Will AI do more harm than good?

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.