AI Ethics Brief #70: Gender and AI art, machines as teammates, stasis in AI ethics, and more ...
AI Ethics Brief #70: Gender and AI art, machines as teammates, stasis in AI ethics, and more ...
Should doxxing be illegal?
Welcome to another edition of the Montreal AI Ethics Institute’s weekly AI Ethics Brief that will help you keep up with the fast-changing world of AI Ethics! Every week we summarize the best of AI Ethics research and reporting, along with some commentary. More about us at montrealethics.ai/about.
⏰ This week’s Brief is a ~15-minute read.
Support our work through Substack
💖 To keep our content free for everyone, we ask those who can, to support us: become a paying subscriber for as little as $5/month. If you’d prefer to make a one-time donation, visit our donation page. We use this Wikipedia-style tipping model to support the democratization of AI ethics literacy and to ensure we can continue to serve our community.
*NOTE: When you hit the subscribe button, you may end up on the Substack page where if you’re not already signed into Substack, you’ll be asked to enter your email address again. Please do so and you’ll be directed to the page where you can purchase the subscription.
This week’s overview:
✍️ What we’re thinking:
Jake Elwes: Constructing and Deconstructing Gender with AI-Generated Art
Stasis in AI Ethics
A roadmap to more sustainable AI systems
🔬 Research summaries:
Machines as teammates: A research agenda on AI in team collaboration
📰 Article summaries:
Foundation models risk exacerbating ML’s ethical challenges
Now That Machines Can Learn, Can They Unlearn?
Should Doxing Be Illegal?
Apple says collision in child-abuse hashing system is not a concern
But first, our call-to-action this week:
Preserving the Ecosystem: AI, Data and Algorithms
The Montreal AI Ethics Institute is partnering with AI Policy Labs for a discussion on AI and the environment.
The discussion will span across how AI is being leveraged for a greener future. With the computational power required, such technology has the possibility to harm the environment, while also holding the key to innovation. Discussions surrounding this paradox through an environmental lens will be the mainstay of this meetup.
📅 September 9th (Thursday)
🕛 Noon –1:30PM EST
Errata: We apologize that in last week’s newsletter we mentioned “free tickets” when in fact this is a paid event.
✍️ What we’re thinking:
AI Application Spotlight:
Jake Elwes: Constructing and Deconstructing Gender with AI-Generated Art
“The idea behind latent space is that there’s this continuous space between the classes. You have these multi-dimensional vectors which relate everything it [the artificial intelligence] has learned about, say, a female face as well as everything it has learned about a male face, and there’s this continuous space in between. It doesn’t actually have those gendered binaries anymore – it’s a continuation, and with unsupervised learning it doesn’t even have the gendered labels…”
To delve deeper, read the full article here.
From the Founder’s Desk:
There is a disconnect currently between the theoretical advances in solving AI ethics challenges and their practical implementation. This talks dives into how that’s led to a stasis in AI ethics and what can be done to get out of that stasis with actionable steps.
The Launch Space - A roadmap to more sustainable AI systems
AI has a sizeable carbon footprint, both during training and deployment phases. How do we build AI systems that are greener? The first thing we need to understand is how to account and calculate the carbon impact of all the resources that go into the AI lifecycle. So what is the current state of carbon accounting in AI? How effective has it been? And can we do better? This conversation will answer these questions and dive into what the future of carbon accounting in AI looks like and what role standards can play in this, especially if we want to utilize actionable insights to trigger meaningful behavior change.
🔬 Research summaries:
Machines as teammates: A research agenda on AI in team collaboration
The importance of collaboration in the AI space is not only between humans but also between humanity and AI. Imagining working with an AI teammate may no longer be imaginary in the future, and understanding how this will affect collaboration will be essential. For, understanding this will highlight the importance of the human cog to the human-AI machine.
To delve deeper, read the full summary here.
📰 Article summaries:
Foundation models risk exacerbating ML’s ethical challenges
What happened: A massive report released recently from Stanford AI researchers titled “On the Opportunities and Risks of Foundation Models” has brought forth fervent discussion on the role that large-scale pretrained and other models are going to play in AI applications downstream that rely on them to build out their systems. An example of this is GPT-3 that now powers hundreds of apps processing billions of words every single day. Any bias in it gets amplified hundreds of times over in all its downstream uses. Such models also create risks of centralization of power in the hands of those who have the compute and data infrastructure to build such models.
Why it matters: Our penchant for larger AI systems has many impacts that exacerbate the problems in the domain of Responsible AI including bias and fairness, privacy, inclusion, accountability, and increasingly an environmental impact as well. Careful analysis needs to be performed and more research funded so that we can construct an in-depth understanding of the risks that such systems pose. The opportunities are quite clear in terms of being able to potentially democratize access to advanced AI capabilities and applying such advances to better humanity but as we’ve seen with most AI systems, there is always a cost that can have sinister consequences.
Between the lines: The newly formed Center for Research on Foundation Models at Stanford can become an example of encouraging cross-domain collaboration trying to answer fundamental questions about how we use AI and what the future holds. It would be interesting also to see how they choose to interact with groups like EleutherAI and HuggingFace which are more community-driven and are building such foundational models that will have an impact on our future.
Now That Machines Can Learn, Can They Unlearn?
What happened: The article covers the nascent area of “machine unlearning” which has the goal of effectively erasing personal information that is captured in parts by AI systems when they are trained and later on the user withdraws consent or wants to have their information erased. This is no easy task since millions of dollars might be spent in training up an AI system and asking to remove certain parts of the data from the training set means, at the moment, retraining the entire system and hence spending all that money again. This disincentivizes organizations from meeting these demands, especially when the financial burden is so high.
Why it matters: While there is a “right to be forgotten” in the EU, most of the current legislations focus on data erasure and consent withdrawal for data, but few talk about the need to also erase traces of the snippets of personal information that are incorporated in the learned representations in AI models. This will become a more essential consideration with more significant legislation coming up in the US and EU and will also be more meaningful as AI systems pervade more parts of our lives.
Between the lines: As pointed out in the article, the techniques of machine unlearning are still in the early days where their efficacy is quite limited. It’s on the same journey as differential privacy where the technique is incredibly promising, tooling is being developed around it, and hopefully we will have more widespread utilization of the technique over time. What remains is for the efficacy to be proven along with it being practically viable, as we get more researchers and practitioners focussing on it, we will build up the tooling and related processes that will make this a more common practice.
What happened: Doc-dropping, shortened to doxxing, is the process of releasing private information about an individual such as their address, phone number, identifiers, and other information with the intent of targeting the internet mob to harass the individual through threats and unwanted contact. The article documents the case of an individual in Montana who suffered tremendously at the hands of such an attack and successfully sued the individual that instigated this attack, though she is yet to receive the court-awarded compensation.
Why it matters: As outlined in the article, the law implemented in several states in the US makes doxxing a civil offense in some, criminal in others. In some cases, the law is also geared towards protecting specific kinds of people from doxxing attacks like reproductive healthcare workers, police officers, etc. Each of the approaches come with their own pros and cons, in the case of civil offenses, the burden of proof remains lighter making it perhaps easier to obtain justice but criminal offenses carry a higher punitive burden offering a stronger deterrent.
Between the lines: In the case of the person mentioned in the article, she believes that such laws would have stemmed the hateful outpour against her by making it clear that perpetrators cannot hide behind a screen and keyboard. These virtual attacks have very real consequences for the victims and stronger legislation that offers protections against such behavior to all citizens has the potential to make our interactions in the virtual world much safer.
Apple says collision in child-abuse hashing system is not a concern
What happened: Apple recently unveiled the NeuralHash perceptual hash algorithm that will be applied to iCloud backed content on Apple devices to detect child sexual abuse material (CSAM). This has been met with backlash from privacy-minded organizations and activists who have called it out for setting a precedent that might allow for more invasive monitoring of people’s private content on their devices. As the details of the system have come to light, researchers have reverse engineered the algorithm and have demonstrated hash collisions (when two images that are different produce the same hash - a representation code for that image) that will befuddle the system into giving out false positives. Apple has mentioned that there are secondary checks in place that will minimize the impact from such false positives through the use of an additional server-side algorithm different from NeuralHash and more than 30 images need to be flagged before they are passed on as an alert for human intervention.
Why it matters: Robustness in systems that detect and automatically flag content is important, especially if there is analysis being performed on private content. Yet, it would appear that there are some flaws in the system as demonstrated by researchers. More importantly, a lack of complete transparency on the secondary systems and what the real-world probabilities of these collisions is going to be like further exacerbate the doubts that people have about the effectiveness of such a system.
Between the lines: While the intention behind the deployment of such a system stands to make the information ecosystem safer, especially as it relates to CSAM, without trust from the users who form that ecosystem, there is bound to be pushback and hesitation in full participation. Apple can of course railroad ahead since they own the software and hardware stack but that will be harakiri in a competitive marketplace where they have always prided themselves on keeping the privacy of their users above all else, even in the face of mounting pressure from law enforcement agencies in the past.
From our Living Dictionary:
‘Machine Learning’
👇 Learn more about why it matters in AI Ethics via our Living Dictionary.
In case you missed it:
Algorithmic Bias: On the Implicit Biases of Social Technology
The paper presents a comparative analysis of biases as they arise in humans and machines with an interesting set of examples to boot. Specifically, taking a lens of cognitive biases in humans as a way of better understanding how biases arise in machines and how they might be combatted is essential as AI-enabled systems become more widely deployed. What is particularly interesting about the paper is also how the author takes a simple k-nearest neighbor (kNN) approach to showcase how biases arise in practice in algorithmic systems. Also, tackling the hard problem of proxy variables is done through the use of illustrative examples that eschew the overused example of zip codes as a proxy for race. Taking multiple different iterations on the same running example helps to elucidate how biases can crop up in novel ways even when we have made genuine efforts to remove sensitive and protected attributes and made other attempts to prevent biases from seeping into the dataset. Finally, the paper concludes with a call to action for people to closely examine both human and machine biases in conjunction to create approaches that can more holistically address the issues of harm for people who are disproportionately impacted by these systems.
To delve deeper, read the full summary here.
Take Action:
Preserving the Ecosystem: AI, Data and Algorithms
The Montreal AI Ethics Institute is partnering with AI Policy Labs for a discussion on AI and the environment.
The discussion will span across how AI is being leveraged for a greener future. With the computational power required, such technology has the possibility to harm the environment, while also holding the key to innovation. Discussions surrounding this paradox through an environmental lens will be the mainstay of this meetup.
📅 September 9th (Thursday)
🕛 Noon –1:30PM EST
Errata: We apologize that in last week’s newsletter we mentioned “free tickets” when in fact this is a paid event.