00:00:11.940
Come back everyone! Hello and welcome. Now the first speaker of this year is Kinsey. She's an engineer at DigitalOcean.
00:00:19.720
She is also the co-founder of Kubmo, a non-profit that helps women find their way in tech. Today, she is going to tell us about breaking the chains of oppressive software. Please give her a warm welcome!
00:00:48.420
Okay! Can everyone hear me? Awesome! Great! Hi everyone, good to see all of you. I'm Kinsey, and I am based in Denver, Colorado.
00:00:51.120
I spend a lot of time in Vail, Colorado, which I don't know if any of you have heard of. It's a mountain town about two hours outside of Denver, and it was actually modeled after Bavaria.
00:01:05.260
Here is a picture of the land. I have to say it looks pretty similar! I'm really excited to be here. I spend a lot of time in Vail because I really enjoy fly fishing. I don't know if any of you have ever been fly fishing or heard of it, but I really like it.
00:01:30.440
When I'm not coding or writing, I'm often in the mountains. This is my dog, Harleigh, at a lake near Vail. So, if you're ever in the United States, come say hi in Colorado; it's beautiful! But I have to say, it's a lot prettier here, maybe just because the mountains are a lot taller.
00:01:47.200
I work for a company called DigitalOcean. Who here has heard of DigitalOcean? Wow, that's really awesome! Do you guys use it to deploy your software? If you have, at least you’ve heard of DigitalOcean.
00:02:06.300
Prior to this, I worked at a company called GoSpotCheck, where I was part of a team that used machine learning to build image recognition software. I'll talk a little bit about AI and those kinds of topics, so I wanted you guys to know that I've been in that world before.
00:02:24.700
Now, I'm going to stop talking about myself, and here's what we'll go through for the next 20 to 30 minutes.
00:02:30.069
I’m going to talk about biases. A bias is a cause to feel or show inclination or prejudice for or against someone or something. You may think that I'm going to come up here to talk about unconscious biases in tech companies.
00:02:40.959
It's a great topic, but we've definitely heard about it before. I really think it’s more than that; it’s actually the biases in our code, specifically in our algorithms.
00:03:02.100
Algorithms are becoming increasingly important. They're making significant decisions in our societies. It's not just search engines; they impact everything from online review systems to educational evaluations, the operation of markets, political campaigns, and even how social services like welfare and public safety are managed.
00:03:25.350
They are driven by data and continually influence our daily lives. A lot of times, I’ve heard algorithms referred to as black boxes because we put something in and get something out, making it impossible to see what’s going on inside.
00:03:39.349
Algorithms have done a lot of good things. They make a lot of things easier, like Lyft and Uber, and different services like that. However, it’s crucial for us to be aware that these algorithms can make mistakes and exercise power over us.
00:04:03.230
My intent here isn’t to demonize algorithms, as they are beneficial and have done much for us. Yet, we must recognize that they operate with biases, just like the rest of us.
00:04:20.350
There are two schools of thought surrounding algorithms: one is that they are rational and objective— but are they really? Is the data we input into them objective? Are they built with the idea that they need to be objective?
00:04:35.610
The other school of thought is that algorithmic reasoning is fundamentally flawed. I truly believe that they are here to stay, regardless of what anyone thinks, but biased algorithms can be dangerous.
00:04:52.730
Now, I want to get into how algorithms have reflected biases with a few examples to share with you all. First, I'm sure all of you have used Google Search, typed something into Google, and noticed its results.
00:05:07.890
Here is a quick video that I wanted to share from a campaign by UN Women. It highlights how poor Google search algorithms can be.
00:05:22.980
[Video Clip]
00:06:42.060
Yeah, I was an advertising major in college before I learned how to code, and I found that campaign really cool and powerful. It brought to light how problematic Google search algorithms can be.
00:07:07.940
It wasn't just about women; you can search for various things and encounter very sexist and racist results. Google uses a technology called Word2Vec, which is a pretty cool innovation, but obviously has flaws.
00:07:26.630
Back in 2013, researchers at Google created a neural network with around 3 million words from Google News. The goal was to identify patterns in how words appear together.
00:07:55.170
This complex system discovered that words with similar meanings occupied similar positions in a vector space of over 300 dimensions. This is when biases became apparent.
00:08:41.920
For example: 'man' is to 'computer programmer' as 'woman' is to 'homemaker'. This clearly reveals blatant sexism and racism.
00:09:05.330
Another example involves biases in image recognition. A while ago, Twitter exploded when Google Photos tagged people of color as 'gorillas', which was undeniably offensive.
00:09:20.070
Rather than addressing the algorithm's flaws or the data associated with the words, Google simply removed every image of primates from their database as their solution.
00:09:39.920
This approach received significant backlash from various media. There are many instances of similar biases against people of color, and addressing these issues will only grow increasingly urgent.
00:09:57.700
As technology such as facial recognition software becomes more prevalent in law enforcement, border security, and even hiring, the stakes are high.
00:10:24.950
Another troubling example was with Beauty AI, an initiative heavily backed by Microsoft. This AI was designed to hold a beauty contest judged by robots using three algorithms trained with deep learning.
00:10:55.960
This system didn't evaluate skin color, and despite receiving over 40,000 submissions, the finalists were predominantly white or of light skin tone.
00:11:15.500
The organizers blamed their dataset for the outcome, but the glaring issue lies within the algorithm that favored specific appearances.
00:11:31.120
There's also an example of bias in the LinkedIn search engine that often prompts users with male suggestions when searching for female contacts.
00:12:01.990
If you were to search for 'Andrea Williams', it might ask if you meant 'Andrew Williams', even if you have a connection named Andrea.
00:12:35.580
LinkedIn defended itself by stating that more males use their platform rather than women, but perhaps that's due to how women perceive their interactions in LinkedIn.
00:12:54.170
Additionally, biases exist within courtroom algorithms, which is quite alarming. The COMPAS algorithm is one of these secret systems, designed to predict the likelihood that someone will reoffend and is used during sentencing.
00:13:19.290
The system requires defendants to answer 137 questions, and their risk of committing another crime is assessed through this algorithm.
00:13:42.670
For instance, Eric Loomis was sentenced to six years in prison based on the COMPAS algorithm's assessment of him being at high risk for reoffending. However, the factors the algorithm considered are unknown, as it operates as a black box.
00:14:12.550
He appealed this ruling, claiming it violated due process since the inner workings of the algorithm remained undisclosed. This case reached the Wisconsin Supreme Court, which ruled against him.
00:14:30.150
Although they noted the sentence would have been the same regardless of the algorithm used, the skepticism surrounding COMPAS endures. Should we be utilizing this predictive algorithm when it's known bias may exist?
00:14:53.620
As more courts adopt software like COMPAS for sentencing, recognizing biases indicates that this is a dangerous practice that threatens the integrity of the judicial system.
00:15:10.010
The director of a criminal law reform project has expressed concern that predictive algorithms could further accentuate existing inequalities in our justice system. He also pointed out that the data used in this sphere is often unreliable.
00:15:31.860
Now, let’s turn our attention to recruiting algorithms. Amazon tested an AI recruiting tool that ended up favoring male engineers. They utilized 500 computer models with 50,000 search terms and used scraped resumes from LinkedIn and other sources.
00:16:04.920
As a result of the biased algorithms, Amazon decided to scrap the entire recruitment project instead of fixing the biases within their system.
00:16:24.160
The situation raises concerns about how many companies might be implementing similar biased algorithms silently.
00:16:49.640
Now, one of the problems exacerbating these issues is the Computer Fraud and Abuse Act, which oddly prohibits investigating algorithms for biases. Criminal charges can arise from creating multiple accounts to examine variations in data.
00:17:03.130
As a result, the ACLU and certain researchers and journalists have filed lawsuits against this law. Moreover, the lack of diversity within tech companies contributes to the perpetuation of biases.
00:17:29.300
Researchers often lack representation, and the programming teams' inherent biases are mirrored in the software they create.
00:17:51.780
The datasets we rely on also possess their own biases, indicating that no matter how advanced our algorithms are, they can’t rectify biased data.
00:18:10.840
A question arises: is the binary system itself biased? The binary system represents the strings of ones and zeros foundational to all computer systems. It enables efficient calculations.
00:18:51.470
You might think bias in the tech industry stems from Silicon Valley tech bros, but the issue traces back to ancient philosophers like Aristotle, who posited duality.
00:19:06.350
These binary classifications, as Aristotle defined them (finite/infinite, odd/even, man/woman), continue to dominate AI today. His ranking of dualities is echoed by millions of engineers.
00:19:44.470
Considering the moral implications of the binary classification is essential since it often fails to address ethical dilemmas. We're living in a world of ones and zeros, with little room for nuanced understanding.
00:20:13.740
The German philosopher Leibniz created the binary system to procure faster yes/no verdicts and to condense large numbers. Rather than establish coherent universal systems, it entrenched Aristotle’s duality.
00:20:38.570
Now, natural language processing (NLP) frameworks and complex machine learning systems rely heavily on binary representations of data.
00:20:58.430
This introduces biases that can hamper understanding the rationale behind human decision-making. Existing frameworks should be capable of adapting to remedy these biases, yet engineers often operate strictly within the constraints of legislative and binary logic.
00:21:42.990
Statistics illustrate how crucial the incorporation of AI is. Currently, 33% of enterprises utilize AI, with predictions indicating continuous growth. A staggering 1,400% increase in active startups using AI has been noted since 2000.
00:22:01.660
55% of HR managers state that AI will become routine in five years, which raises alarm bells considering companies like Amazon have faced substantial hurdles with their algorithms.
00:22:24.330
Predictions state that global AI spending could increase to $7.3 billion annually by 2022, highlighting ongoing investments in AI for business differentiation and faster service delivery.
00:22:54.870
However, the chief victims of this technological advancement will be impoverished communities and minorities. Researchers believe this technology may exacerbate societal disparities.
00:23:15.430
As societies continue to evolve with technological advancements, certain groups will undoubtedly suffer. The implications of AI on issues of geopolitical conflict are significant.
00:23:37.800
AI plays a crucial role in many life-altering decisions, such as whom to interview for jobs, who gets parole, and credit approvals. It's embedded in popular technologies, including speech recognition, Google services, Netflix recommendations, and Amazon's algorithms.
00:23:58.590
Imagine a future where algorithms prevent innocent people from boarding flights or escalate wrongful insurance premium hikes—these risks are very real.
00:24:17.200
As we engage in this crucial dialogue, we must recognize that technologists are now shaping political and economic structures through their work.
00:24:40.720
With this understanding, we must ask ourselves how the growing reliance on algorithms might intensify societal prejudice, discrimination, and intolerance of differences.
00:25:00.290
So what can you do about it? One potential solution is designing non-binary categorization systems within AI. However, practicality becomes a concern, as most of us rarely have enough time to tackle such endeavors.
00:25:19.680
We need to emphasize programs that implement algorithms ethically and responsibly, including pondering whether morality can be programmed in.
00:25:34.240
It’s crucial for us to begin discussing and acting on these questions. We can advocate for policies supporting research into AI ethics, emphasizing the importance of guidance when deploying algorithms.
00:25:53.400
We have the opportunity to support from policymakers who also recognize the importance of checking AI and science and can influence legislation that promotes the ethical use of algorithms.
00:26:19.930
We can establish a code of conduct for AI, aligned with the methodologies used in other domains like open-source and conference contributions, ensuring that testing is done within ethical bounds.
00:26:36.360
Fostering diversity within the engineering field responsible for creating algorithms is equally important. The more inclusivity we have, the more awareness can develop against inherent biases.
00:26:55.380
Moreover, making ourselves aware of our personal biases contributes positively. I highly suggest taking the implicit test available online. It surprised me, revealing biases I didn't even realize I had.
00:27:10.720
Incorporating bias reduction processes in our code reviews is another vital avenue. Mozilla now offers a Firefox extension to anonymize pull requests, allowing contributors to focus solely on code quality.
00:27:39.660
We must also demand algorithmic accountability. Leading tech companies like Google, Microsoft, and IBM are beginning to express their commitment to this cause.
00:28:05.720
There are corporations emerging focused solely on auditing algorithms, which is a refreshing development. You can also leverage fairness tools developed by research organizations like Accenture and the Alan Turing Institute to combat biases.
00:28:35.510
Be transparent regarding your datasets and algorithms; don’t let them remain black boxes. Your stakeholders should encourage this type of openness to address concerns.
00:28:54.860
How can we ensure that our algorithms are accountable? First, can we trust our own data? Bring these questions to the forefront in meetings, code reviews, and project conversations.
00:29:12.880
Engaging in ethical discussions around AI and integrating these values into our planning and execution stages promotes conscientiousness in our technology.
00:29:37.040
So to summarize, it’s critical to address the ethical and moral implications of our code. Our responsibility as developers carries weight, and we possess the power to invoke change.
00:29:52.400
Let’s use our power wisely, set a precedent for future generations of developers, and strive for a world that is just and equitable.
00:30:15.980
Thank you for listening, and thanks to Ruby on Ice for hosting such an amazing conference! Here are some articles and papers I referenced, and I can send out the implicit bias test if anyone is interested.