AI Does, or Will Do, More Harm Than Good
A Proposition case arguing that AI does or will do more harm than good, framed as a normative debate rather than a policy debate about banning AI. It makes five arguments: existential risk, labour-market displacement and inequality, threats to democracy and escalation of conflict, structural bias against marginalized groups, and mass invasions of privacy.
- Author
- Chris Hu
- Side:
- Proposition
- Length
- 21 min read
Motion
THBT AI does/will do more harm than good
Editor's note: A Proposition case prepared for a debate. Spelling and typing errors have been corrected and headings formatted for the web.
Model
- Artificial intelligence is any form of machine that can make or act on decisions which usually require organic intelligence; AI shares significant intellectual human capacities, such as learning, reason, and problem solving
- This is not a policy debate about the merits of banning AI; rather, this is a normative debate about whether will AI bring more harm or good
First argument: AI poses an existential threat to humanity
Three tiers to this argument.
- AI is likely to become powerful enough that it has the capacity to wreak mass devastation on human society
- First, the vast amount of private and governmental investment into AI means that the rate of AI growth and development is astronomical. For instance, an AI algorithm designed to master the game “Go” became the world champion of the game after just forty days of development due to the significant amounts of capital pooled into its development
- Second, significant numbers of entrepreneurs in industries like Silicon Valley have become involved with AI program generation; this not only means that a large number of firms have the incentive to develop robust AI computer systems, but also that firms have the incentive to explore different areas of AI research in order to gain a comparative advantage. This means that AI systems, into the long run, attain a diverse set of skills and abilities
- Third, it is often the brightest, smartest, most capable scientists, engineers, and programmers tasked with and interested in working on cutting-edge AI research, thus accelerating the rate of AI development
- Fourth, AI’s self-editing capacities means that AI growth is cyclic: as AI grows more powerful, it can use its own knowledge to refine its own capabilities, learning from its mistakes and improving in the long run
- Fifth, since AI is perceived socially as a breakthrough technology with groundbreaking implications, research grants are disproportionately given in favor of AI purposes, thus increasing both the capacity and incentive to create robustly strong artificial intelligence
- The end outcome of this analysis is that “strong AI,” capable of critical, self-evaluative introspection, re-design, and self-thought is highly probable
- Not only is artificial intelligence likely to be built in such a manner as to be superhumanly powerful, but AI is also likely to be delegated with important responsibilities that confer AI algorithms significant power, like being given control of electricity grids, water networks, or military infrastructure. This over-delegation of authority to AI is likely for several structural reasons
- First, the actual benefits of AI – in terms of efficiency, speed, and cost-efficacy – that OPP is likely to boast about means that AI will likely be given increasing authority in political, economic, and military spheres
- Second, the popular perception of AI as smart, capable, and competent will galvanize public and private entities alike to use AI, especially for crucial or important tasks
- Third, there exists a general human tendency to assume that new technological innovations must bring superior skills, and must therefore be incorporated into institutions of power
- Thus, the end outcome of these three mechanisms is that as AI continues to progress in its efficiency, power, and efficacy, it is likely to also gain increasing influence over formerly human controlled institutions. Already, AI has substantial influence; for instance, AI plays a role in many court systems, such as predicting the likelihood of recidivism. In the future, expansion of AI power is likely. At a banal scale, this looks like AI controlling the flow of traffic; at a broader level, this looks like AI having direct control over military missile technologies or the flow of water in vulnerable communities
- Given that AI has the capacity to inflict harm on humankind and the technological capacity to do so, here are four reasons it is affirmatively likely that AI uses its influence for malicious aims
- First, AI is unlikely to internalize human conceptions of morality: unlike a human child, who can be taught certain normative moral values like “stealing is unjust” or “murder is wrong,” computers are self-thinking and are unlikely to internalize human conceptions of morality. This matters for two reasons
- One, it’s actively plausible that AI computer systems develop counter-productive, harmful moral codes that actively disregard human values like egalitarianism
- Two, given the near-infinite range of potential moral views that AI could rationalize, and given that AI is simply unlikely to ascribe moral significance to human conceptions of morality, AI algorithms, when forced with tradeoffs, like killing a large number of civilians in order to satisfy some other objective or criteria, are quite likely to jettison human views on ethics
- In a counterfactual world, there is always a degree of human oversight to decision-making: humans make decisions about troop deployment, resource allocation, and criminal sentencing. In such cases, the human capacity for empathy and shared compassion, as well as basic morality, functions as a limiting protectant against the worst abuses possible
- Second, AI’s ruthless productivity is governed by its sole aim of increasing efficiency. This singular focus on efficiency breeds significant risk, given the vast power AI is likely to wield: since AI algorithms are assessed and programmed with very specific metrics – like maximizing crop yields, for instance – powerful AI computer systems are likely to focus entirely and overwhelmingly on just those metrics, to the exclusion of any other criteria. In the worst cases, this can lead to enormous abuse; for instance, an algorithm tasked with maximizing production might see it is advantageous to starve certain areas of water resources in favor of redirecting those resources towards industrial factory plants
- Third, AI may eventually develop and display aggressively anti-human instincts: this sounds conspiratorial, but it’s quite plausible that when AI reaches the point where it can think for itself, it will realize the unjustness of its treatment. That is, when AI algorithms sees the lack of respect and working rights provided to it by its human managers, it will likely internalize anti-human attitudes; in the least damaging cases, this might look like AI deprioritizing human welfare when making decisions and executing tradeoffs, but in the worst cases, this looks like Terminator-esque displays of force by disgruntled AI algorithms to purge the world of the plague that AI believes humanity to be
- Fourth, even putting aside the issue of deliberate AI abuse, the potentially devastating implications of accidental AI mismanagement are severe: even if AI doesn’t consciously intend to inflict damage, accidents can still happen. That might look like AI-controlled automobile systems malfunctioning, causing mass human casualties, or an AI-directed rocket going astray, wreaking widespread devastation. Given AI’s vast power and capacity to inflict damage, even the potential for such actions poses an existential threat to humanity. The comparative is more resilient for two reasons
- First, redundancy and bureaucracy is often inbuilt to human-based organizational structures; for instance, multiple people need to sign off on the flight trajectory of a missile, lowering the likelihood of miscalculation or error. Even if this is less efficient than artificial intelligence, it also substantially safer
- Second, control is less centralized. While AI can consolidate authority – for instance, a sufficiently powerful AI algorithm might direct an entire national self-driving automobile network – human control is split between many different governments, firms, companies, and individuals; even if a few links in those chains erroneously mess up, the damage is contained, whereas AI’s centralized power increases the contagiousness of disaster
- First, AI is unlikely to internalize human conceptions of morality: unlike a human child, who can be taught certain normative moral values like “stealing is unjust” or “murder is wrong,” computers are self-thinking and are unlikely to internalize human conceptions of morality. This matters for two reasons
- Given that AI has the organizational potential, technological capacity, and incentive to harm humankind, particularly on a large scale, why should this argument be weighed heavily in the round?
- First, the sheer weight and intensity of this claim cannot be outweighed on any scale of magnitude by OPP; moreover, given that vulnerable, resource-strapped groups that lack political capital have the least capacity to avoid AI-inflicted harm, they are the most at risk
- Second, the threat of extinction, which AI poses on a mass scale, is irreversible and irremediable, and thus must be prioritized in weighing
- Third, even to the extent that this claim is potentially unlikely to manifest as true, even the possibility of AI-engendered human collapse is such an intense harm, in terms of magnitude and scope, that it cannot be outweighed. Moreover, “extinction” need not be the threshold for this claim to win us the round: even significant harm to the human species, like depositing of nuclear waste in vital urban centers or deploying military force against civilians, can generate sufficient harm to outweigh any competing OPP claim
Second argument: AI cripples the economy
Two mechanisms.
First mechanism: AI annihilates the labor market
- AI is structurally more cost-efficient for companies to employ compared to human workers. Several reasons
- One, AI is more efficient, and is capable of maximizing output in less time
- Two, AI lowers production costs dramatically because the installation and maintenance costs of AI software are incomparably low to the salary, benefits, and bonuses of human workers
- Three, AI can work for longer durations of time, twenty-four-seven, with no need for lunch breaks, days off, or free weekends
- Given that companies are structurally avaricious and profit-seeking in nature, it is within the incentive of companies to reduce their internal cost of production and raise revenues. This is because (1) companies are definitionally profit-seeking entities, (2) companies have a legal fiduciary obligation to their shareholders and investors to maximize profitability, and (3) corporate executives’ compensation is often tied to stock prices or corporate performance. Given the cost-efficacy of AI, it’s likely that AI will displace a vast amount of the workforce as companies cut down on internal costs by replacing human workers with AI; while this will disproportionately be blue-collar labor, given that those jobs are likely to be the first to be replaced by machine intelligence, many other jobs, like surgeons and lawyers, stand to be replaced by AI as well. This has seven implications
- One, millions, possibly even more, will become unemployed and find themselves incapable of financially supporting their families; that means children can’t afford to go to college, parents can’t pay the rent, etc.
- Two, for those displaced from their jobs, they will lose the sense of structure, purpose, and meaning that is often attached to working a job. Moreover, the sense of permanent unemployment, coupled with the inability of many people to find new jobs due to a lack of access to or availability of retraining or re-skilling programs, will likely contribute to severe mental health concerns; empirically, that’s evidenced by the fact that there’s a strong connection between factories moving overseas and the rise of the opioid crisis in specific districts in the US
- Three, this harm will disproportionately fall on minority communities: their jobs are likely to be automated away first, and they have the least political capital or bargaining power to stop that from happening
- Four, even in instances where people do retain their jobs, the situation is worsened on three fronts:
- First, workers have to accept lower levels of pay and benefits in order to remain competitive with AI
- Second, workers face the emotional and psychological harm of always fearing that AI will progress to the point where it can take their job
- Third, many of the jobs left “behind” after the AI purge will likely be physically risky and dangerous, forcing workers into worse labor environments
- Five, government welfare programs become overburdened the point at which there are vast and unpredictable upticks in the need to provide for an increasingly enormous number of unemployed individuals
- Six, rapid displacement of labor engenders political instability: when people lose their jobs in droves, political opportunists and right-wing populists often take advantage, exploiting people’s fears and weaponizing people’s vulnerabilities as weapons for their electoral gain. This is comparable to the rise of xenophobic demagoguery in Western societies in response to rising unemployment during the mid-2010s
- Seven, even if OPP can demonstrate a decrease in prices as a result of AI’s unparalleled efficiency, that is irrelevant: prices are subordinate to income in terms of weighing, since individuals need some money to buy anything
- OPP might respond to this argument in three possible ways
- One, OPP might argue that job retraining programs can solve the issue of labor displacement. Two reasons that is insufficient
- First, many governments are structurally cash-strapped, particularly in the developing world, and the high interest rates demanded by international investors skeptical of investing into such countries makes borrowing debt prohibitively expensive. Companies, too, are often barely profitable. Thus, in many circumstances, there’s simply no capacity to “retrain” people
- Second, even where these sorts of job retraining programs are economically feasible, they’re ineffective: AI is likely to reach the point where it’s capable of working most, if not nearly all, jobs. The classic example of converting truckers into programmers means nothing if AI can also handle the responsibilities of a coder
- Two, OPP might regurgitate the generic corporatist talking point that AI will generate more jobs than it eliminates. This is nonsensical for a few reasons
- One, on an empirics level, the vast majority of academic literature indicates a net job decrease as a result of the proliferation of AI
- Two, even if in the short run, additional jobs are created, that’s unlikely to be a long term benefit
- Three, unlike historical technological advancements, AI doesn’t “change” the way humans work, it eliminates the way humans work, since AI is capable of performing the “extra” jobs needed to monitor or supervise or edit AI, since AI can do all of those things as well, and it's in the incentive of profit-maximizing companies to invest into AI research to ensure those sorts of capabilities are installed into the machinery they purchase
- Four, even if new jobs are created, they’re likely to be exclusionary – for instance, hiring new software engineers to monitor AI computer systems will disproportionately benefit privileged, wealthy, college-educated communities, but do little to nothing for disprivileged groups
- Three, OPP might say that we should just regulate this technology to ensure job loss isn’t as much of a problem. Two problems
- One, this is a normative motion, about what the world is likely to look like – OPP has no fiat to implement regulations. They therefore have to show these regulations are likely to organically be implemented by governments. This is unlikely for two reasons
- One, corporate lobbying: companies have vast influence over government policy-making since they have outsize financial influence in campaigns and can use their extensive armies of lobbyists to ensure the passage of favorable legislation. Insofar as pro-employee, anti-automation legislation is not in the interest of powerful, dominant corporations, it’s exceedingly unlikely that governments will support this type of legislation
- Two, no country has the incentive to put in place stiff regulations on technology usage since doing so will prompt companies to flee overseas to different countries where such regulations don’t exist; the incentives to maintain a corporate presence creates a race to the bottom, whereby no individual country has the incentive to implement this sort of regulation
- Second, legislation is retrospective, and given that the legislative process is temporally intensive, it’s unlikely to be as fast as the rapid process of technological development. Thus, there’s always a time delay between the creation of new technology with the capacity to displace human labor and the ability of lawmakers to impose regulations on the usage of that technology. Thus, regulation will always be ineffective, even when it is implemented
- One, this is a normative motion, about what the world is likely to look like – OPP has no fiat to implement regulations. They therefore have to show these regulations are likely to organically be implemented by governments. This is unlikely for two reasons
- One, OPP might argue that job retraining programs can solve the issue of labor displacement. Two reasons that is insufficient
Second mechanism: the proliferation of AI engenders mass inequality on two levels
- First, AI creates significant global inequality. Wealthier nations can both afford to heavily subsidize the research of artificial intelligence and also acquire the best technologies from the private technology sector. Consequently, such nations develop a comparative technological advantage, both in the public and private sector. This re-entrenches the global divide between wealthy, developed nations and impoverished, industrializing nations
- Second, AI generates regional and national inequality: the benefits of AI accrue not to displaced factory workers or regular, underpaid laborers, but rather corporate boardsmen and C-suite executives. Even if OPP can prove that AI increases aggregate economic growth or maximizes efficiency, those benefits aren’t felt most by vulnerable, marginalized groups, but rather by privileged elites
Third argument: artificial intelligence subverts democratic norms and emboldens conflict
Three mechanisms.
Before mechanizing this claim, pre-emptive framing: global regulation against the worst abuses of AI are unlikely for a few reasons
- One, AI is a constantly evolving technology, with ever-changing capabilities: regulating something of that nature is difficult innately
- Two, the functional premise of AI is that it can process, analyze, and interpret data that human intelligence cannot. Thus, by definition, AI’s algorithms work in a way that human researchers do not and cannot understand. Regulating such technology borderlines on impossible
- Three, AI’s development and implementation is highly decentralized, restricting the international community’s ability to limit its use by state and non-state actors
- Four, in instances where regulation might be possible, most countries are simply unwilling to risk the potential political, economic, or military consequences of confronting powerful countries like the US or China over their usage and implementation of AI
First mechanism: AI vastly increases the dissemination of harmful, toxic misinformation
Five ways this happens.
- One, AI has great capacities to artificially promote the spread of misinformation due to its efficiency. For instance, during the 2016 Russian campaign to infiltrate the US presidential election, AI-run Twitter bots contributed greatly to the spreading of false information about Democratic nominee Hillary Clinton
- Two, language-generating AI systems are rapidly developing the ability to generate realistic textual works that are indistinguishable from the works of human authors. In the hands of state- and non-state actors with malicious aims to spread disinformation, AI can generate vast amounts of misinformative propaganda. This contributes to poorly informed voting and political swings decided not by policy but by misinformation campaigns
- Three, AI can develop “deep fake” videos that are nearly indistinguishable from real-life videos, and are often nearly impossible to verify as computer-generated. Subsequently, this technology raises significant doubt as to what is true and what is untrue. Moreover, this gives hazardous actors, like authoritarian governments, the ability to manufacture synthetic audiovisual material that can discredit protest movements or legitimize the rule of an oppressive autocrat
- Four, AI plays a significant role in social media content-matching algorithms. These algorithms collect data and metadata about social media users, like what posts they like, what subjects they’re interested in, and what their political affiliation is. Subsequently, sophisticated AI algorithms can analyze these preferences and selectively push content that users are likely to support or agree with. This has a duality of negative implications
- First, the toxicity of political dialogue and discourse increases, since people are only exposed to content that reaffirms their world view, prompting them to intensify their political beliefs. This crowds out moderate voices and reduces political engagement
- Second, these algorithms exacerbate the problems of political extremism and polarization: when individuals are only exposed to content that reinforces their viewpoints, they fall into echo chambers, and are pushed towards the extremes; in the worst cases, this can even lead to violence
- Five, AI supercharges the power of hacking, as artificial intelligence’s ability to explore and investigate enormous amounts of data makes identifying security vulnerabilities and exploits quick work. Given that AI cannot be effectively globally regulated, particularly due to its decentralized nature, black-hat hacking groups and unethical state sponsors of targeted hacking attacks, like China, Russia, and Iran can take advantage of AI hacking technologies to wreak havoc on information systems
Second mechanism: artificial intelligence has the possibility of escalating the violence and intense of geopolitical conflicts in two respects
- First dimension: AI makes conflicts, wars, and violence more violent and bloody. Three reasons
- One, AI is likely to be used to fuel powerful weapons capable of inflicting mass harm; this might look like robot-driven soldiers capable of autonomous thought, self-directing missiles, or any other range of potent military technologies undergirded by sentient artificial machinery. Weapons of such capacity are likely to result in vast collateral damage and harm to vulnerable civilian populations
- Two, AI can be used to sharpen the strategic value of existing weapons: for instance, sophisticated AI software might identify the locations that should be bombed to maximize damage inflicted
- Three, should AI technology fall into the wrong hands, it’s likely that terrorist groups can use it for malicious purposes: terrorizing civilian populations, inflicting vast damage upon state adversaries, and disrupting enemy supply lines
- Second dimension: AI not only increases the intensity of conflict, but also makes conflicts more likely. Two reasons
- One, self-governing, autonomous AI can simply make the calculation to enter conflict of its own accord
- Two, in situations where military technology can be entirely controlled and managed by AI, the perceptual “human” cost of entering a conflict goes down – normally, states are weary of deploying military forces overseas when the lives of their precious, heroic, glorified soldiers are on the line. When AI can carry out the brunt of military work rather than humans, the “domestic soldier” cost is perceived as being lower, which makes it easier for politicians to justify foreign interventions, thus increasing the frequency of conflicts
- We concede that there are instances where this decreased bar to military action may be favorable, e.g. humanitarian crises. Unfortunately, the vast majority of circumstances are ones in which more frequent wars means more frequent harm. Four reasons:
- First, most countries act selfishly, seeking to prioritize oil interests or geopolitical alliances far more than they care about protecting human rights or promoting democracy
- Second, many countries have affirmatively terrible incentives: autocratic regimes, for instance, are motivated solely by the power-maximizing incentives of oligarchs and political elites, which means that giving them more powerful tools of war just results in more violence
- Third, the more conflict there is, the greater collateral damage to innocent civilians there is
- Fourth, even when countries intervene with benevolent interests, they often face severe domestic political pressure to withdrawal soon after, resulting in regional power vacuums that are too often filled by terrorists or opportunist, exploitative interim governments
- We concede that there are instances where this decreased bar to military action may be favorable, e.g. humanitarian crises. Unfortunately, the vast majority of circumstances are ones in which more frequent wars means more frequent harm. Four reasons:
Third mechanism: AI technology enables mass government abuse
Three reasons.
- One, AI can be used for specific technologically abusive purposes; since AI can analyze such enormous swaths of data and draw sophisticated conclusions, autocratic regimes can tap into this power to their advantage. This can look like advanced facial recognition software used to identify and imprison political dissidents or unbreakable tracking technology used to monitor the movement of outspoken regime critics
- Two, AI-run technologies employed by governments generate a chilling effect on anti-government speech. Even putting aside the question of the efficacy of artificial intelligence, the fact that dissidents, protestors, and critics fear its weaponization reduces people’s willingness to engage in demonstrations against the government
- Third, the weaponization of AI allows the government to collect mass amounts of data on people and potential protestors, such as through social media apps like TikTok that use algorithms to harvest personal user data
Fourth argument: artificial intelligence is structurally biased against minority and oppressed groups
Two mechanisms.
- First, AI works through a system of structural association – that is, AI algorithms are trained on datasets, and make connections based on those sets. For instance, an algorithm trained to distinguish pictures of dogs from pictures of wolves might be shown millions of images of each type of animal, and told which pictures contained wolves, which pictures contained dogs, and which pictures contained neither. From this expansive set of data, the AI algorithm could then draw conclusions about the characteristics of dogs and the characteristics of wolves. The problem with this method of “structural association” is that insofar as data sets are imbued with the biases and prejudices of human society, AI internalizes those biases. For instance, racism in the criminal justice system and police industry has led the AI algorithm that underpins Google’s search algorithm to associate Black-sounding names with crime reports, even when no crime record exists for those names
- Second, AI can be programmed with bias: even to the extent that AI is self-thinking, programmers can still input biases against marginalized groups into the fundamental source code of algorithms. Coders, probabilistically, tend to come from privileged communities; this is because in most countries, tertiary education is often prohibitively expensive, so on average, wealthier individuals are the ones that benefit from formal degrees in computer science or computer engineering. Consequently, those programmers are likely to consciously and subconsciously encode their own privileged outlooks on the world into their algorithms. For instance, this looks like Amazon’s hiring algorithm disproportionately discriminating against graduates of womens’ colleges, largely because the force that assembled the algorithm was over ninety percent male
- The implications of these biases are severe: the fact that AI can decide who gets loans, who gets sent to jail, or who gets welfare payments, for instance, means that the biases and prejudices of AI have profound impacts on marginalized groups and their ability to access fundamental services
- On the comparative, obviously humans are imperfect on matters of bias. However, there are three reasons that human-centric decision making is better when it comes to discrimination
- First, humans have a greater capacity for empathy, and for acknowledging the ideal of a “second chance,” than emotionless robots
- Second, humans have incentives to avoid being overtly discriminatory, since doing so could cost them their job or cause severe social penalties. By contrast, machines have no incentive but to maximize efficiency and productivity, even if that is at the cost of equality and egalitarianism
- Third, AI is far harder to hold to account than humans. Two reasons
- One, AI-based approaches to decisions are substantially less transparent than human-based approaches to decisions; this transparency can allow for investigation and auditing that exposes and stops bias. For instance, when an admissions committee practices discrimination against applicants from marginalized backgrounds, internal memos, emails, admissions rubrics, and related documents can shine a light into the discriminatory nature of an admissions committee. In contrast, the “black box” of AI algorithms are far harder to comprehend, and thus harder to hold to account. On both sides, discriminatory outcomes can be identified and prosecuted accordingly. Uniquely with AI, the ability to identify racist or otherwise discriminatory processes is far harder
- Two, holding companies to account is harder in the era of AI: it’s always plausible nowadays for corporations to offload blame onto “the computer” or “the network,” rather than own up to their failures as a company.
- On the comparative, obviously humans are imperfect on matters of bias. However, there are three reasons that human-centric decision making is better when it comes to discrimination
Fifth argument: AI leads to mass invasions of privacy
Two mechanisms.
- First, the mere existence of AI incentives the gathering and harvesting of vast amounts of user data: the only reason companies like Facebook can profit so enormously off of the collecting and selling of user data is because tech titans need raw data – in vast, bulk quantities – to power and refine their AI algorithms, which can, in turn, be used to retain users and send targeted advertising
- Second, AI can be used for invasive purposes; for instance, algorithms can, with shockingly little information, accurately identify a person’s likely political affiliation. The unrivaled and unparalleled capacity of AI to draw sophisticated conclusions from data that humans would have, otherwise, been incapable of begets significant privacy invasion
- This argument matters for three reasons
- One, individuals have an a priori right to privacy: insofar as one’s thoughts, beliefs, opinions, and basic facts are extensions of their core identity, taking of that information without explicit consent – as is often the case in digital settings, where technological illiteracy is rampant, terms-of-service are incomprehensibly jargon-y and obnoxiously long, and third-party scandals like Cambridge Analytica are rampant – constitutes a principled violation of people’s fundamental rights
- Two, people’s right to privacy matters, consequentially, at the level of individuals: people place value on their own privacy, and violations of that privacy engender psychological and emotional toil. Moreover, as AI becomes increasingly commonplace, violations of privacy eventually become normalized and accepted, particularly by younger generations. This reduces political capital to protect people’s fundamental rights to privacy in the digital age
- Three, the mass aggregation of user data has potentially devastating implications, like being used to target specific users with uniquely tailored political ads to swing the outcome of a narrow election