Understanding algorithmic bias
Rebecca Berry
Helping leaders transform their organisations through inclusive leadership | Inclusion Expert | Equity & Diversity Writer
You can watch a video of the content of this article here .
Regular readers (hey everyone! ??) will know that I regularly cite the LinkedIn algorithm as a suppresser of social justice opinions and ideas.?
Then I read a post about using algorithms to make decisions more effectively, based on 'Algorithms to Live By' Brian Christian and Tom Griffiths. The premise is that when you need to make a decision, explore 37% of your options then be prepared to commit to the first option that's better than the first 37%.?
Great for making significant personal purchasing decisions such as buying a car or a house. Not so great for decisions involving people, such as deciding who to hire. Interviewing 37% of your shortlist, then hiring the next person who performs better than the first 37% isn't an inclusive approach.
We know that hiring managers are more likely to hire people who look like them than to hire for diverse perspectives. Interviewing fewer people reduces the opportunity to do that even further.?
When algorithms are applied to people decisions, they return the ‘it’s always been like this’ result. So nothing changes. We need to challenge people algorithms, not live by them. And to be fair, the authors do say that algorithms have their limits.
I found myself thinking about that post a lot. I DO grumble about algorithms and AI a lot, but my knowledge of them is superficial. So I decided to dig into it further, challenging my confirmation bias (because I knew I'd be looking for evidence that algorithms and AI perpetuate all the isms), and share my learning and insights.?
I started by learning more about algorithms and how they work. What exactly IS an algorithm, and how is it related to AI?
Algorithms are the building blocks of AI.?An algorithm is a series of instructions telling a computer how to transform a set of facts into useful information. That’s it. Turning data into useful information.?But useful to whom?
Businesses, of course! Businesses want that data to track trends and take advantage of them to make more money. That could be by using our browsing history to show us tailored adverts, or remembering us when we return to a website to show us information based on our previous browsing.?
But algorithms on their own don’t make decisions. They just produce information. That’s where AI comes in.?
AI simulates human intelligence and decision-making processes. By gobbling up VAST quantities of data from algorithms, AI learns to analyse it for correlations and patterns, then make decisions about what happens next.
An example I saw online: say you check the price of a holiday flight to get an idea of prices so you can budget for your fortnight of fun in the sun. You go back a few days later to book the flight, and the price has shot up. Maybe the AI recognises that you’ve been there before and deduces that you’re now committed and ready to buy. Or maybe lots of other people are booking the same flights, increasing their value.
The algorithm reports your return to the flight website, or the increase in flight bookings. Artificial intelligence picks this up and hikes prices accordingly, again using an algorithm to decide what the increase should be.
But it’s not just about making money. Algorithms are used to help with medical diagnosis, for example, and AI can learn to detect certain medical conditions with more accuracy than humans. Although that didn't work with a dermatology AI system. It learned that all cancerous lesions had a ruler next to them, so the AI primarily looked for the presence of a ruler for malignancy. You can guess the rest.
Algorithms on social media track our interests so that AI can show us content it thinks we’ll like. They report our track record in content engagement so AI can analyse it and decide who to show it to. And algorithms on social media detect content that AI doesn’t want us to see for whatever reason. More about this in my next two posts.
So, algorithms can exist without AI, but AI can’t make decisions without data from algorithms. So why do we talk about algorithmic bias rather than AI bias? If AI makes decisions to show or suppress information based on the data it gets from algorithms, why do we blame algorithms for the bias that can result? Why don’t we blame AI?
It’s because AI can only learn and make decisions using the exact data it gets from the algorithms. It doesn’t think hmm, we need a bit more interpretation and nuance here (at least, not yet). It doesn’t ask itself what perspectives might be missing.
So the algorithm is the source of bias. But if the algorithm is only *collecting* data, how on earth can it be biased??
Because people write the code that generates the algorithms.
For example, imagine I’m coding an algorithm (just a sec, that’s quite a lot for me to imagine… OK, I’m ready). My algorithm drives an AI system that decides what kind of care people receive for Parkinson’s Disease. I know lots about it because I care for my mum, who has PD, but I’m a white woman caring for a white woman. I can research the experiences of others who have PD in an attempt to debias my algorithm, but there will always be unknown unknowns. Things that would only come up from the direct lived experience of people with PD. So my algorithm may exclude crucial information, meaning that the AI system won’t know about it and take it into account in its decision making. So someone of a different gender or the Global Majority may not get the right decisions made about their care.
PWC’s report ‘Algorithmic Bias and Trust in AI’ says:
“People write the algorithms, people choose the data used by algorithms and people decide how to apply the results of the algorithms. Without diverse teams and rigorous testing, it can be too easy for people to let subtle, unconscious biases enter, which AI then automates and perpetuates. That’s why it’s so critical for the data scientists and business leads who develop and instruct AI models to test their programmes to identify problems and potential bias.”
How does this show up?
A study found that facial recognition AI misidentifies people from the Global Majority more often than white people. If used by the police, AI facial recognition increases the risk of them unjustly apprehending people from the Global Majority. Wrongful arrests due to mistaken AI facial recognition have already happened.
In USA financial services, several mortgage algorithms have systematically charged Black and Latine borrowers higher interest rates.
And of course the famous example of Amazon scrapping its recruitment AI because it learned that women didn’t get the senior roles, so it automatically rejected female candidates.
So businesses have a responsibility to ensure their algorithms are ethical and non-discriminatory. It’s reportedly one of the top concerns for business leaders. Let’s see how that works out…
Let’s talk about algorithms and social media.
Until I started investigating how to get the best reach on LinkedIn, I naively thought that everything I posted would be shown to all my connections and followers.
Nope!
Social media platforms share the general themes of their algorithms (it’s in their interests for people to understand how to build their reach) but not the algorithms themselves.
领英推荐
On LinkedIn, the algorithm likes posts that:
?? Are easy to read; short paragraphs, simple language, clearly written
?? Include 3 - 5 hashtags, which helps the algorithm to categorise who to show it to
?? Keep people on LinkedIn; external links in your main post will reduce its reach, although links in the comments seem to escape the algorithm
?? Include key words relevant to your niche. For example, I write about equity and inclusion, and my posts are littered with key words. Demonstrating your expertise makes it more likely that your post will be shown to more people.
Once you’ve posted, the algorithm will initially show the post to people you interact with the most, and those who are in your immediate network (e.g. colleagues from the same company). It will then keep an eye on engagement. If your post quickly attracts likes, comments and shares, it will be shown to more people. And there’s an engagement hierarchy; comments are most valuable, followed by reactions, followed by shares.
If there’s little response, or (yikes!) if your post is reported as spam or people hide future posts from you, LinkedIn stops it right there.
So how can you maximise the chances of speedy engagement?
The first/easiest thing to do is to post when people are at work. According to Influencer Marketing Hub the best time to post content is:
?? Wednesday from 8–10 AM
?? Thursday at 9 AM and 1–2 PM
?? Friday at 9 AM
Secondly, build thought leadership in your area of expertise. Post four or five times a week. The most dedicated LinkedIn influencers post twice a day or more.
As comments are what the LinkedIn algorithm is looking for, you need to post content that will draw comments. Doesn’t matter if they’re good or bad. I unexpectedly went a little bit viral a few weeks ago posting about white supremacy, and most of the comments were telling me what a disgusting human being I am. Meh.?But the algorithm kept the post going. Last time I looked, it had over 436,000 impressions and more than 3,100 comments.
Finally, LinkedIn is looking for engagement probability, so build up a body of consistent content that people engage with, and LinkedIn will show your posts to more people. If you’re just starting out, don’t be too proud to ask people around you for likes and comments. Liking and commenting on your own posts counts, although beware commenting too much, as the algorithm may classify it as spam.
What we’re NOT told about the LinkedIn algorithm is that it appears to suppress social justice content. In my own experience, having spent a while analysing the engagement on my posts, with one notable exception I know that anything I post about anti-racism or pro-trans tends to be shown to far fewer people.?
And I’ve been shadow banned, for no reason that I could see. I only knew I’d been shadow banned because?when I looked at my LinkedIn notifications nothing came up. My posts were shown to nobody for a few days. Then everything went back to normal.
What's shadow banning??
It’s the practice of blocking someone’s content because they haven’t complied with the rules of the platform. It happens to lots of people for various reasons, so it’s not just social justice content that attracts a shadow ban.
But talk to anyone who posts regularly about about social justice and they’ll have experienced it.
One example is Saira Rao, who promoted her book, ‘White Women: Everything You Already Know About Your Own Racism and How to Do Better’ on LinkedIn. And Saira found herself instantly banned.?She eventually got to talk to one of the senior folk at LinkedIn, who said it had been a mistake and had Saira’s account reinstated. But he was vague about the reason for the ban.
Shadow bans aren’t decided by people. They’re decided by AI, acting on algorithms created by people. The power to suppress some views and promote others is a worrying AI capability. Someone should be keeping them ethical and accountable, shouldn't they? And if it’s widely acknowledged across industry and governments that algorithms drive biased AI, something needs to be done, right??
Regulation of AI is the subject of international debate right now. Sam Altman, OpenAI’s CEO, recently testified before the USA senate about the future of AI. OpenAI is a research laboratory that states its aim as developing ‘friendly’ AI. Its key partner is Microsoft, which has invested over $10 billion.?
AI regulation polarises opinion, some seeing it as impossible, and too early. AI is still in its infancy and we have no idea what it will look like even five years from now. Others see it as essential to avoid global chaos as AI secretly meddles in elections and decides how and what information is presented to the world. The question is, who should regulate it? Should it be a global effort, like GDPR? Should countries regulate their own AI in collaboration with each other? All of this knowing that big tech can run rings around those who don't deeply understand AI and how it works. And being the only ones who understand it, they stand to hold incredible power as the gatekeepers of AI regulations.
The impact of of biased algorithms is much bigger than I imagined at the beginning of my week of research. From what I can see, Pandora’s box is well and truly open. I await developments with interest!
Here's a list of the resources I used to write this article: