Skip to main content Skip to secondary navigation

SQ10. What are the most pressing dangers of AI?

Main content start

As AI systems prove to be increasingly beneficial in real-world applications, they have broadened their reach, causing risks of misuse, overuse, and explicit abuse to proliferate. As AI systems increase in capability and as they are integrated more fully into societal infrastructure, the implications of losing meaningful control over them become more concerning.1 New research efforts are aimed at re-conceptualizing the foundations of the field to make AI systems less reliant on explicit, and easily misspecified, objectives.2 A particularly visible danger is that AI can make it easier to build machines that can spy and even kill at scale. But there are many other important and subtler dangers at present.


One of the most pressing dangers of AI is techno-solutionism, the view that AI can be seen as a panacea when it is merely a tool.3 As we see more AI advances, the temptation to apply AI decision-making to all societal problems increases. But technology often creates larger problems in the process of solving smaller ones. For example, systems that streamline and automate the application of social services can quickly become rigid and deny access to migrants or others who fall between the cracks.4

When given the choice between algorithms and humans, some believe algorithms will always be the less-biased choice. Yet, in 2018, Amazon found it necessary to discard a proprietary recruiting tool because the historical data it was trained on resulted in a system that was systematically biased against women.5 Automated decision-making can often serve to replicate, exacerbate, and even magnify the same bias we wish it would remedy.

Indeed, far from being a cure-all, technology can actually create feedback loops that worsen discrimination. Recommendation algorithms, like Google’s page rank, are trained to identify and prioritize the most “relevant” items based on how other users engage with them. As biased users feed the algorithm biased information, it responds with more bias, which informs users’ understandings and deepens their bias, and so on.6 Because all technology is the product of a biased system,7 techno-solutionism’s flaws run deep:8 a creation is limited by the limitations of its creator.

Dangers of Adopting a Statistical Perspective on Justice

Automated decision-making may produce skewed results that replicate and amplify existing biases. A potential danger, then, is when the public accepts AI-derived conclusions as certainties. This determinist approach to AI decision-making can have dire implications in both criminal and healthcare settings. AI-driven approaches like PredPol, software originally developed by the Los Angeles Police Department and UCLA that purports to help protect one in 33 US citizens,9 predict when, where, and how crime will occur. A 2016 case study of a US city noted that the approach disproportionately projected crimes in areas with higher populations of non-white and low-income residents.10 When datasets disproportionately represents the lower power members of society, flagrant discrimination is a likely result.

Sentencing decisions are increasingly decided by proprietary algorithms that attempt to assess whether a defendant will commit future crimes, leading to concerns that justice is being outsourced to software.11 As AI becomes increasingly capable of analyzing more and more factors that may correlate with a defendant's perceived risk, courts and society at large may mistake an algorithmic probability for fact. This dangerous reality means that an algorithmic estimate of an individual’s risk to society may be interpreted by others as a near certainty—a misleading outcome even the original tool designers warned against. Even though a statistically driven AI system could be built to report a degree of credence along with every prediction,12 there’s no guarantee that the people using these predictions will make intelligent use of them. Taking probability for certainty means that the past will always dictate the future.

An original image of low resolution and the resulting image of high resolution
Image-generation GANs can be used to perform other tasks like translating low-resolution images of faces into high resolution images of faces. Of course, such a transformation is not recovering missing information so much as it is confabulating details that are consistent with its input. As an example, the PULSE system tends to generate images with features that appear ethnically white, as seen in this input image of former US President Barack Obama. From:…

There is an aura of neutrality and impartiality associated with AI decision-making in some corners of the public consciousness, resulting in systems being accepted as objective even though they may be the result of biased historical decisions or even blatant discrimination. All data insights rely on some measure of interpretation. As a concrete example, an audit of a resume-screening tool found that the two main factors it associated most strongly with positive future job performance were whether the applicant was named Jared, and whether he played high school lacrosse.13 Undesirable biases can be hidden behind both the opaque nature of the technology used and the use of proxies, nominally innocent attributes that enable a decision that is fundamentally biased. An algorithm fueled by data in which gender, racial, class, and ableist biases are pervasive can effectively reinforce these biases without ever explicitly identifying them in the code. 

Without transparency concerning either the data or the AI algorithms that interpret it, the public may be left in the dark as to how decisions that materially impact their lives are being made. Lacking adequate information to bring a legal claim, people can lose access to both due process and redress when they feel they have been improperly or erroneously judged by AI systems. Large gaps in case law make applying Title VII—the primary existing legal framework in the US for employment discrimination—to cases of algorithmic discrimination incredibly difficult. These concerns are exacerbated by algorithms that go beyond traditional considerations such as a person’s credit score to instead consider any and all variables correlated to the likelihood that they are a safe investment. A statistically significant correlation has been shown among Europeans between loan risk and whether a person uses a Mac or PC and whether they include their name in their email address—which turn out to be proxies for affluence.14 Companies that use such attributes, even if they do indeed provide improvements in model accuracy, may be breaking the law when these attributes also clearly correlate with a protected class like race. Loss of autonomy can also result from AI-created “information bubbles” that narrowly constrict each individual’s online experience to the point that they are unaware that valid alternative perspectives even exist.

Disinformation and Threat to Democracy

AI systems are being used in the service of disinformation on the internet, giving them the potential to become a threat to democracy and a tool for fascism. From deepfake videos to online bots manipulating public discourse by feigning consensus and spreading fake news,15 there is the danger of AI systems undermining social trust. The technology can be co-opted by criminals, rogue states, ideological extremists, or simply special interest groups, to manipulate people for economic gain or political advantage. Disinformation poses serious threats to society, as it effectively changes and manipulates evidence to create social feedback loops that undermine any sense of objective truth. The debates about what is real quickly evolve into debates about who gets to decide what is real, resulting in renegotiations of power structures that often serve entrenched interests.16

Discrimination and Risk in the Medical Setting

While personalized medicine is a good potential application of AI, there are dangers. Current business models for AI-based health applications tend to focus on building a single system—for example, a deterioration predictor—that can be sold to many buyers. However, these systems often do not generalize beyond their training data. Even differences in how clinical tests are ordered can throw off predictors, and, over time, a system’s accuracy will often degrade as practices change. Clinicians and administrators are not well-equipped to monitor and manage these issues, and insufficient thought given to the human factors of AI integration has led to oscillation between mistrust of the system (ignoring it) and over-reliance on the system (trusting it even when it is wrong), a central concern of the 2016 AI100 report.

These concerns are troubling in general in the high-risk setting that is healthcare, and even more so because marginalized populations—those that already face discrimination from the health system from both structural factors (like lack of access) and scientific factors (like guidelines that were developed from trials on other populations)—may lose even more. Today and in the near future, AI systems built on machine learning are used to determine post-operative personalized pain management plans for some patients and in others to predict the likelihood that an individual will develop breast cancer. AI algorithms are playing a role in decisions concerning distributing organs, vaccines, and other elements of healthcare. Biases in these approaches can have literal life-and-death stakes.

In 2019, the story broke that Optum, a health-services algorithm used to determine which patients may benefit from extra medical care, exhibited fundamental racial biases. The system designers ensured that race was precluded from consideration, but they also asked the algorithm to consider the future cost of a patient to the healthcare system.17 While intended to capture a sense of medical severity, this feature in fact served as a proxy for race: controlling for medical needs, care for Black patients averages $1,800 less per year.

New technologies are being developed every day to treat serious medical issues. A new algorithm trained to identify melanomas was shown to be more accurate than doctors in a recent study, but the potential for the algorithm to be biased against Black patients is significant as the algorithm was trained using majority light-skinned groups.18 The stakes are especially high for melanoma diagnoses, where the five-year survival rate is 17 percentage points less for Black Americans than white. While technology has the potential to generate quicker diagnoses and thus close this survival gap, a machine-learning algorithm is only as good as its data set. An improperly trained algorithm could do more harm than good for patients at risk, missing cancers altogether or generating false positives. As new algorithms saturate the market with promises of medical miracles, losing sight of the biases ingrained in their outcomes could contribute to a loss of human biodiversity, as individuals who are left out of initial data sets are denied adequate care. While the exact long-term effects of algorithms in healthcare are unknown, their potential for bias replication means any advancement they produce for the population in aggregate—from diagnosis to resource distribution—may come at the expense of the most vulnerable.

[1] Brian Christian, The Alignment Problem: Machine Learning and Human Values, W. W. Norton & Company, 2020



[4]; (“Automating inequality”)


[6] Safiya Umoja Noble, Algorithms of Oppression: How Search Engines Reinforce Racism, NYU Press, 2018 

[7] Ruha Benjamin, Race After Technology: Abolitionist Tools for the New Jim Code, Polity, 2019



[10] Kristian Lum and William Isaac, “To predict and serve?” Significance, October 2016,

[11] Jessica M. Eaglin, “Technologically Distorted Conceptions of Punishment,” 

[12] Riccardo Fogliato, Maria De-Arteaga, and Alexandra Chouldechova, “Lessons from the Deployment of an Algorithmic Tool in Child Welfare,” 



[15] Ben Buchanan, Andrew Lohn, Micah Musser, and Katerina Sedova, “Truth, Lies, and Automation,” 

[16] Britt Paris and Joan Donovan, “Deepfakes and Cheap Fakes: The Manipulation of Audio and Visual Evidence,” 




Cite This Report

Michael L. Littman, Ifeoma Ajunwa, Guy Berger, Craig Boutilier, Morgan Currie, Finale Doshi-Velez, Gillian Hadfield, Michael C. Horowitz, Charles Isbell, Hiroaki Kitano, Karen Levy, Terah Lyons, Melanie Mitchell, Julie Shah, Steven Sloman, Shannon Vallor, and Toby Walsh. "Gathering Strength, Gathering Storms: The One Hundred Year Study on Artificial Intelligence (AI100) 2021 Study Panel Report." Stanford University, Stanford, CA, September 2021. Doc: Accessed: September 16, 2021.

Report Authors

AI100 Standing Committee and Study Panel 


© 2021 by Stanford University. Gathering Strength, Gathering Storms: The One Hundred Year Study on Artificial Intelligence (AI100) 2021 Study Panel Report is made available under a Creative Commons Attribution-NoDerivatives 4.0 License (International):