solution would be to prevent it from happening in the first place. So how do we keep bias out of AI?

Ironically, one of the most exciting possibilities of AI is a world free of human biases. For example, when it comes to hiring, an algorithm could give men and women equality when applying for the same job or prevent racial prejudice in policing.

Consciously or not, though, the machines we create do reflect how people see the world, and thus can adopt similar stereotypes and world views. As AI becomes increasingly ingrained in our lives, it’s something we do need to be mindful of.

When it comes to AI, there’s the added challenge that bias doesn’t come in just one form; there are multiple types of bias. These include interaction bias, latent bias, selection bias, data-driven bias, and confirmation bias.

Next: Yet more types of bias


“Interaction bias” is where the user biases an algorithm based on the way they interact with it. When machines are taught to learn from those around them, they can’t decide what data to keep or discard, or what is right or wrong. Instead, they simply consume everything they are given — the good, the bad and the ugly — and base their decision-making on it. Tay, the chatbot mentioned above, was an example of this type of bias. It was influenced by a community that taught it to be racist.  

“Latent bias” is where the algorithm incorrectly correlates ideas with things such as race, gender, or sexuality. For example, when searching for an image of a doctor, an AI might present a male doctor before a female one, or vice versa when searching for a nurse. 

“Selection bias” sees the data used to train the algorithm over-represent one population or group, making it work in their favour and at the cost of others. With the example of hiring, if AI is trained to recognise CVs only from men, then female candidates won’t be successful in the application process.

“Data-driven bias” is where the original data used to train the algorithm is already biased. Machines are like children: they don’t question the data they are given, but simply look for patterns within it. If the data are skewed at the outset, the output will reflect this.

The final example, and similar to data-drives bias, is “confirmation bias,” which involves favouring information that verifies pre-existing beliefs. It affects how people gather information and also how we interpret it. For example, if you believe that people born in August are more creative than those born at any other time of year, there’s a tendency to look for data that reinforces this thinking.

When reading these examples of how bias can infiltrate AI, it can seem concerning. But it’s important to take stock and remember that the world is a biased place and so, in some instances, the results we receive from AI aren’t surprising. That doesn’t make it right, and it highlights the need to have a process for testing and validating AI algorithms and systems so that biases are caught early – ideally,  during development and before deployment.

Unlike humans, algorithms can’t lie and so, if the results are biased, there must be a reason for it: the data it has been given. Humans can lie about the reasons for not hiring someone, but AI can’t. With algorithms, we can potentially know when they’re biased and tweak them so that in the future, they overcome this.

AI learns as it goes, and so mistakes will be made. Often it’s not until an algorithm is used in the real-world that any built-in biases are discovered, because they are amplified. Rather than seeing algorithms as a threat, they can present a unique opportunity to address any bias and correct where necessary.

Next: Constantly question


We can build systems to detect biased decision-making and act on them. Compared to humans, AI is particularly well-suited to applying Bayesian methods to determine the probability of a given hypothesis, without all the potential for human bias. It’s complicated, but possible, and when you consider how important AI already is (and this will only increase in the years to come) it’s a responsibility we shouldn’t shirk. 

As AI systems are built and deployed, it’s vital that we understand them so that we can design them to have awareness and avoid potential bias issues in the future. We forget that AI is still very much in its infancy, despite all the rapid advancements. There’s still so much to learn and improvements to be made. This tweaking will go on for some time, but in the meantime, AI will get smarter and we’ll identify increasing numbers of ways to overcome issues such as bias.  

It is vital that the technology industry constantly question how and why machines do what they do. While most AI operates in a black box, with the decision-making process hidden, transparency in AI is key to building trust and dispelling myths.

There’s a lot of research going on to help identify biases, such as that being done by the Fraunhofer Heinrich Hertz Institute. They are looking into identifying the different types of biases, such as those mentioned earlier, but also more “low-level” biases, and issues that can occur during the training and development of AI.

Another consideration is unsupervised training. At the moment, most AI models are generated through supervised training: datasets with labels provided by humans. With unsupervised training, no labels are given and the algorithm has to classify, identify, and cluster the data by itself. While this method is typically many orders of magnitude slower than supervised learning, the approach would limit human involvement and, therefore, eliminate any conscious or unconscious human biases that can work their way into the data.

There are also things that can be done at the grassroots level. Technology companies need to involve a range of people when creating a new product, site, or feature. Diversity will mean that algorithms are fed a wider variety of data that is unintentionally biased. There’s also a greater chance that any bias will be spotted if a number of people are analyzing the output. 

Algorithmic auditing also has a role to play. In 2016, a Carnegie Mellon research team discovered algorithmic bias in online job advertisements. When they replicated people searching for jobs online, Google ads showed men the listings for high-income jobs nearly six times more often than to women. The team concluded that carrying out internal auditing would’ve helped to reduce this type of bias.

Simply put, machine bias is human bias. Bias in AI can develop in multitude ways, but the reality is, it comes from one source: us.

The key to dealing with the issue is dependent on technology companies, engineers, and developers all taking visible steps to safeguard against accidentally creating an algorithm that discriminates. By carrying out algorithmic auditing and maintaining transparency at all times, we can be confident of keeping bias out of our AI algorithms.

Related links and articles:

www.imgtec.com

News articles:

Does AI have a ‘white guy’ bias?

Time is ripe for machine learning, says IMEC

The Internet is made of facts, lies and propaganda

 

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News

Share:

Linked Articles
10s