We had the incredible opportunity to interview Professor Ameet Soni, an associate professor at Swarthmore College and the newly appointed Associate Dean of Faculty, shared his profound insights into the rapidly evolving field of Artificial Intelligence (AI), particularly its application in biomedical technology and the ethical considerations it entails. With his extensive background in Machine Learning, Biomedical Applications of AI, and Ethics of Technology, Professor Soni offers a unique perspective on the challenges and opportunities presented by AI in the modern world.

The Risks and Regulations of AI in Biomedicine:
A significant focus of our discussion was on the risks associated with AI technologies, especially in sensitive areas like biomedicine. Professor Soni emphasized the importance of existing regulatory frameworks, such as those provided by the FDA, in overseeing medical technologies. He advocated for working within these frameworks while also enhancing them to address the unique challenges posed by AI, such as the issue of "black box" algorithms that are not easily explainable.

Balancing Innovation and Safety:
A key challenge in regulating AI, as Professor Soni pointed out, is balancing the need for innovation with the imperative of safety. He suggested a dynamic approach where regulation does not merely precede the deployment of technology but continues as an ongoing process of monitoring and adaptation. This approach ensures that AI systems remain safe and effective even as they evolve and are used in real-world scenarios.

Ethics: At the Heart of AI Development
One of the most notable aspects of our conversation was Professor Soni's emphasis on the central role of ethics in AI development. He stressed the need for responsible AI, where ethical considerations are embedded in the development process from the outset. In his machine learning course at Swarthmore, he encourages students to consider all possible stakeholders and maintain engagement with them throughout the lifecycle of a technological product, not just during the development phase.

Embedding Ethics in Computer Science Education
Professor Soni highlighted his efforts, along with those of other institutions, to integrate ethical discussions into the computer science curriculum. This approach, known as 'embedded ethics', aims to ensure that students encounter these critical questions throughout their education, rather than treating ethics as a separate or one-off topic. He shared examples from his courses where students are encouraged to apply theoretical knowledge to real-world problems, considering the ethical implications of their technological choices.

Looking Ahead: The Future of AI and Ethics in Education
In conclusion, Professor Soni expressed optimism about the progress being made in embedding ethics in computer science education. He noted the initiatives at leading institutions like Harvard, Stanford, and the University of Toronto in developing models for integrated ethics education. These efforts, he believes, are crucial for preparing a new generation of technologists who are not only skilled in AI but are also deeply aware of its societal impacts.

Conclusion:
Our conversation with Professor Ameet Soni provided invaluable insights into the complex world of AI, especially in the realm of biomedical technology. His emphasis on the importance of ethical considerations and the need for ongoing regulation and stakeholder engagement offers a framework for developing AI technologies that are not only innovative but also responsible and beneficial to society. 

Thank you so much for your time and thoughtful conversation Professor Soni. 

Recently, we had the opportunity to interview Ricardo Baeza-Yates, the Director of Research at Northeastern’s Institute for Experimental AI. His journey, beginning with a foundational Ph.D. in computer science from the University of Waterloo, has seen a significant evolution, leading him to his current focus on the pressing issues of AI ethics and responsible AI. As he shared, this transition was driven by his desire to delve into areas with more profound societal impacts, moving away from his initial focus on algorithms and web search.

During our conversation, Baeza-Yates provided an in-depth look into the data mining process crucial for AI training, highlighting the complexities and challenges involved. He detailed the intricate process of selecting appropriate web sources for data curation, emphasizing the difficulty in cleaning data to remove biases and hate speech. A particularly striking example he mentioned was the challenge of detecting hidden biases, which are often not overtly apparent in the data. This aspect underscores the nuanced nature of data curation in AI, where even with meticulous efforts, some biases can remain concealed.

Baeza-Yates also brought to light the privacy risks associated with AI, especially in the context of large language models like ChatGPT that rely on vast amounts of web-sourced data. He pointed out the risk of inadvertently including hate speech or sensitive information in these models, despite rigorous efforts to filter such content. This discussion highlighted the ongoing struggle in balancing the need for comprehensive data with the imperative of maintaining privacy and ethical standards.

The conversation then shifted to the regulatory challenges in AI and data mining. Baeza-Yates provided insightful examples to underscore these challenges. He suggested that AI models should respect the terms of usage of websites from which they mine data, but he acknowledged the practical challenges in enforcing such compliance. Additionally, he discussed the idea of watermarking AI-generated content to distinguish it from human-generated content. However, he expressed skepticism about its effectiveness, particularly in preventing intentional misuse.

At the Northeastern Institute of Experimental AI, Baeza-Yates and his team are actively engaged in pioneering projects focused on responsible AI. He shared their collaboration with companies to establish AI ethics principles and governance structures, mentioning their work on bias audits in health insurance as a key example of their efforts to promote ethical AI practices. This work is crucial in guiding AI development towards more responsible and ethical practices.

Looking towards the future of AI, Baeza-Yates expressed both enthusiasm and caution. He anticipates the development of more sophisticated language models capable of understanding semantics and verifying facts against knowledge bases. However, he also warned of the potential for AI to create convincing but fake news, images, and videos, posing significant challenges to our trust in digital media. This dual perspective underscores the complex landscape of AI development, where advancements bring both exciting opportunities and formidable challenges.

In conclusion, Ricardo Baeza-Yates’ insights provided a comprehensive overview of the ethical challenges and considerations in AI development and usage. His focus on responsible AI, privacy concerns, and the need for effective regulation offers crucial guidance for the future of AI technology. This conversation on "Biased Brains" was not only enlightening but also a beacon for those navigating the complex terrain of AI ethics and data privacy. As we continue to explore the evolving world of AI, these discussions will remain pivotal in shaping a future where AI benefits society while minimizing potential harms.

Thank you Mr. Baeza-Yates for your time and insight! 

ELIZA, developed in the 1960s, represents a foundational moment in the history of artificial intelligence, particularly in the field of natural language processing. Its ability to simulate conversation, especially with its limited technological resources, is a testament to early AI innovation. This blog post delves into the specific mechanics of how ELIZA functioned and the principles behind its conversational capabilities.

The Core Mechanism of ELIZA: 
ELIZA, developed in the 1960s, functioned primarily through a clever combination of pattern matching and substitution methodologies. At its heart was a script, the most famous being the DOCTOR script, which simulated a Rogerian psychotherapist. ELIZA began by scanning the user's input for specific keywords or phrases that were predefined in its programming. Once a keyword was identified, ELIZA employed a set of decomposition rules to break down the user's statement into manageable segments. These segments were then reassembled using reassembly rules to form a coherent response. This process often involved mirroring or paraphrasing the user's input, effectively reflecting the statement back to the user in a new form, typically as a question or a prompt for further discussion. The genius of ELIZA lay in its ability to maintain a flow in conversation using these scripted responses, despite having a minimal understanding of the content. It did not possess real comprehension or contextual awareness; its responses were solely based on the mechanical application of its programmed rules. This simple yet effective mechanism allowed ELIZA to create an illusion of empathy and understanding, engaging users in what seemed like a meaningful conversation, but was, in reality, a sophisticated pattern of linguistic mirroring.

The Illusion of Intelligence and Empathy
ELIZA's effectiveness in creating the illusion of understanding and empathy was surprising, especially considering its simple operating principle. It was this ability to engage users in a dialogue, making them feel heard and understood, that marked ELIZA as a significant milestone in AI development. However, it is essential to remember that ELIZA's conversations, though appearing empathetic, were limited to its programmed scripts and lacked true emotional understanding.

Conclusion
ELIZA's legacy in AI is not just about its technological ingenuity but also about the broader implications and ethical considerations it introduced in AI-human interactions. By exploring the inner workings of ELIZA, we gain valuable insights into the challenges and potentials of conversational AI. ELIZA demonstrated the possibilities of AI in mimicking human interaction, albeit within the confines of its programmed capabilities. As we continue to develop more sophisticated AI systems, ELIZA serves as a reminder of both the achievements and limitations of these technologies in emulating human conversation and emotions.

Source:​
Weizenbaum, J. (1966). "ELIZA - A Computer Program For the Study of Natural Language
Communication Between Man And Machine." Communications of the ACM.

AlphaFold, developed by DeepMind, is an innovative artificial intelligence program designed to predict the three-dimensional structures of proteins from their amino acid sequences with remarkable accuracy. The process begins with the input of a protein sequence, a chain of amino acids, essential for determining protein function. AlphaFold uses Multiple Sequence Alignments (MSAs) to analyze evolutionary relationships between the target protein and its homologs across different organisms, providing insights into structural constraints based on evolutionary conservation. At the heart of AlphaFold is an advanced neural network with an attention mechanism, which processes the MSA and infers spatial relationships between the amino acids. This neural network predicts distances and angles between amino acid pairs, crucial for mapping their positions in three-dimensional space. Using these predictions, AlphaFold constructs a detailed three-dimensional model of the protein, iteratively refining it for accuracy. It also generates confidence scores for each part of the structure, indicating the reliability of its predictions. The accuracy and utility of AlphaFold have been extensively validated, including through comparisons with experimentally determined structures and its outstanding performance in the Critical Assessment of Structure Prediction (CASP) competitions. Notably, its success in CASP14 marked a significant milestone in protein structure prediction. AlphaFold's ability to accurately predict protein structures opens up new possibilities in biological research and medicine, offering deep insights into biological processes and aiding in the development of novel therapeutic strategies.

DeepMind's AlphaFold has catalyzed a new era in the field of computational biology with its groundbreaking capability to predict protein structures, offering transformative implications across various sectors. In biomedical research, AlphaFold dramatically accelerates the understanding of biological processes at the molecular level, pivotal for developing innovative treatments for diseases. Its ability to efficiently predict protein structures, which traditionally requires time-intensive experimental methods, speeds up scientific discoveries significantly. This rapid prediction is particularly crucial for understanding diseases like cancer and neurodegenerative disorders, as it aids in unraveling the mechanisms of protein malfunction, paving the way for new treatment strategies. In the realm of enzyme design, AlphaFold opens new possibilities in creating enzymes with specific properties for use in industries such as pharmaceuticals, biofuels, and food processing.

The perspective piece, "AlphaFold – A Personal Perspective on the Impact of Machine Learning," written by Alan R. Fersht, a seasoned expert in protein science, provides a compelling insight into the profound significance of AlphaFold in the field. As a distinguished protein scientist, Fersht's narrative begins in the year 1968, a time when computational biology and AI were still in their infancy. He takes readers on a reflective journey through his career, marked by significant milestones like the rise of X-ray protein crystallography, DNA sequencing, and the integration of computational methods in protein analysis. What makes Fersht's perspective particularly engaging is his ability to draw parallels between his passion for board games like chess and the realm of AI. He delves into the historical challenges faced by AI in mastering complex strategy games and how technology gradually outpaced human expertise, all while adding a personal touch through his reference to Demis Hassabis, a chess prodigy turned AI expert.

Fersht's narrative seamlessly transitions into the heart of his article, where he explores the Protein Folding Problem, encompassing the prediction of three-dimensional protein structures and the unraveling of folding pathways. He underscores the remarkable achievements of AlphaFold, emphasizing its capability to discern patterns in primary sequences, much like how chess engines analyze positions, and to construct precise protein structures. Fersht's appreciation for the power of Machine Learning in this field is evident, and he envisions its potential to revolutionize drug design and structural biology.

The article culminates in Fersht's anticipation of a future where AlphaFold could catalyze enzyme design, automate drug discovery, and even venture into designing entirely novel protein folds. His respect and admiration for experimentalists and theoreticians shine through as he eagerly anticipates the synergy between human ingenuity and AI capabilities, much like how chess players integrate AI insights into their strategies.
​
In conclusion, Alan R. Fersht's perspective is a compelling narrative that not only unveils the profound impact of Machine Learning in the world of computational biology but also offers a broader view of AI's evolution, the potential for human-machine collaboration, and the exciting frontiers awaiting exploration. His reflections serve as a testament to the ever-expanding horizons of scientific discovery, with technology serving as a guiding force propelling us toward new realms of knowledge and innovation.

Sources: 
Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2
AlphaFold - A personal perspective on the impact of machine learning
J. Mol. Biol. (2021), Article 167088, 10.1016/j.jmb.2021.167088

The Emergence of AI: 
Alan Turing, a pioneer in the field of computer science, laid the groundwork for what we now call artificial intelligence (AI). In his groundbreaking paper, "Computing Machinery and Intelligence," Turing proposed the idea of machines that could think and reason like humans. This concept has evolved into the modern field of AI, which encompasses everything from simple automated responses to complex machine learning algorithms.

The Turing Test: 
The Turing Test is a method for determining whether a machine can exhibit intelligent behavior that is indistinguishable from that of a human. The original test involves a human evaluator who communicates with an unseen interlocutor, which could be either a human or a machine (a computer program). The communication is typically text-based, to prevent the evaluator from determining the interlocutor's nature through their appearance or voice.

How the Test Works:

1. The evaluator (or interrogator) interacts with both a human and a machine through a computer interface. The evaluator is aware that one of the two entities they are communicating with is a machine, but they do not know which one.
   
2. The evaluator engages in a natural language conversation with both parties. They can ask any question or bring up any topic in an attempt to determine which participant is human and which is the machine.
   
3. If the evaluator cannot reliably tell the machine from the human, the machine is said to have passed the Turing Test. Passing the test suggests that the machine's responses are indistinguishable from those of a human, implying a level of intelligence or at least a convincing simulation of human intelligence.

The Evolution of AI: Beyond Turing's Imagination
Since Turing's era, AI has transformed from theoretical models to practical applications. Today, AI is becoming integrated into various aspects of our lives, from virtual assistants to autonomous vehicles. This rapid development has surpassed even Turing's predictions, demonstrating the limitless potential of AI.

AI in the Modern World: Opportunities and Challenges
Modern AI systems can process and analyze vast amounts of data, learn from experiences, and make decisions with minimal human intervention. While these capabilities offer immense benefits, they also present unique challenges, such as ensuring fairness, transparency, and accountability in AI decisions.

Source: 
Turing, A. M. “Computing Machinery and Intelligence.” Mind, vol. LIX, no. 236, Oct. 1950, pp.
433–460.​

Unconscious biases, as explored by Shankar Vedantam in "The Hidden Brain," are an integral concept to understand when exploring AI ethics.  These biases exert a pervasive influence over our perceptions and actions, often operating without our awareness, and their ramifications extend deeply into our personal and societal interactions.

An eye-opening example from the book underscores the profound impact of unconscious bias. Research conducted at a Montreal day-care center revealed that even toddlers as young as three years old displayed racial categorization tendencies. These young children consistently associated white faces with positive attributes and black faces with negative ones, providing early evidence of the emergence of racial bias in their cognitive processes. Vedantam emphasizes that these associations, though observed in very young children, are not inherently biological but are predominantly shaped by the cultural and environmental influences they encounter during their formative years.

Another powerful illustration presented in the book revolves around a study involving job applications. The applications were identical in qualifications, but they carried different names indicating various races. Perhaps unsurprisingly, applications with white-sounding names received significantly more callbacks for interviews compared to those with African-American-sounding names. This experiment starkly underscores how unconscious bias can sway hiring decisions, even among individuals who consciously uphold principles of equality and fairness.

Moreover, Vedantam adeptly portrays how unconscious biases permeate our daily interactions. Our brains, in their quest for efficiency, often rely on shortcuts and stereotypes to rapidly process information, resulting in biased judgments. For example, we might unconsciously associate specific clothing styles with negative traits, leading to unjust treatment or unwarranted fear of individuals who pose no genuine threat.

Even in the realm of healthcare, where doctors are typically well-intentioned and highly trained, unconscious biases can insidiously influence diagnosis and treatment decisions. Research indicates that racial disparities in treatment exist, with African American patients being less likely to be referred for specific medical procedures despite presenting similar symptoms. Additionally, Vedantam highlights the role of media in perpetuating unconscious biases. Biased portrayals of certain groups in the media reinforce stereotypes and significantly shape public opinion, thereby nurturing prejudice and discrimination.

Despite the omnipresence of unconscious biases, Vedantam offers a glimmer of hope by emphasizing our capacity to confront and overcome them. This journey begins with introspection and a willingness to challenge our own biases, as self-awareness serves as the crucial first step in recognizing and questioning automatic associations and judgments. Actively seeking diverse perspectives and experiences serves as a valuable strategy to broaden our understanding of others and mitigate the influence of unconscious bias. He also discusses effective strategies and interventions to mitigate the impact of unconscious biases. Techniques such as blind auditions and blind evaluations in hiring processes have proven successful in removing the influence of biases, focusing solely on an individual's talent and merit.

As we navigate a world increasingly shaped by artificial intelligence, it's essential to recognize that AI systems can also inherit and perpetuate these unconscious biases. Just as we strive to confront and mitigate our own biases, we must design and train AI algorithms to do the same. A conscious effort to create AI systems that are free from the biases ingrained in our society is pivotal to building a fair, just, and equitable future where both humans and machines coexist harmoniously. Vedantam's work inspires us to confront our own hidden biases and actively strive for a more conscious and just world, both in our human interactions and in the technology we create.

Source: 
Vedantam, S. (2009). The Hidden Brain: How Our Unconscious Minds Elect Presidents, Control Markets, Wage Wars, and Save Our Lives. 
Buy now! 

​​Before we can delve into the intricacies and specifics of AI ethics and biases, it is crucial to establish a foundational understanding of the biases inherent in AI and the reasons for their existence. This blog post draws insights from two informative articles to shed light on this complex issue. The first article, authored by Karen Hao and published in MIT Technology Review in February 2019, underscores that AI bias cannot be solely attributed to biased training data; instead, it has nuanced origins throughout the deep-learning process. The second article by Jake Silberg and James Manyika at McKinsey, published in June 2019, explores opportunities to mitigate human biases through AI and the pressing need to improve AI systems to prevent the perpetuation of human and societal biases. These articles emphasize that while AI holds the potential to alleviate biases, it also carries the risk of exacerbating them if not managed carefully, making it essential to understand the mechanics of AI bias.

Cause of AI Bias:
AI bias is a multifaceted challenge originating from various sources beyond biased training data. It arises from human biases that impact decision-making, both consciously and unconsciously. Biased data, reflecting historical prejudices or societal inequities, can perpetuate these biases when used to train AI models. Additionally, biases can infiltrate data collection processes, such as oversampling specific demographics due to over-policing. The choices made during algorithm development, like selecting attributes for consideration, can introduce bias, affecting model predictions and fairness. Defining fairness in AI is complex, as there are various definitions with inherent trade-offs between them, making it challenging for AI systems to conform to multiple fairness metrics simultaneously.

Addressing AI Bias:
Mitigating AI bias is an ongoing challenge that demands careful consideration. It involves various strategies and considerations. Using AI to reduce human bias is one approach, enabling more objective decision-making by relying on relevant data rather than subjective factors. Transparency and accountability are vital, requiring organizations to establish processes for testing and mitigating bias in AI systems, including auditing data and models for fairness. Collaboration between humans and AI is essential, with human judgment complementing AI recommendations in decision-making processes. Interdisciplinary collaboration across fields, including ethics and social sciences, is necessary to develop standards for bias and fairness. Encouraging diversity in the AI community can provide unique insights and perspectives in addressing bias issues.

Conclusion:
AI bias is a multifaceted challenge that necessitates a comprehensive approach. While AI has the potential to reduce human biases, it also carries the risk of amplifying them. Achieving fairness and ethics in AI requires ongoing research, interdisciplinary collaboration, transparency, and accountability.

Sources:
Karen Hao, "This is how AI bias really happens—and why it’s so hard to fix," MIT Technology Review, February 4, 2019.
Jake Silberg and James Manyika, "Tackling bias in artificial intelligence (and in humans)," McKinsey Global Institute, June 6, 2019.