By: Alexandra Manfield

Published: April 14, 2026

Artificial intelligence (“AI”) is often described as objective, neutral, and data-driven—these notions are misleading. ^[1]  AI systems do not emerge in a vacuum; they are designed, trained, and deployed by humans who operate within legal, cultural, and institutional frameworks shaped by racial hierarchy. ^[2] As a result, AI often replicates structural racism by embedding inequality in datasets, implementing racist policies, and presenting bias as an unavoidable technical fact. ^[3] In critical areas like health care, policing, and digital media, the false idea of AI neutrality has real impacts on communities of color. ^[4]

This problem has taken on new constitutional significance. In Moody v. NetChoice, the Supreme Court recognized that AI algorithms produce expressive content and extended First Amendment protections to algorithmic content curation.[5] The Court explained that algorithms function not merely as technical tools, but also as expressive mechanisms that generate and organize content in ways that reflect editorial judgment.[6] Accordingly, the Court held that a platform’s algorithmic content curation is protected by the First Amendment, reasoning that choices about what to display or suppress convey the platform’s own expressive viewpoint.[7]

Statutory law, on the other hand, treats technology as a tool rather than an independent source of structural influence.[8] Section 230 of the Communications Decency Act,[9] enacted to promote the growth of the early internet, shields online platforms from liability for third-party content and grants them discretion to moderate material.[10] The statutory language naturally reads to protect internet companies in two ways: (1) when they unknowingly fail to edit or remove harmful third-party content and (2) when they decide to exercise their editorial functions in good faith.[11] Courts have often applied Section 230 broadly, extending its protections to algorithmic recommendation systems on the theory that these systems merely facilitate the display of user-generated content.[12] In doing so, the statute has functioned as a foundational rule governing online technologies; however, this framework was developed long before the emergence of large-scale AI systems.

AI systems learn from data which reflects the world as it is—not as it should be. Historical and contemporary datasets are shaped by unequal access to health care, discriminatory policing practices, racially segregated housing, and exclusionary labor markets.[13] When AI models are trained on this information, they do not simply identify neutral patterns; they internalize and scale existing inequities. Machine-learning systems operate through probabilistic inference rather than fixed rules, producing outputs that appear objective because no individual human decision is visible.[14] Yet, bias enters these systems at multiple points: during dataset construction, data labeling, model tuning, and human feedback.[15] Once embedded, these patterns are difficult to trace or audit, giving rise to the so-called “black box” problem.[16]

Critical Race Theory (“CRT”) offers a valuable perspective on why this illusion continues. CRT highlights that systems can appear neutral yet produce racially unequal outcomes because inequality is embedded in institutional structures.[17] The effects of algorithmic bias are not just hypothetical; they are extensively documented across various sectors. Facial recognition technology is a prime example. Research consistently shows these systems misidentify Black individuals more often than white ones, mainly because training datasets overrepresent lighter-skinned faces.[18] In law enforcement, such errors lead to wrongful surveillance, false arrests, and increased criminalization of communities of color—reinforcing decades of racially biased policing.[19] Predictive policing models rely on past crime data that is influenced by racially biased policing, not actual crime rates.[20] When these algorithms categorize neighborhoods as “high risk,” they’re not detecting new threats, but perpetuating harmful stereotypes that associate communities of color with criminality, leading to feedback loops of over-policing.[21]

Similar issues arise with health care algorithms. Diagnostic algorithms often fail to accurately detect conditions in Black patients because they were trained on predominantly white datasets.[22] Healthcare algorithms that underestimate risk for Black patients mirror deep-rooted narratives that downplay or misrepresent Black illness and pain.[23] A common example of this disparity is seen in a typical AI system that dermatologists use to detect skin cancer in patients.[24] The specialized type of AI system, called a convolutional neural network (“CNN”), is often trained on datasets predominantly composed of white patients.[25] Black patients’ skin lesions may differ from those of white patients, making automated diagnosis less accurate.[26] This is concerning because Black patients have the highest melanoma mortality, with only 70% of patients with a five-year survival rate, compared to 94% in white patients.[27] Misdiagnoses and health care barriers hinder treatment and lead to advanced skin cancer.[28]

In the media industry, content moderation systems frequently suppress the speech of marginalized groups while simultaneously amplifying harmful stereotypes or misinformation.[29] Whether content is evaluated by a human or an AI algorithm, bias can enter at many points in a platform’s content review process.[30] One study revealed that models for automatic hate speech detection were 1.5 times more likely to mark tweets by self-identified Black people as offensive or hateful.[31] Tweets using African American English were over twice as likely to be labeled as “offensive” or “abusive.”[32] Another study examined five commonly used datasets for hate speech research and discovered a “consistent, systemic, and significant racial bias in classifiers trained on all five datasets.”[33] The study reported that in “almost every case, black-aligned tweets are classified as sexism, hate speech, harassment, and abuse at higher rates than white-aligned tweets.”[34] These industry examples demonstrate that AI does not serve to challenge existing inequalities; instead, it tends to reinforce them under the guise of technical legitimacy.[35]

This situation becomes especially dangerous when algorithmic outputs are protected by constitutional law, thereby heightening the ongoing tension between technological progress, corporate influence, and racial hierarchies present in AI systems. The Moody Court highlighted that decisions about what content to promote or suppress reflect a platform’s viewpoint, even when made through automated processes.[36] While the ruling focused on state regulation of online platforms, its implications go well beyond content moderation. Regarding algorithmic outputs as protected speech simply because they are expressive, while also implementing policies that treat AI technology as neutral, enables legal institutions to overlook the fundamental roots of algorithmic harm.[37] This leads to mischaracterizing discrimination as merely incidental, instead of recognizing it as a systemic issue.[38]

When the products of biased algorithms are regarded as neutral—or worse, as protected speech—the law legitimizes discrimination. As AI becomes more deeply embedded in law, medicine, policing, and public discourse, the stakes of that choice grow higher. A more equitable legal framework must prioritize transparency and accountability and recognize that AI is a product of social conditions—not an escape from them.

 

 

[1] Margret Hu, Critical Data Theory, 65 Wm. & Mary L. Rev. 839, 861 (2024).

[2] Stefan Milne, Q&A: New AI Training Method Lets Systems Better Adjust to Users’ Values, U. Wash. News (Dec. 18, 2024), https://www.washington.edu/news/2024/12/18/ai-user-values-preferences-rlhf/; Harini Suresh & John Guttag, A Framework for Understanding Sources of Harm Throughout the Machine Learning Life Cycle, Ass’n for Computing Mach., Nov. 2021, at 5-6.

[3] Milne, supra note 2.

[4] Milne, supra note 2.

[5] Moody v. NetChoice, LLC, 603 U.S. 707, 716 (2024).

[6] Id. at 730-31; see also Turner Broad. Sys., Inc. v. FCC, 520 U.S. 180, 185 (1997) (holding a private party’s collection of third-party content into a single speech product is itself expressive).

[7] Moody, 603 U.S. at 743.

[8] Hu, supra note 1, at 863 (noting policymakers encourage the growth of systematic decision-making on the basis that it is inherently non-discriminatory).

[9] 47 U.S.C. § 230.

[10] Id.; Universal Commc’n Sys., Inc. v. Lycos, Inc., 478 F.3d 413, 422 (1st Cir. 2007) (finding § 230 protects a platform from any cause of action implicating the platform’s decisions with respect to third-party posts, and for its inherent decisions about how to treat posts).

[11] 47 U.S.C. §§ 230(c)(1)-(2) (2018); see Malwarebytes, Inc. v. Enigma Software Grp. USA, 141 S. Ct. 13, 14-15 (2020) (Thomas, J., concurring in denial of certiorari) (explaining courts broadly interpret § 230); cf. Barnes v. Yahoo!, Inc., 570 F.3d 1096, 1102-03 (9th Cir. 2009) (recognizing the statute encourages websites to actively police themselves, not offer an excuse for inaction).

[12] See, e.g., Universal Commc’n Sys., Inc., 478 F.3d at 422 (finding § 230 immunity should be broadly construed); Malwarebytes, 141 S. Ct. at 16 (Thomas, J. concurring in denial of certiorari) (criticizing current overly broad § 230 interpretations).

[13] Suresh & Guttag, supra note 2, at 4-6.

[14] Tanner Kohler, How AI Models are Trained, Nielsen Norman Grp. (May 2, 2025), https://www.nngroup.com/articles/ai-model-training/ (explaining the model learns by trying to predict the next “token” in a sequence).

[15] Id.

[16] Joshua A. Kroll et al., Accountable Algorithms, 165 U. Penn. L. Rev. 633, 633-34 (2017) (explaining the “black box” where AI considers only the inputs and outputs of the system, not how the AI system reached its conclusion).

[17] Hu, supra note 1, at 845-46.

[18] Sonia M. Gipson Rankin, Technological Tethereds: Potential Impact of Untrustworthy Artificial Intelligence in Criminal Justice Risk Assessment Instruments, 78 Wash. & Lee L. Rev. 647, 651 (2021).

[19] Id. at 650-52; Suresh & Guttag, supra note 2, at 5-6.

[20] Suresh & Guttag, supra note 2, at 6.

[21] See id. (noting Northpointe’s COMPAS had significantly higher false positive rates for Black defendants versus white defendants); Larry Hardesty, Study Finds Gender and Skin-Type Bias in Commercial Artificial-Intelligence Systems, MIT News (Feb. 11, 2018), https://news.mit.edu/2018/study-finds-gender-skin-type-bias-artificial-intelligence-systems-0212.

[22] Natalia Norori et al., Addressing Bias in Big Data and AI for Health Care: A Call for Open Science, 2 Patterns, Oct. 2021, at 2 (“[W]hen tested with images of Black patients, the networks have approximately half the diagnostic accuracy compared with what their creators originally claimed”).

[23] Id. (noting diagnostic algorithms for diseases with a genetic components like skin cancer that are only trained on data from white patients may fail to generalize patients of other ethnicities); see also Chaz Arnett, Dystopian Dreams, Utopian Nightmares: AI and the Permanence of Racism, 112 Georgetown L.J. 1299, 1326 (2024) (explaining the need for a framework to amplify the voices and power of the communities most vulnerable to AI harm).

[24] Norori et al., supra note 22, at 2.

[25] Id. (stating the proportion of Black patients in the dataset is typically around 5% to 10%).

[26] Id.

[27] Id.

[28] Id.

[29] Ángel Díaz & Laura Hecht-Felella, Double Standards in Social Media Content Moderation, Brennan Ctr. For Just. 10 (Aug. 4, 2021), https://www.brennancenter.org/our-work/research-reports/double-standards-social-media-content-moderation.

[30] Id.

[31] See id. at 11 (citing Maarten Sap et al., “The Risk of Racial Bias in Hate Speech Detection,” Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, July 28-Aug. 2, 2019, 1671, https://www.aclweb.org/anthology/P19-1163.pdf).

[32] Id.

[33] Id. (citing Thomas Davidson et al., “Racial Bias in Hate Speech and Abusive Language Detection Datasets,” Proceedings of the Third Abusive Language Workshop at the Annual Meeting for the Association for Computational Linguistics, Florence, Italy, Aug. 1-2, 2019, 6, https://arxiv.org/pdf/1905.12516.pdf).

[34] Id. (quoting Davidson et al., “Racial Bias in Hate Speech and Abusive Language Detection Datasets,” at 6).

[35] Suresh & Guttag, supra note 16, at 5 (explaining historical biases in AI).

[36] Moody v. NetChoice, LLC, 603 U.S. 707, 716-17 (2024).

[37] Hu, supra note 1, at 850-51.

[38] Id.