Radical Reads

Ripple Effect – How Inclusive Design Benefits Everyone

By Leah Morris, Senior Director, Velocity Program

http://Screenshot%202024%2007%2002%20at%209.23.54 AM

As AI transforms industries and reshapes our culture, advances in the technology will help AI better represent marginalized communities, including the LGBTQI+ (Lesbian, Gay, Bisexual, Transgender, Queer, and Intersex) community. In the spirit of Pride Month, this week we share considerations for how technology can be developed with inclusivity in mind.

In the US, 7.6% of all adults do not identify as heterosexual (this figure rises to one in five among Gen Z adults). A recent analysis indicated that large language models may exhibit discriminatory behavior more frequently towards LGBTQI+ individuals, showing a 7.2% difference in the QueerBench score of harmfulness (hate speech, discrimination and prejudice). 

Despite these challenges, AI also presents opportunities to address known issues faced by LGBTQI+ individuals. To harness AI’s potential while mitigating its risks, adopting best practices emerging from within industry and the research community can ensure AI benefits all users more equitably. This includes:

  • Diverse data collection: Ensuring datasets are inclusive and representative of the LGBTQI+ community. This involves collaborating with LGBTQI+ organizations to understand their unique challenges and perspectives. For example, this is relevant when AI is applied in tools that generate medical and psychological advice.
  • Data labeling: Accurately labeling data in line with best practices and experiences among the LGBTQI+ community can help close the representation gap. There is a pressing need for participatory research to include a range of diverse perspectives on issues of data collection and AI design. 
  • Addressing harmful bias: Bias can emerge anywhere in the system, with race and gender the least obvious biases to detect (Mozilla Fellow Deborah Raji is a leading researcher on how misconceptions about race and gender biases can be reflected in algorithms). Studying the origins of biases can significantly enhance our capacity to develop technology that ensures equity for a broader range of users.
  • Participatory development and feedback: Participatory design can play a larger role in empowering marginalized communities and those at their intersections to take an active role in constructing research agendas and outputs. Involving the LGBTQI+ community in the development and deployment of AI systems ensures that different experiences and identities are considered.

Learn More: Queer in AI works to create awareness of queer issues in AI/ML, foster a community of queer researchers, and celebrate the work of queer scientists. Visit Queer in AI to view their resources.

AI News This Week

  • Training AI music models is about to get very expensive   (MIT Technology Review)

    AI music startups Suno and Udio are facing litigation for using copyrighted music in their training data without permission. Major labels, including Sony Music, Warner Music Group, and Universal Music Group, allege that the companies trained their AI models on copyrighted music allowing the AI to generate songs that imitate human recordings. The industry may be moving towards a licensing model, with YouTube reportedly negotiating deals with these labels to use their music legally for AI training. These lawsuits highlight the challenges of training AI music models due to concentrated music rights ownership and the technical difficulty of generating high-quality music.

  • Ray Kurzweil on how AI will transform the physical world  (The Economist)

    Computer scientist Ray Kurzweil predicts that by the time today’s newborns start kindergarten, AI will surpass human capabilities in all cognitive areas. His unchanged forecast from 1999, once viewed as overly optimistic, now appears conservative due to rapid advancements in AI. These developments promise to transform not just digital realms but physical ones as well – making energy abundant, reducing manufacturing costs, and revolutionizing medicine. AI’s role in accelerating research in photovoltaics, battery technology, and molecular biosimulation is key to this transformation. Ultimately, Kurzweil suggests AI will extend life expectancy, where aging no longer increases mortality, signalling a future of longer, healthier lives free from traditional human limitations.

  • These companies are leveraging AI to improve Canada’s stressed healthcare system  (Globe and Mail)

    In Canada, entrepreneurs are leveraging AI to tackle healthcare system challenges. One example is Radical Ventures’ portfolio company, Signal 1. In collaboration with Unity Health Toronto, Signal 1 developed CHARTwatch, a machine-learning tool that anticipates patient deterioration to enhance hospital outcomes. Additionally, they designed a system to pinpoint patients ready for discharge, boosting the efficiency of transitions. Innovations such as those from Signal 1 are crucial for advancing healthcare delivery and operational efficiency.

  • How AI revolutionized protein science, but didn’t end it  (Quanta Magazine)

    Three years ago, Google’s AlphaFold2 revolutionized protein science by accurately predicting 3D protein structures from amino acid sequences, solving the protein folding problem. This breakthrough accelerated research and drug development by providing experimental validation for complex structures. Scientists have used AlphaFold2 creatively, integrating it with techniques like cryo-EM and leveraging its predictions to design new proteins. The application of deep learning is expected to solve problems throughout the field of structural biology including RNA and biomolecular complexes. 

  • Research: LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs  (University of Waterloo)

    Researchers have introduced LongRAG, a novel framework for Retrieval-Augmented Generation (RAG) that balances the load between the retriever and reader components in AI systems. Traditional systems search through many small pieces of information, making the process slow and less effective. LongRAG uses larger chunks of information, reducing the number of pieces the system needs to search through by 30 times. This change makes the system much faster and better at finding the right answers. LongRAG has shown impressive results, correctly finding answers 71% on the NQ dataset and 72% on HotpotQA. By using long-context language models, LongRAG achieves remarkable performance improvements in zero-shot answer extraction, offering insights into future RAG system designs.

Radical Reads is edited by Leah Morris (Senior Director, Velocity Program, Radical Ventures).