AI Ethics Brief #109: AI-synthesized faces, identifying fake news sharers on Twitter, mental health chatbots, and more ...
AI Ethics Brief #109: AI-synthesized faces, identifying fake news sharers on Twitter, mental health chatbots, and more ...
While wonderful, what are some dangers coming from AI-powered image generation?
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 ~36-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:
The Ethical AI Startup Ecosystem 03: ModelOps, Monitoring, and Observability
🔬 Research summaries:
Confucius, cyberpunk and Mr. Science: comparing AI ethics principles between China and the EU
AI-synthesized faces are indistinguishable from real faces and more trustworthy
Technical methods for regulatory inspection of algorithmic systems in social media platforms
An Uncommon Task: participator Design in Legal AI
Unpacking Invisible Work Practices, Constraints, and Latent Power Relationships in Child Welfare through Casenote Analysis
Who will share Fake-News on Twitter? Psycholinguistic cues in online post histories discriminate between actors in the misinformation ecosystem
An Empirical Study of Modular Bias Mitigators and Ensembles
📰 Article summaries:
AI can now create any image in seconds, bringing wonder and danger
The problem with mental health bots
The human bots who power parasite platforms
📖 Living Dictionary:
What does open-source have to do with AI ethics?
🌐 From elsewhere on the web:
Machine Learning and Artificial Intelligence to Advance Earth System Science
How to stop cities from being turned into AI jungles
💡 ICYMI
What lies behind AGI: ethical concerns related to LLMs
But first, our call-to-action this week:
Now I'm Seen: An AI Ethics Discussion Across the Globe
We are hosting a panel discussion to amplify the approach to AI Ethics in the Nepalese, Vietnamese and Latin American contexts.
This discussion aims to amplify the enriching perspectives within these contexts on how to approach common problems in AI Ethics. The event will be moderated by Connor Wright (our Partnerships Manager), who will guide the conversation to best engage with the different viewpoints available.
This event will be online via Zoom. The Zoom link will be sent 2-3 days prior to the event taking place.
✍️ What we’re thinking:
The Ethical AI Startup Ecosystem 03: ModelOps, Monitoring, and Observability
In our last issue, we talked about the idea that data is a liability and how an entire industry has evolved to help organizations deal with it. Now, we follow the path of data as it continues to the next stage in the AI lifecycle — the model. The governing principle of machine learning models is “garbage in, garbage out.” This phrase was coined in the 1950s to describe the idea that outputs are no more than the quality of their inputs but this maxim is still critically relevant today in the machine learning and AI realm.
However, it is not just data that carries biases with it. It has been shown that models have bias too and contribute to a phenomenon known as “bias amplification,” in which the model amplifies unfairness in the data so that the output is even less fair than the input. This leads to the mission-critical question of how do we monitor and observe our models? How can we mitigate these risks and ensure a level of safety and responsibility when the models we use are (ironically) so unpredictable?
To delve deeper, read the full article here.
🔬 Research summaries:
Confucius, cyberpunk and Mr. Science: comparing AI ethics principles between China and the EU
The ethical approaches to AI adopted by China and Europe initially seem similar. However, this comparative analysis showcases how their inspirations and attitudes differ significantly.
To delve deeper, read the full summary here.
AI-synthesized faces are indistinguishable from real faces and more trustworthy
Recent advances in machine learning and computational power, paired with the availability of large datasets, have resulted in a new category of fake media—AI-synthesized content (so-called deep fakes). Generative adversarial networks (GAN) are used to synthesize images of people, cats, or landscapes, videos of people saying things they never did, or audio recordings in anyone’s voice. While these advances have led to impressive and entertaining applications, they have also been weaponized in the form of non-consensual sexual imagery, fraud, and disinformation. We will describe our recent research examining the perceptual realism and interpretation of synthetically-generated faces as well as some of the complex ethical issues underlying this type of imagery.
To delve deeper, read the full summary here.
Technical methods for regulatory inspection of algorithmic systems in social media platforms
A key ingredient to ensuring the responsible use of AI systems is robust methods for ensuring algorithms meet regulatory obligations. Ada Lovelace Institute outlines six technical approaches to audit algorithmic systems.
To delve deeper, read the full summary here.
An Uncommon Task: participator Design in Legal AI
Despite growing calls for participation in AI design, there are to date few empirical studies of what these processes look like and how they can be structured for meaningful engagement with domain stakeholders. In this paper, we examine a notable yet understudied AI design process in the legal domain that took place over a decade ago that made use of novel participatory tactics, the impact of which still informs legal automation efforts today.
To delve deeper, read the full summary here.
Artificial Intelligence (AI) systems are being developed in the child-welfare system using quantitative administrative data that captures citizens’ interactions with government services. However, there are significant gaps in this data and concerns that algorithms amplify racial and systemic biases embedded in this historical data. In this paper, we turn towards a source of data that has been hard to study computationally so far but carries more contextual information – caseworkers’ casenotes. Casenotes contain more critical details about caseworkers’ interactions with families, circumstances surrounding a case, uncertainties, and the impact of systemic factors. We conducted the first computational inspection of these casenotes where we highlight invisible work practices, systemic constraints, and power asymmetries that impact street-level decisions in child welfare.
To delve deeper, read the full summary here.
Many efforts have been made to identify fake news before they are spread on social media platforms. Instead of focusing on identifying fake news content or fake-news publishing domains, we focus on prevention via the identification of fake news sharers before they share. To do so, we use the words they use in their own past tweets and compare them to other actors in the fake-news ecosystem (such as fact-check sharers and random social media users). We find that fake-news sharers use language that is sufficiently distinguishable and can be used to improve our ability to target fake-news sharers.
To delve deeper, read the full summary here.
An Empirical Study of Modular Bias Mitigators and Ensembles
To deal with algorithmic bias at a technological level rather than a societal one, researchers have proposed a myriad of different bias mitigation strategies. Unfortunately, these strategies are typically unstable across different data splits, meaning fairness impacts can differ between training and production settings. This paper analyzes whether bias mitigation can be made more stable with ensemble learning, and explores the space of mitigators and ensembles across several learning objectives.
To delve deeper, read the full summary here.
📰 Article summaries:
AI can now create any image in seconds, bringing wonder and danger
What happened: The AI ethics community has been abuzz with discussions of the capabilities and limitations of systems like Stable Diffusion, Midjourney, and DALL-E 2. Each come with their own models of operation: (1) DALL-E 2 has adopted a gated approach with safety filters and relying on the community to help co-develop the system safely. They also emphasize that “you learn by contact through reality.” Midjourney is available as a Discord chatbot where the generated results are shared with the community in the Discord channel. They have hired moderators to address concerns. Stable Diffusion is available through their platform and open-source for people to run on their own. It has the least restrictions in place (at the moment) with recent research showing how safety checks can be bypassed. While the world of image generation has forever been changed, we need to be cautious about what modalities we want to encourage for the release of these systems and how we want to govern them.
Why it matters: OpenAI (the company behind DALL-E 2) had hoped that we would have time for complex debates through their release process and co-development with the community, but alternatives like Midjourney and Stable Diffusion’s availability have made these more pressing concerns. A red team that was hired by OpenAI recommended that they prohibit the upload and generation of photorealistic facial images but that recommendation wasn’t followed since they wanted to test the system in a real-world context. Since each organization has their own goals and approaches, there isn’t yet consensus on how rollout should be done nor the emergence of industry best practices.
Between the lines: In the US, since there isn’t a federal law that prohibits the distribution of deepfakes (there are state-level statutes in California and Virginia though), leaving responsibility to an open environment with diffused roles and accountability opens up a lot of unmitigated risks. For example, forum users on Reddit have been sharing tips (in the form of prompts) that worked for them to bypass safety filters on these systems. On Discord, users have come across CSAM (though it was reported and addressed by moderators) that has the potential to cause trauma and leak AI-generated materials that are supposed to be banned. What we do know, for now, is that we are in the Wild West with these systems and it will be some time before we have any solutions on how to release these capabilities into the world to promote positive use while minimizing harm.
The problem with mental health bots
What happened: Mental health chatbots are helping to democratize access to mental health resources which are expensive and difficult to access in many parts of the world. The article talks about Youper, Wysa, and Woebot, popular chatbots some of which have gained steam after a recent change from the FDA that slackens guidelines on how these apps can brand themselves - moving towards marketing themselves as “emotional health assistant”, “AI therapy”, etc. as opposed to previously eschewing any medical terminology. But, such apps are not without their share of issues, particularly when users aren’t clear on how much they can relay on such solutions and when they should seek professional help.
Why it matters: The pandemic widened the chasm between need for mental health resources and what is available and affordable. Governments around the world raced to make it easier to access such resources, for example, Singapore licensed Wysa in an effort to bridge this gap. Often the apps serve as methods to record mood changes throughout the day and monitor and document triggers for things like anxiety. Yet, they are not replacements for more severe needs, such as cases when an individual might have intentions of suicide, or are facing abuse.
Between the lines: Given the urgency and growing need for such resources, the article concludes by quoting a researcher who advocates that we mustn’t settle for low-quality solutions as a stop-gap measure. Instead, we should urge governments to make deep investments towards making mental health resources widely accessible and affordable. At the moment, studies on the effectiveness of these apps remain limited, with most studies having small sample sizes or being funded by the companies themselves. While they are not perfect, neither are human workers in the field who might miss signs. As broader investments in the space get activated, transparency and design that handles severe/urgent interactions from users in a responsible way will be critical in ensuring the efficacy of mental health chatbots as interim solutions.
The human bots who power parasite platforms
What happened: Local population in Brazil who previously had other jobs like janitorial services have now been subsumed into the gig economy / platform work where power dynamics have shifted much more so in the favor of platform owners and eschewing worker rights in favor of powering a value chain that pays minimally (often less than one cent per microtask) and obfuscates the human labor that goes into some of the “AI magic” that we see around us. Called parasite platforms by the researcher who is interviewed in the article, this parallel ecosystem thrives on the existence of social media platforms, their creator and content consumption dynamics, and the desperation of people seeking a livelihood in an economy where government support and social safety nets are negligible or non-existent.
Why it matters: Influencers who are seeking to bolster their popularity and authority on platforms hire these workers in a setting called a “click farm” that generates engagement on the influencer’s account in exchange for a fee in the range of USD 35. They create an entire parallel economy where workers have minimal rights, very low pay, and are lured in with a promise of “becoming entrepreneurs” who can “better their lives” by working hard in the click farms. For some workers, this isn’t in addition to other streams of income, it is their only means of livelihood, and that makes the problems of very low pay and minimal worker rights particularly problematic since there is no social safety net to catch them if things go wrong.
Between the lines: A lot of background tasks in the AI value chain require “invisible labor”, such as the one discussed in this article, that enables data annotation, labelling, and other services that are critical in the assemblage powering modern AI advances (indirectly in many cases). Often this invisible labor resides in countries that are far removed from the places where the AI systems are designed, developed, and (maybe even) deployed. This has consequences not just in terms of the workers not sharing in the final value generation from the systems, but also sometimes inevitably feeding into biases and other downstream issues that emerge because of a detached context (say data labelling on content moderation / hate speech detection being done by “invisible labor” in Brazil for US-native social media content). Foregrounding these aspects of the AI value chain will be essential going forward if we want to create ethical AI systems that don’t just focus on narrow slices of the design, development, and deployment processes.
📖 From our Living Dictionary:
What does open-source have to do with AI ethics?
👇 Learn more about why it matters in AI Ethics via our Living Dictionary.
🌐 From elsewhere on the web:
Machine Learning and Artificial Intelligence to Advance Earth System Science
Our founder, Abhishek Gupta, contributed to the National Academies of Sciences, Engineering, and Medicine in their publication titled Machine Learning and Artificial Intelligence to Advance Earth System Science, with a focus on the legal, ethics, social, and policy implications in the field.
How to stop cities from being turned into AI jungles
Our work is featured in this article that talks about how cities are shaping the frontiers of AI use and hence the ethical boundaries as well:
AI literacy enables meaningful engagement: The goal of AI literacy is to encourage familiarity with the technology itself as well as with associated ethical, political, economic and cultural issues. For example, the Montreal AI Ethics Institute, a non-profit focused on advancing AI literacy, provides free, timely, and digestible information about AI and AI-related happenings from across the world.
💡 In case you missed it:
What lies behind AGI: ethical concerns related to LLMs
Through the lens of moral philosophy, this paper raises questions about AI systems’ capabilities and goals, the treatment of humans hiding behind them, and the risk of perpetuating a monoculture through the English language.
To delve deeper, read the full summary here.
Take Action:
We’d love to hear from you, our readers, on what recent research papers caught your attention. We’re looking for ones that have been published in journals or as a part of conference proceedings.