AI in Society
The blog is connected to my role of the head of the National Institute on AI in Society
at California State University Sacramento. However, opinions and positions expressed therein are mine, and do not represent the university's opinions or positions.
Tuesday, March 5, 2024
Let's hallucinate, take the kids along
Interestingly, this property of AI models reveals something profound about the human mind as well. Our remarkable ability to imagine, create and solve problems is inextricably linked to our capacity for error. Just like LLMs, human thinking is fundamentally generative and probabilistic - we are constantly making predictions and filling in gaps based on prior knowledge and contextual cues. And in doing so, we inevitably make mistakes, jumping to conclusions and seeing patterns where none exist. In a sense, "hallucination" is a built-in feature of human cognition, not a bug.
This insight has important implications for how we approach education and learning. Too often, our educational systems are overly focused on eliminating errors and inculcating "correct" answers. While accuracy is certainly important, an excessive emphasis on being right all the time can stifle creativity and limit our ability to generate novel ideas and solutions. To truly tap into the power of the human mind, we need to create space for productive mistakes and flights of imagination.
So perhaps we should spend less time trying to prevent students from ever being wrong, and more time teaching them how to recover from errors, distinguish fact from fantasy, and harness their imaginative abilities in positive ways. By embracing a bit of beneficial "hallucination", we may actually enhance our ability to discover truth and expand the boundaries of human knowledge. The key is striking the right balance - letting our minds roam free, while also exercising our critical faculties to rein in our fantasies when needed. In this way, we can learn from the foibles of AI to better understand and cultivate the powers of our own marvelous minds.
Saturday, March 2, 2024
Prompt as a magic incantation
In engagements with AI, the crafting of prompts—a crucial interface between human intention and machine output—has acquired an almost mystical significance for some users. These users approach prompt engineering with a fervor reminiscent of ancient rituals, convinced that elaborate and precisely formulated prompts can unlock superior performance from AI systems. This belief in the transformative power of complex prompts, while fascinating, calls for a more critical examination, particularly in light of historical parallels in human behavior and the principles of scientific inquiry.
The comparison to B.F. Skinner's 1948 study, "Superstition in the Pigeon," is particularly apt. Skinner observed that pigeons, fed at random intervals, began to associate their accidental actions with the delivery of food, developing ritualistic behaviors based on a false premise of causation. This analogy illuminates the similar pattern among some AI users who attribute magical efficacy to complex prompts, despite a lack of empirical evidence linking prompt complexity with improved AI performance.
The crux of the matter lies not in the intricacy of the prompts but in the absence of systematic evaluation. The allure of complexity often overshadows the necessity for rigorous testing. Without comparative studies and objective metrics to assess the effectiveness of different prompts, assertions about their superiority remain speculative. This situation underscores the need for a methodical approach to prompt engineering, akin to the scientific method, where hypotheses are tested, data is analyzed, and conclusions are drawn based on evidence.
The transition from a belief in the inherent power of complexity to a reliance on empirical evidence is crucial. Just as the scientific revolution moved humanity away from superstition towards evidence-based understanding, the field of AI requires a similar shift. Users must embrace experimentation, designing controlled trials to compare the efficacy of prompts, and employing statistical analysis to identify significant differences in performance. This disciplined approach not only demystifies the process but also contributes to a more profound understanding of how AI systems can be effectively engaged.
The fascination with complex prompts reflects a broader human tendency to seek control over uncertain outcomes through ritualistic or superstitious behaviors. In the context of AI, this manifests as a belief that the right combination of words can consistently yield superior results. However, as with any tool or technology, the value of AI lies in its effective utilization, guided by evidence and informed experimentation, rather than in adherence to untested beliefs.
Wednesday, February 28, 2024
Hackers vs. Handlers: The Battle for Equity in the Generative AI Revolution
Handlers, often backed by deep pockets and corporate interests, are the architects of proprietary systems. They package, distribute, and regulate access to AI technologies, aiming to create comprehensive solutions that cater to market demands. Their approach, while fostering innovation and ensuring quality, often leads to restricted access and a consolidation of power, raising concerns about equity and inclusivity in the technological realm. The curious fact is that many handlers are former hackers, who made it in the startup world.
Hackers, in contrast, are the rebels of the digital age. They advocate for a more open and collaborative approach to AI development, believing that technology should be a public good, accessible to all. They prefer the do-it-yourself, scrappy solutions. Their efforts are not driven by profit but by a conviction that broader access to AI tools can level the playing field, enabling a more diverse set of voices to contribute to and benefit from technological advancements.
The clash between hackers and handlers is emblematic of a larger debate about the future of technology and its role in society. While handlers bring structure and scalability, hackers inject diversity, creativity, and a sense of community. The balance between these forces is crucial. An overemphasis on control and commercialization risks stifling innovation and perpetuating inequalities, while unchecked openness may lead to issues of quality and security.
The generative AI revolution presents an opportunity to recalibrate this balance. Supporting hackers and their open-access ethos can foster a more equitable technological landscape, where innovation is not the exclusive domain of the well-funded. This means championing open-source projects, recognizing community-driven initiatives, and creating legal frameworks that protect the principles of openness and collaboration.
As we stand at the precipice of this AI revolution, the choices the societies make will have far-reaching implications. Supporting the hacker ethos without alienating the handlers, and promoting broader access to AI technologies can ensure that the benefits of this revolution are shared by all, not just the privileged few. It is time to shift the balance in favor of equity, inclusivity, and the collective advancement of society.
Saturday, February 17, 2024
Curb your enthusiasm
Do we learn how to use the current versions of AI, or wait for them to get much better very soon? The excitement around AI's exponential growth mirrors a pattern we've seen with other technologies: a burst of initial progress followed by the hard reality of limitations. History offers lessons from nuclear fusion to space exploration, where initial optimism ran into practical and technological barriers.
Nuclear fusion, which began its journey as a promising energy solution in the 1950s, has yet to deliver on its promise of endless clean energy. The technical and financial challenges have proven to be more complex and enduring than anticipated. Similarly, space exploration, once thought to usher in an era of human settlement in outer space, has been tempered by the harsh realities of cost, distance, and survival in a hostile environment.
As AI technologies, particularly generative AI like ChatGPT, race ahead, they too may face significant hurdles. The rapid development and deployment of these technologies have revealed challenges, notably the increasing demand for computing power. This situation is exacerbated by the competitive push from tech giants like Google and Meta, highlighting the difficulty of sustaining rapid advancement.
One potential game-changer on the horizon is quantum computing. This emerging field promises to revolutionize computing power, potentially overcoming current limitations in a way we can barely imagine. The impact of quantum computing on AI could be profound, offering solutions to problems that are currently intractable and opening new avenues for advancement.
Yet, even with quantum computing, it's wise to temper our expectations, at least until practical and cheap quantum computers become a reality. Each technological leap brings its own set of challenges and unknowns. Rather than waiting for miraculous breakthroughs, a more pragmatic approach is to focus on optimizing current AI technologies. Understanding and working within their limitations can lead to significant improvements and applications that are both practical and impactful now.
This approach doesn't mean halting innovation but rather balancing the pursuit of new technologies with the efficient exploitation of existing ones. By learning from the past and being mindful of the inherent challenges in technological progress, we can navigate the complexities of innovation more effectively. Quantum computing may indeed provide the next significant leap, but until then, making the most of current AI capabilities is both a wise and necessary strategy.
Friday, February 9, 2024
The Advising Bot Dilemma
In educational organizations, the integration of AI, particularly through automated advising tools like chatbots, embodies a strategic advancement yet introduces a complex dilemma. These digital advisors, designed to navigate queries ranging from academic programs to student services, highlight a pivotal choice between precision and broad utility.
At one pole, AI bots can be meticulously engineered to handle vaguely formulated inquiries, but only providing correct answers manually curated by humans. This approach, while ensuring a high level of fidelity, is marked by a slow and expensive development process. For entities with vast knowledge bases or intricate operations, the manual input required could significantly dilute the efficiency gains such tools promise to deliver.
Conversely, AI advisors programmed for wider application operate by not only interpreting queries, but also sourcing answers from a pre-existing repository of documents and websites. This method, though expedient, compromises on accuracy, a drawback that becomes more pronounced within the context of large and diverse information repositories.
A balanced strategy proposes the coexistence of both high and low-fidelity bots within the educational sphere. Low-fidelity bots offer an expedient first layer of support, adept at managing basic inquiries through triage advising. Tailoring these bots to specific domains and incorporating clear disclaimers could mitigate the risk of misinformation, directing students towards accurate resources while alleviating the administrative burden on staff.
For situations where accuracy is paramount, a semi-automatic model emerges as a superior alternative, at least for now. This model envisions a symbiotic relationship between AI systems and human advisors, with AI proposing potential responses and the advisor ensuring their validity. Such a configuration enhances efficiency without compromising the integrity of the advice provided.
AI imperfections sometimes may be tolerated. AI adoption required a pragmatic cost-benefit analysis. The evaluation hinges on whether the operational efficiencies gained through deploying lower-fidelity systems justify the associated risks. We must compare them not to very expensive and very reliable alternative, but to not getting any advicу at all, or receiving it from roommates and random sources. The decision on whether to limit these systems to straightforward queries or to implement them within defined subject areas requires careful consideration.
Addressing these trade-offs is crucial for harnessing AI's potential in educational settings. This nuanced approach, advocating for a judicious blend of high and low-fidelity advising tools, underscores the importance of strategic planning in AI deployment. It offers a pathway to leverage technological advancements, ensuring they complement rather than complicate the educational mission.
Tuesday, February 6, 2024
AI undermines linguistic privilege
The tremors of unease felt across the echelons of privilege are not solely due to the fear of technological unemployment or the unsettling pace of change. Rather, they stem from a deeper, more introspective anxiety: the threat AI poses to the use of language as a bastion of privilege. For centuries, mastery over the nuanced realms of oral and written speech has served as a subtle yet potent tool of social stratification, a way to gatekeep the corridors of power and influence. But as AI begins to democratize these linguistic capabilities, it inadvertently challenges the very foundations of societal hierarchies, provoking a backlash draped in ethical rhetoric that masks a more self-serving agenda.
Language, in its most refined forms, has long been a marker of education, sophistication, and belonging. To speak with the clipped accents of an upper-class Englishman, to wield the jargon of academia, or to navigate the complex conventions of professional communication has been to hold a key to doors otherwise closed. These linguistic markers function as tacit gatekeepers, delineating who belongs within the inner circles of influence and who remains outside, their voices deemed less worthy. The assertion that one must speak or write in a certain way to be considered intelligent or capable reinforces societal power structures and perpetuates inequities. It's a subtle form of oppression, one that privileges certain dialects, accents, and syntactical forms over others, equating linguistic conformity with intelligence and worthiness.
Enter the realm of artificial intelligence, with its natural language processing capabilities and machine learning algorithms. AI, with its inherent impartiality to the accents, dialects, and syntactical structures it mimics, does not discriminate based on the traditional markers of linguistic prestige. It can generate scholarly articles, craft professional emails, or compose poetic verses with equal ease, regardless of the socioeconomic or cultural background of the user. This leveling of the linguistic playing field poses a direct challenge to those who have historically leveraged their mastery of language as a means of maintaining status and privilege.
Critics of AI often cloak their apprehensions in the guise of ethical concerns, voicing fears about data privacy, algorithmic bias, or the dehumanization of communication. While these issues are undoubtedly important, they sometimes serve to obscure a more uncomfortable truth: the democratizing impact of AI on language threatens to undermine traditional power dynamics. The reluctance to embrace this technology fully may, in part, stem from a reluctance to relinquish the privilege that comes with linguistic mastery.
This resistance to change is not a new phenomenon. Throughout history, technological advancements have often been met with skepticism by those whose status quo they disrupt. The printing press, the telephone, and the internet all faced initial pushback from those who feared the loss of control over information dissemination. Similarly, AI's impact on language is merely the latest battleground in the ongoing struggle between progress and privilege.
Yet, the equalizing potential of AI should not be viewed with apprehension but embraced as an opportunity for societal advancement. By breaking down the barriers erected by linguistic elitism, AI can facilitate more inclusive, diverse forms of communication. It can empower individuals from all backgrounds to express themselves effectively, participate in scholarly discourse, and compete in professional arenas on equal footing. In doing so, AI can help to dismantle some of the systemic barriers that have perpetuated inequality and hindered social mobility.
The anxiety surrounding AI's impact on language reflects broader concerns about the erosion of traditional forms of privilege. As AI continues to advance, it challenges us to reconsider the values we ascribe to certain forms of linguistic expression and to question the fairness of societal structures built upon them. Embracing the democratizing influence of AI on language could lead to a more equitable and inclusive society, where intelligence and capability are recognized in all their diverse expressions, rather than gauged by adherence to arbitrary linguistic norms. In the end, the true measure of progress may not be in the sophistication of our technologies but in our willingness to let go of outdated markers of privilege.
Friday, February 2, 2024
AI in Education Research: Are We Asking the Right Questions?
A recent preprint titled " Generative AI Can Harm Learning " has attracted significant attention in education and technology circl...
-
In the ongoing narrative of education's transformation, AI's integration has prompted a profound reassessment of what constitutes un...
-
The notion of measuring "AI-generated text" as a fixed percentage of an academic submission is fundamentally flawed. This metric i...
-
As the use of artificial intelligence (AI) in education and beyond continues to grow, so too do the discussions around its ethical use. Howe...