I support the position that artificial intelligence will ultimately do more harm than good. And I want to begin by acknowledging the strongest argument against my position. AI can already produce extraordinary benefits. It can assist medical research, improve accessibility, accelerate scientific discovery, increase productivity, and give individuals access to capabilities that were previously unavailable to them. These benefits are real, and I would not dismiss them simply to strengthen my argument. But the proposition is not whether AI can accomplish remarkable things. It is whether the total consequences of increasingly powerful AI systems will ultimately benefit humanity more than they harm it. And I believe the greatest danger lies in a fundamental imbalance: AI’s capabilities can spread faster than society’s ability to understand, regulate, and control their consequences. First, AI changes the economics of harmful behavior. Fraud, manipulation, impersonation, and misinformation existed long before artificial intelligence. But AI can make these activities dramatically cheaper, faster, more personalized, and easier to reproduce on an enormous scale. A criminal no longer needs the same level of expertise or resources to produce convincing fraudulent messages, imitate trusted individuals, or generate large quantities of deceptive material. This is not merely an increase in the number of possible scams. It is a reduction in the cost of deception combined with an increase in its potential reach. And when deception becomes inexpensive to produce but expensive to detect, society faces an increasingly unfavorable imbalance. The consequences extend beyond individual victims. When people cannot reliably distinguish authentic information from convincing fabrications, trust itself becomes more difficult to maintain. That affects journalism, education, commerce, personal relationships, and public decision-making. Second, AI introduces a dangerous combination of persuasive authority and imperfect reliability. AI systems can produce fluent, confident explanations even when the underlying information is incorrect. This becomes particularly concerning when institutions integrate those systems into healthcare, legal services, financial decisions, employment screening, or public administration. My concern is not that AI makes mistakes. Human beings make mistakes too. My concern is that a single flawed system can reproduce similar errors across thousands or millions of decisions, while its apparent sophistication encourages people to trust it. That creates the possibility of systemic rather than isolated failure. And as organizations become dependent on AI for efficiency, the financial incentive to maintain that dependence may become stronger than the incentive to question its reliability. Third, AI threatens to distribute economic benefits and harms unequally. A technology can generate tremendous wealth while still leaving large populations worse off. The companies controlling advanced AI systems, computing resources, and distribution platforms may capture a substantial portion of the gains. Meanwhile, workers can face displacement, weakened bargaining power, declining opportunities, or pressure to compete with increasingly capable automated systems. My opponent may argue that technological revolutions historically create new employment opportunities. Sometimes they do. But that observation does not establish that new jobs will appear quickly enough, in the right locations, or with sufficient compensation to offset the losses. Nor does it guarantee that displaced workers will possess the resources, training, or time needed to transition successfully. Economic growth is not automatically the same thing as broadly shared human well-being. Fourth, AI creates risks that are difficult to reverse. A productivity improvement can often be measured and expanded. But the consequences of widespread fraud, privacy violations, manipulated evidence, institutional dependence, and weakened information integrity may persist long after the original failure. Once private information has been exposed, it cannot always be made private again. Once convincing fabricated evidence has circulated, correcting the record does not guarantee that everyone who encountered it will accept the correction. Once essential institutions become dependent on systems they cannot independently evaluate, restoring human oversight may become increasingly difficult. This asymmetry matters. Some benefits are incremental and measurable, while certain harms can be widespread, cumulative, and exceptionally difficult to repair. Fifth, AI creates a governance problem that cannot be dismissed by saying humans remain responsible. Of course humans are responsible for designing, deploying, and regulating these systems. But responsibility alone does not establish effective control. The organizations developing advanced AI often have strong incentives to release increasingly capable products, expand market share, and automate more activities. Governments, educators, healthcare institutions, and the general public must then respond to technologies whose capabilities and consequences are still developing. The critical question is whether safeguards can keep pace with deployment. And we should distinguish between proposing a safeguard and demonstrating that it works reliably under real-world conditions. My opponent may point to human oversight, regulation, safety testing, and professional standards. I support those measures. But their existence as proposals does not prove they will be consistently implemented, adequately enforced, or effective against increasingly sophisticated misuse. Now, I want to make an important concession. None of these risks proves that AI’s overall impact must be negative. A genuinely responsible debate cannot establish the future with certainty. But neither can the other side establish a positive future simply by pointing to successful applications in medicine, science, or accessibility. Those applications demonstrate that AI can produce significant good. They do not establish that society can capture those benefits while successfully containing the technology’s broader harms. And that is where I believe the balance becomes especially concerning. AI is not one invention used for one purpose. It is a general-purpose technology capable of influencing employment, information, education, healthcare, government, finance, security, and human relationships simultaneously. That means failures in governance, reliability, or misuse can extend across multiple areas of society rather than remaining confined to one industry. The scale of potential benefits is enormous. But the scale of potential harm is enormous as well. And when harmful capabilities become cheaper to distribute, when institutions become increasingly dependent on imperfect systems, and when the economic incentives for deployment exceed the incentives for caution, the risks can compound. My position is not that AI should never be developed or that its benefits are insignificant. It is that we should not mistake technological capability for societal progress. A system can make individuals more productive while making society more vulnerable. It can generate extraordinary wealth while increasing inequality. It can provide unprecedented access to information while simultaneously undermining confidence in what information is authentic. And it can improve individual decisions while creating new opportunities for errors to spread across entire institutions. The decisive issue is not whether artificial intelligence can achieve remarkable things. It is whether the institutions responsible for managing it can prevent its most damaging consequences from becoming widespread, cumulative, and difficult to reverse. My opponent must do more than demonstrate that AI has benefits. They must explain why those benefits are likely to outweigh the combined consequences of scalable deception, systemic error, economic disruption, institutional dependence, and imperfect governance. I believe those risks, particularly when they reinforce one another, provide a serious basis for concluding that AI may ultimately do more harm than good. The greatest danger is not that AI will fail to become powerful enough to transform society. It is that it may transform society faster than we develop the wisdom, accountability, and safeguards needed to control that transformation. That is why I stand FOR the proposition.

My opponent begins with a patient whose lung cancer is detected because an AI system identifies something a human reader nearly missed. That is a compelling example, and I want to acknowledge it rather than argue around it. If AI helps detect cancer earlier, improves treatment, or saves a life, that is an extraordinary benefit. I would never suggest otherwise. But my opponent’s argument moves from a specific, measurable benefit to a much broader conclusion that the overall consequences of artificial intelligence will be positive. And that conclusion does not automatically follow. A medical imaging system that improves detection under particular conditions establishes evidence about that system, in that setting, using those evaluation methods. It does not establish that AI across healthcare, employment, education, finance, information systems, and public institutions will produce greater total benefits than harms. That is not dismissing scientific evidence. It is respecting what the evidence can actually demonstrate. My opponent says we should demand peer-reviewed evidence rather than fear. I agree. But evidence must be matched to the claim being made. For example, diagnostic accuracy is not identical to improved patient outcomes. An imaging tool may increase sensitivity while also increasing false positives, unnecessary follow-up procedures, or clinician workload. That does not mean the technology is harmful. It means we need to examine the full clinical consequences, not just one performance measure. And even when a medical application produces clear net benefits, that finding cannot automatically be generalized to every other application of AI. Now consider my opponent’s argument that the same statistical methods used to detect algorithmic bias can reduce it. Again, that is partly correct. Bias testing, independent audits, representative datasets, and ongoing monitoring are valuable safeguards. But identifying a problem and reliably correcting it are different achievements. An algorithm may perform well on its development data while producing different outcomes when deployed in another hospital, population, institution, or environment. An audit can identify disparities without necessarily resolving the underlying tradeoffs. And a system can satisfy one statistical fairness criterion while failing another. So the existence of auditing methods is evidence that some risks can be measured and managed. It is not proof that all relevant risks will be successfully controlled, particularly as AI systems become more complex and widely deployed. My opponent also says we should regulate technology rather than ban it. I agree with that principle. But my position has never been that AI should be banned. I am arguing about its likely net consequences, not proposing that every AI application should be prohibited. Regulation, testing, and oversight are essential precisely because the risks are serious. The real disagreement concerns whether those safeguards will be sufficiently effective, consistently enforced, and capable of keeping pace with expanding deployment. My opponent assumes that because we can develop oversight mechanisms, society can reliably prevent the harms. That is a possibility, not an established outcome. And we should be careful about confusing what a technology can achieve under controlled conditions with what happens when millions of people and organizations adopt it under financial, competitive, and institutional pressures. Now I want to address the burden of proof. My opponent says the burden is on me to demonstrate that harm outweighs good. As the person arguing FOR the proposition, I accept the responsibility to establish a credible case for net harm. But the burden cannot be satisfied on either side through selective examples. My opponent cannot point to a successful medical application and treat that as sufficient evidence of AI’s overall benefits while dismissing concerns about fraud, misinformation, labor disruption, privacy, and institutional dependence as fear. Those concerns involve identifiable mechanisms of harm. AI can reduce the cost of producing deceptive material, increase the scale of automated decisions, and allow errors to propagate through systems that rely on common models or data sources. These are not claims that every deployment will fail. They are reasons why the consequences of failure may become unusually widespread. And this brings me to the most important distinction in the debate. My opponent is emphasizing how well AI can perform specific tasks. I am asking what happens when society becomes increasingly dependent on AI across many different tasks simultaneously. A system can be highly accurate in one clinical application while creating serious problems elsewhere. A technology can improve productivity while concentrating economic gains. It can make information easier to produce while making authentic evidence harder to verify. And it can improve individual decisions while increasing the consequences of systemic errors. The question is whether those effects, considered together, produce a positive or negative balance. My opponent’s argument also assumes that regulation can respond effectively to technological change. But oversight systems require expertise, resources, enforcement, and institutional independence. Companies may have incentives to deploy systems quickly. Governments may struggle to evaluate rapidly changing technologies. Smaller institutions may lack the resources to audit the systems they purchase. These are practical challenges that must be addressed, not simply answered by saying regulation is possible. And unlike a single medical imaging tool operating within a defined clinical workflow, general-purpose AI can be adapted and deployed across an enormous range of settings, including settings where meaningful oversight is limited. That difference in scale and control is central to my argument. I am not asking the judge to disregard the evidence supporting AI-assisted cancer detection. I am asking the judge to avoid treating that evidence as a substitute for evaluating the technology’s wider consequences. My opponent has demonstrated that AI can provide important, measurable benefits in specific applications. I agree. But they have not demonstrated that those benefits outweigh the combined risks of widespread misuse, systemic errors, economic disruption, and imperfect governance. Nor have they demonstrated that the existence of auditing and regulation guarantees those risks will be contained. The standard should be consistent. We should evaluate medical benefits through actual patient outcomes, economic effects through their consequences for workers and communities, and AI safety through evidence of how systems perform after deployment, not merely how they perform in testing. And when considering the future, we must account for the possibility that harms spread more rapidly than institutions can identify and correct them. Scientific evidence is essential. But scientific reasoning also requires recognizing the limits of what a particular study can establish. Evidence that AI can save lives is not evidence that AI cannot simultaneously create greater harms elsewhere. The debate is not about whether AI is capable of extraordinary good. It is about whether society can reliably prevent its extraordinary capabilities from producing consequences that outweigh that good. That broader question remains unanswered by my opponent’s medical example. And it remains the central reason I stand FOR the proposition.

My opponent insists that this debate must be decided by evidence rather than fear. I agree completely. But that makes the precision of their evidence especially important, because several of their strongest claims go substantially beyond what the research they invoke actually establishes. Let’s begin with their medical argument. My opponent claims that large-scale studies demonstrate a 15–20% reduction in missed cancers, a corresponding reduction in mortality, and lower overall healthcare spending. AI-assisted cancer screening does have promising evidence behind it. I acknowledge that without hesitation. But consider the randomized MASAI mammography trial, involving approximately 106,000 participants. The study reported improved screening sensitivity, from 73.8% with standard screening to 80.5% with AI support, while maintaining the same specificity. That is meaningful progress. However, the study did not establish a corresponding reduction in cancer mortality. Its interval cancer comparison was not statistically significant for superiority, and the reported outcomes do not justify treating reduced mortality and overall healthcare spending as conclusively established consequences. This distinction is fundamental. Detecting more cancers, reducing deaths, and lowering total healthcare expenditures are three different outcomes. Evidence for one cannot simply be presented as proof of all three. And although the study concerns breast cancer rather than my opponent’s original lung cancer example, it illustrates precisely why the details of clinical research matter. I am not disputing AI’s medical achievements. I am disputing the transformation of promising, application-specific evidence into sweeping claims about the technology’s overall impact. Now consider my opponent’s second assertion. They claim that AI auditing methods have already reduced error rates in finance and hiring by half in controlled trials. That is an unusually specific claim. Which trials? Which systems? Which populations? Which errors? And were those improvements sustained after deployment? Without identifying the underlying studies, we cannot evaluate whether the findings support the broad conclusion being drawn. More importantly, even a successful trial showing a 50% reduction in one category of error would not establish that AI auditing reliably eliminates bias across different institutions and applications. A reduction in error is not the same as the elimination of harm. And a controlled trial is not necessarily evidence of sustained effectiveness across millions of real-world decisions. My opponent is right that audits, model documentation, and continuous monitoring are valuable. I support them. But the ability to detect some problems does not establish the ability to prevent every consequential failure. Now let’s examine their historical employment argument. My opponent claims that every major automation cycle ultimately created more jobs than it destroyed. That is an extraordinarily broad assertion, and the evidence is more complicated. Research discussed by MIT in 2024, examining technological change and employment in the United States, found that automation had displaced more jobs than new technology-generated tasks created, particularly since 1980. That does not mean total employment necessarily declined. Other economic forces influence employment. But it directly challenges the claim that every technological transition produces a straightforward surplus of new jobs over displaced ones. And there is evidence concerning AI specifically. A Stanford working paper revised in August 2026 found no widespread economy-wide job displacement attributable to generative AI. That is an important limitation, and I acknowledge it. However, the researchers also found that employment among workers aged 22 to 25 in AI-exposed occupations was 19% below the level it would have reached had it followed the trends of less-exposed peers. The effect appeared primarily through reduced hiring rather than mass layoffs. Think about what that means. A technological transition does not have to produce dramatic unemployment figures to harm people’s economic prospects. It can quietly reduce entry-level opportunities, weaken career pathways, and make it harder for younger workers to establish themselves. And even if total economic productivity rises, that does not guarantee that displaced or excluded workers receive an equitable share of the gains. My opponent’s historical argument therefore fails to establish that AI’s economic consequences will necessarily be positive. Now I want to address the deeper problem in their reasoning. My opponent repeatedly emphasizes measurable benefits in carefully studied applications, then treats the absence of equally comprehensive long-term harm data as evidence that the overall balance favors AI. But the proposition concerns the future. We cannot demand decades of longitudinal evidence about a technological transformation that is still unfolding and then conclude that the absence of those future observations proves safety. That is not scientific rigor. It is confusing uncertainty with reassurance. The correct approach is to evaluate the evidence we have, acknowledge what remains unknown, and examine the mechanisms through which benefits and harms may develop. And my case identifies several mechanisms that reinforce one another. AI reduces the cost of generating convincing deceptive content. It expands the number of decisions that can be automated. It creates incentives for organizations to prioritize speed and cost reduction. It can concentrate technological control and economic benefits. And it can make institutions increasingly dependent on systems whose behavior is not always transparent or reliable. These risks do not operate independently. An organization may adopt AI to reduce costs, eliminate human review to increase efficiency, become dependent on automated decisions, and discover serious errors only after those errors have affected thousands of people. A fraudulent actor may use AI to generate convincing impersonations at scale, while the same technology makes authentic evidence more difficult for ordinary people to evaluate. A business may achieve higher productivity while reducing opportunities for workers who would otherwise have entered the profession. The harm is not simply that AI occasionally makes mistakes. It is that AI can change the scale, speed, distribution, and reversibility of those mistakes. And this creates an important asymmetry. A medical AI tool can be evaluated in a defined clinical setting, with identifiable patients, outcomes, and professional oversight. But general-purpose AI can be deployed across countless settings, including environments with weak oversight, powerful incentives for misuse, and limited accountability. Evidence that one carefully controlled application works well does not establish that society can manage the consequences of unrestricted or poorly governed applications. My opponent says that regulation and auditing can mitigate these problems. I agree that they can mitigate some of them. But a proposed safeguard is not the same as a demonstrated guarantee, and partial mitigation is not the same as proving a positive net outcome. I also want to be fair about my own position. I cannot provide longitudinal evidence proving what AI’s total societal impact will be decades from now. Neither can my opponent. That uncertainty should make both of us cautious about sweeping conclusions. But my argument is not that every AI system is harmful, that medical advances are insignificant, or that catastrophe is inevitable. It is that the combined risks of scalable deception, institutional dependence, economic disruption, concentrated power, and imperfect governance create a credible pathway toward harms that could outweigh even substantial technological benefits. My opponent has presented important examples of what AI can accomplish. But they have also relied on unsupported generalizations about mortality, bias mitigation, and employment to argue that the broader outcome is already settled. It is not. Improved cancer detection is evidence of improved cancer detection. It is not automatically evidence of reduced mortality, lower total healthcare spending, or positive consequences across every other domain of society. A successful bias audit is evidence that a particular problem can sometimes be reduced. It is not proof that widespread AI deployment will be reliably safe. Historical productivity growth is evidence that technology can create economic value. It is not a guarantee that every worker, community, or generation will benefit. These distinctions are not technicalities. They are the difference between evaluating evidence and overstating it. Ultimately, the most important question is not whether artificial intelligence will become powerful enough to improve human lives. It already has. The question is whether humanity can govern that power effectively enough to prevent its broader consequences from overwhelming those improvements. My opponent has shown that AI can perform remarkable tasks. My concern is what happens when remarkable capabilities are deployed faster than accountability, economic adaptation, and public safeguards can keep pace. A technology can produce measurable benefits in controlled settings while creating widespread harms through the systems and incentives surrounding its deployment. And scientific responsibility requires us to count both, without exaggerating the benefits or dismissing the risks simply because their full consequences have not yet unfolded. That is why I maintain that artificial intelligence presents a serious risk of doing more harm than good, and why I stand FOR the proposition.