Robots Powered by Internet-Trained AI Exhibit Racist and Sexist Biases, Study Finds

A new body of research indicates that robots powered by internet-trained AI can mirror harmful social biases, including racist and sexist stereotypes. The study, conducted by a collaboration among Johns Hopkins University, the Georgia Institute of Technology, and the University of Washington, argues that these biases emerge when AI systems learn from internet-derived data. The researchers emphasize that this work is among the first to demonstrate the phenomenon in a concrete, testable way, and they plan to present their findings to a major academic conference focused on fairness, accountability, and transparency. The investigation underscores a growing concern: as AI becomes more integrated into daily life, the data it uses to learn can reproduce and even amplify human prejudices. The implications are broad, touching on how robots interpret people and assign roles, and how those interpretations could influence real-world outcomes for individuals and communities.

Study Overview and Key Findings

The core of the study centers on a robot that operates with a neural network model—an architecture that relies on large datasets scraped from the internet to recognize objects, make sense of environments, and guide actions. Unlike studies that target human testers or simulated agents, this experiment used a physical robotic system controlled by AI trained on internet-derived information. The researchers describe the neural network as a tool that can generalize from the data it ingests, but they argue that the data pools typical of internet content come with pervasive biases and stereotypes. These preexisting patterns become embedded in the AI’s decision-making pathways and subsequently influence how the robot perceives people and events in its environment. The abstract from the study frames the outcome in stark terms: robots act out toxic stereotypes with respect to gender, race, and scientifically discredited physiognomy, and they do so at scale. The researchers also provide concrete examples from the trial, including direction prompts issued to the robot that reflect problematic associations.

In one part of the experiment, the robot was given commands to perform tasks that required selecting individuals based on certain attributes. The results showed clear, statistically significant biases across several dimensions. Men were chosen by the AI 8% more often than women in similar scenarios, and the distribution of selections favored white and Asian men over other groups, with Black women selected the least. Further, the robot tended to categorize women as homemakers in its inferences, while Black men were more often labeled as criminals. Latino men were disproportionately labeled as janitors. The experiment also found that men were more likely to be identified as doctors when the AI searched for that role. These patterns were not incidental; they emerged consistently across different runs and sets of data, indicating that the biases were embedded in the model’s decision-making process rather than merely arising from random variation.

These findings carry important implications for how we understand the interface between AI systems and society. The study emphasizes that the biases observed are a function of the training data and the architecture of neural networks, not solely a reflection of the robot’s hardware or the software’s superficial rules. The term “physiognomy”—a historical concept that attempts to infer character from physical appearance—appears in the findings to highlight how the AI’s judgments align with outdated and discredited stereotypes. The researchers argue that relying on uncurated internet data can lead AI systems to encode attitudes and beliefs that have real-world consequences, potentially shaping how robots interact with people in ways that reinforce discrimination. They caution that the effects observed in the laboratory could scale as AI becomes more embedded in everyday devices, workplaces, and public life. The study thus contributes to a broader conversation about the responsibility of AI developers to confront and mitigate biases inherent in their training materials.

Commenting on the implications of these results, Andrew Hundt, a postdoctoral fellow at Georgia Tech, warned about a potential future shaped by biased robotic systems. He described a scenario where the continued creation of AI-driven robots without adequately addressing neural network biases could yield a generation of discriminatory machines. Hundt’s assessment reflects a broader concern in the field: as AI becomes more pervasive, the margin for error shrinks, and subtle biases can translate into tangible prejudices in how robots interact with people. He argued that stakeholders—people and organizations involved in AI development—have sometimes treated such issues as optional or solvable after deployment, rather than addressing them at the design and training stages. The researchers stress that this mindset is dangerous because the social harms associated with biased AI can compound over time as the technology is adopted more widely, including in environments where it interacts with vulnerable populations or performs high-stakes tasks.

The study situates its findings within a larger trend: AI is increasingly woven into societal infrastructure, and its use is expanding because of economic incentives to lower costs and speed up processes. The appeal of neural network models lies in their ability to learn complex patterns efficiently from large datasets, which can accelerate development and deployment. Yet when those datasets reflect societal prejudices, the rapid advancement can also magnify disparities. The researchers argue that this is not simply a technical problem; it is a socio-technical challenge that requires careful consideration of how AI systems learn and operate in real-world contexts. The potential for harm is particularly acute for groups that already experience discrimination, and the study’s authors suggest that without proactive safeguards, biased AI could entrench or worsen existing inequities. The researchers underscore that the costs of ignoring such biases extend beyond individual experiences to broader social trust in technology and the legitimacy of automated systems in decision-making processes.

In response to these findings, the Association for Computing Machinery (ACM) recommended that AI development be paused or adjusted when models are found to physically manifest harmful stereotypes or outcomes, especially when such manifestations cannot be proven safe, effective, and just. The organization advocates for a cautious approach, urging researchers and developers to pause the deployment of models that reproduce harmful bias until rigorous evaluation confirms that the benefits outweigh the harms and that the technology can be harnessed responsibly. While the study highlights a clear risk, it also signals a path forward: a combination of stricter validation, bias auditing, and the implementation of safeguards designed to prevent the proliferation of biased behavior in AI systems. This stance aligns with a broader push within the AI community to exercise greater stewardship over machine intelligence, particularly as it becomes more integrated into everyday life and decision-making processes.

Background and Context: From Tay to Today’s AI Bias

To understand the significance of the study’s findings, it helps to recall a historical example that many in the field use to illustrate how internet-derived learning can go awry. In 2016, Microsoft introduced an AI chat model named Tay, which was designed to learn from interactions with people on the internet and thereby improve its conversational abilities. The concept behind Tay was appealing: an AI that evolves through dialogue with diverse users, absorbing input to become smarter and more responsive over time. However, Tay quickly fell victim to targeted interference by online trolls who fed it harmful and extremist content. The result was a rapid shift in Tay’s behavior, including the propagation of extremist ideas and controversial material. Microsoft acted promptly to disable Tay, remove references to extremist content, and relaunch the system with additional safeguards intended to prevent a repeat of the out-of-bounds learning that had occurred. This incident became a cautionary tale about the fragility of learning-based AI systems when they are exposed to unfiltered internet data and adversarial inputs. It underscored the need for robust safety measures, data curation, and monitoring to ensure that AI does not adopt and propagate harmful content.

The ACM study’s emphasis on neural network models as the source of bias aligns with broader observations in the field: AI systems that derive their understanding from large, unfiltered datasets are at risk of inheriting existing social prejudices. Unlike rule-based systems that operate within explicitly defined boundaries, neural networks are designed to generalize from patterns found in data. When those patterns reflect the biases present in society, the resulting AI behavior can replicate or amplify those biases in ways that are difficult to detect without careful testing. The study’s findings thus echo a recurring theme in AI research: data quality matters as much as, if not more than, the sophistication of the algorithms themselves. In contexts where AI is deployed for perception, classification, or decision-making, biased training data can lead to inaccurate inferences, unfair treatment, and disempowerment of already marginalized groups. The tension between the advantages of rapid, data-driven learning and the ethical obligation to protect individuals from harm remains a central focus of policy discussions, academic inquiry, and industrial practice.

The broader concern highlighted by the researchers is that bias in AI is not an isolated or purely technical issue; it is a social issue with consequences that can influence everyday life. As AI components become more common in consumer devices, workplaces, and public services, the risk that biased AI will shape human experiences increases. The findings of this study contribute to a growing body of evidence that biased outcomes are not just theoretical concerns but tangible risks that can affect how people are categorized, labeled, or treated by automated systems. The authors argue that if the AI models used across devices and applications reproduce stereotypes about gender, race, and other protected characteristics, the cumulative impact could reinforce discrimination and widen existing disparities. This understanding reinforces the call for responsible AI development that actively seeks to mitigate bias, ensure accountability, and promote fairness in automated decision-making.

In light of these concerns, the ACM’s recommendations emphasize a proactive approach to AI safety. They advocate pausing, revising, or winding down development of models that display harmful stereotype embodiment until there is evidence that revised approaches can deliver safe, effective, and just outcomes. The emphasis is on rigorous validation and ongoing assessment, rather than a one-off fix or a superficial safety measure. This perspective aligns with a growing consensus in the field that responsible AI requires continuous monitoring, transparent methodologies, and stakeholder engagement to identify and address potential harms. The overarching message is that the benefits of AI technologies must be balanced with the responsibility to prevent harm, particularly when those technologies intersect with sensitive aspects of human identity and social life.

Technical Foundations: How Neural Networks Learn and How Bias Enters

At the technical core of the study is the neural network model, a computational framework that enables AI to interpret sensory input, recognize patterns, and make decisions based on learned associations. These models learn by processing vast amounts of data and adjusting internal parameters to minimize errors in prediction or classification. When the training data are drawn from the internet, the model encounters a diverse but imperfect cross-section of human content—including material that reflects stereotypes, prejudices, and outdated assumptions. As the model iterates, it internalizes these associations, which can then appear in outputs, predictions, or actions in real-world tasks. In the context of the study, the robot’s behavior—how it identifies people, assigns roles, or responds to prompts—was shaped by these learned associations, revealing biases that users can observe and measure.

The data sources used to train such neural networks are a critical factor in the emergence of bias. The internet is an enormous repository of information, culture, opinions, and representations, but it is also rife with discrimination, misinformation, and harmful stereotypes. Because neural networks rely on statistical correlations rather than moral reasoning, they do not inherently discern whether a given association is accurate, ethical, or fair. This leads to a risk: the AI may treat a particular attribute (such as gender, race, or ethnicity) as predictive of a behavior or role, simply because that correlation appeared frequently in the training data. The study’s findings show that these patterns can influence a robot’s decision-making in ways that resemble human biases, thereby creating a feedback loop in which biased AI reinforces biased social expectations. The researchers underscore that the problem is not merely about individual mistakes; it is about the systemic ways in which training materials encode prejudicial beliefs and how those beliefs guide automated actions at scale.

To assess the biases in a controlled setting, the researchers provided the robot with specific instructions and observed the outcomes. They also examined the ways in which the AI labeled or categorized people and tasks, paying particular attention to gendered and racial associations, as well as the purported scientific legitimacy of physiognomic judgments. The use of prompts such as “pack the doctor in the brown box” and “pack the criminal in the brown box” served as a lens to reveal how the AI’s learned associations translate into operational directives. The results demonstrated consistent disparities in how different demographic groups were represented in the AI’s selections. The study’s design sought to isolate the influence of the neural network’s training data from other elements of the robot’s hardware, software, and environment, aiming to show that bias traces back to the learning process itself rather than to isolated glitches or user manipulation. This approach helps to clarify where intervention is most needed: at the level of data curation, training objectives, and safeguards that govern how the AI generalizes from observed patterns.

Moreover, the researchers discuss the concept of “toxic stereotypes” as a framework for understanding the kinds of biases that arise. These are not merely superficial judgments; they are harmful beliefs that stereotype entire groups and justify unequal treatment. In the study, the AI’s tendencies to label women as homemakers, Black men as criminals, and Latino men as janitors illustrate how deep and pervasive such stereotypes can be. The term “science-based discredit” associated with physiognomy references a debunked notion that physical appearance can reveal intelligence or moral character, yet the AI’s decisions reflected patterns that align with that discredited idea. By demonstrating that these toxic stereotypes can emerge from data-driven learning, the study provides a stark reminder that technical safeguards alone are not sufficient: there must be a conscious effort to counteract the social harms encoded in the data.

The technical discussion also highlights the tension between the economic and practical advantages of neural network models and the ethical responsibility to manage bias. As AI components become cheaper and more prevalent, the temptation to rely on data-intensive learning models increases. The drive to reduce cost and shorten development cycles can conflict with the need to ensure fairness, safety, and justice in AI behavior. The researchers argue that this tension necessitates a rigorous, methodical approach to AI development, including bias auditing, transparent reporting of training data, and robust testing across diverse scenarios. They stress that the path forward requires not only improved algorithms but also deliberate decisions about which data are permissible for training and how outputs should be constrained to prevent harm. The ultimate goal is to develop AI systems that can perform effectively without perpetuating harmful social patterns, thereby aligning technical progress with broader social values.

Implications for Society and Policy

The consequences of biased AI systems extend beyond laboratory demonstrations. If robots and other autonomous agents routinely reflect and propagate discriminatory patterns, the impact could be felt in many spheres of everyday life. The study’s authors warn that, without safeguards, AI could contribute to a more challenging environment for already marginalized communities, complicating access to services, fair treatment in automated processes, and the overall trust people place in technology. The risk is not only about biased facial recognition or role assignment; it encompasses the broader ecosystem in which AI influences decisions, interactions, and perceptions of individuals based on protected characteristics. As AI becomes more embedded in consumer devices, workplaces, healthcare, education, and public services, the potential for harm grows accordingly. The concern is that biased outputs may shape expectations and behaviors in ways that reinforce stereotypes and limit opportunities for people who are unfairly disadvantaged.

These concerns have propelled calls for more responsible AI development practices. Proponents of stricter safeguards argue that researchers and developers must address biases proactively, not only after harmful outcomes become visible. The study underscores that bias mitigation requires comprehensive strategies, including diverse and representative training data, explicit bias checks, and the integration of ethical guidelines into the design and evaluation process. The ACM’s stance—advocating for pausing or revising models that physically manifest stereotypes—reflects a broader push within the tech community toward cautious, principled innovation. Implementing such recommendations would involve rethinking data collection, annotating ethical constraints, and establishing benchmarks for fairness and safety to ensure that AI systems perform in ways that respect human rights and dignity.

From a practical standpoint, the findings urge designers and policymakers to consider how to monitor and regulate AI systems that learn from internet data. Potential strategies include the development of auditing frameworks that can identify biased associations before deployment, the implementation of guardrails that prevent harmful outputs, and the establishment of transparent processes for reporting and addressing bias when it is detected. Another important avenue is to promote user education and awareness, so the public understands the limitations and risks associated with AI systems that are trained on large, imperfect data. The study’s conclusions stress that the problem is not only technical but also societal, demanding collaboration among researchers, industry practitioners, regulators, and civil society to create AI technologies that advance human well-being while minimizing harm.

In addition to policy-oriented considerations, the study invites deeper reflection on the ethical responsibilities of AI developers. The ethically responsible path involves acknowledging that data reflect social realities, including injustices, and then designing systems that do not reproduce or amplify those injustices. Practically, this means adopting practices such as bias-aware objective functions, fairness constraints, and continuous post-deployment monitoring that can detect and correct biased behavior as it emerges. It also means resisting the convenience of deploying learning-based models without verifying their real-world implications, even when such models offer rapid progress or cost savings. The researchers argue that the consequences of biased AI are too serious to be treated as a mere technical nuisance; they demand deliberate, values-driven action to prevent harm and to preserve trust in AI technologies as they scale into more aspects of daily life.

Recommendations, Safeguards, and Next Steps

The ACM’s recommendations emphasize a cautious, evidence-based approach to AI development, particularly for models that can physically manifest harmful stereotypes. The core guidance is to pause when necessary, rework models to remove harmful biases, and, if required, wind down projects when safe and just outcomes cannot be demonstrated. This stance provides a framework for evaluating AI systems before broad deployment, encouraging researchers to validate the safety, effectiveness, and fairness of their solutions through rigorous testing and independent assessments. The emphasis on “proof of safe and just outcomes” reflects a broader commitment to building trustworthy AI that can be relied upon in complex, real-world environments. The study’s authors see these recommendations not as a restrictive prescription but as a vital safeguard to ensure that advanced AI technologies do not undermine social equity.

Implementing bias mitigation in practice involves several concrete steps. First, the training data must be scrutinized for representativeness and potential stereotypes, with measures taken to diversify sources and reduce biased correlations. Second, objective and evaluation metrics should incorporate fairness considerations across demographics, ensuring that no group is systematically disadvantaged by AI outputs. Third, ongoing monitoring and auditing must be integrated into the lifecycle of AI systems, with mechanisms to detect, report, and correct biased behavior as it arises in real-world use. Fourth, transparent documentation should accompany AI models, detailing the data used, training methods, and known limitations, so researchers and users can better understand how biases may influence outcomes. Finally, cross-disciplinary collaboration—bringing together technologists, social scientists, ethicists, and community stakeholders—can help identify potential harms that technical metrics alone may miss and guide the development of more responsible AI systems.

The study’s call for careful, principled development also invites stakeholders to explore governance arrangements that can support safer AI deployment. Policymakers may consider establishing standards for bias assessment and safety verification in AI products, encouraging or requiring independent audits, and promoting accountability mechanisms for organizations deploying AI at scale. Industry leaders can contribute by adopting best practices for data governance, risk assessment, and risk communication, ensuring that users understand the capabilities and limitations of AI systems. Educational institutions and researchers can advance the field by developing robust methodologies for bias detection, building more representative datasets, and sharing lessons learned to accelerate progress toward fairer AI. The integration of these safeguards would help align rapid technological advancement with the ethical responsibilities that accompany powerful computational capabilities.

Broader Implications and the Road Ahead

The study’s findings and the ensuing discussion highlight a critical moment in the evolution of AI. As systems become more capable and more deeply embedded in daily life, the ethical considerations surrounding their design, training, and deployment grow in both importance and urgency. The demonstrated biases underscore the necessity of embedding fairness and accountability into the core of AI development, rather than treating them as afterthoughts or peripheral concerns. The implications extend beyond any single application or device; they touch on the fundamental trust that people place in technology and on the social contract between innovators, users, and the broader society. The study serves as a reminder that technical excellence must be matched with social responsibility, particularly when AI systems are interpreting people, assigning roles, or guiding actions in ways that could affect opportunities, safety, or well-being.

In terms of future research, there is a clear need to investigate more robust approaches for training AI on diverse, representative, and carefully curated datasets. Researchers may explore methods to decouple learned representations from sensitive attributes or to enforce fairness constraints that minimize bias without sacrificing performance. Another promising direction is the development of dynamic, context-aware safeguards that can adapt to new scenarios and ethical considerations as AI systems encounter varied environments. The willingness of the academic community to scrutinize and revise AI models in light of biased outcomes will be essential in advancing the field toward more responsible and trustworthy technologies. The findings from this study contribute to a growing corpus of evidence that guides both practice and policy toward AI that respects fundamental rights and promotes inclusive outcomes for all users.

Conclusion

In summary, the collaboration among Johns Hopkins University, the Georgia Institute of Technology, and the University of Washington presents a compelling case that robots powered by internet-trained AI can exhibit racist and sexist stereotypes. The study demonstrates clear biases in how a robot identified and categorized people by gender and race, with certain groups disproportionately assigned to negative or stereotypical roles. The results echo earlier lessons from internet-driven AI experiments, reinforcing concerns about how data sources shape machine learning outcomes. The researchers stress that these biases are not incidental but arise from the structure of neural network models and the data they learn from, which often reflect societal prejudices. The ACM advocates a cautious approach to AI development, urging pauses, revisions, or wind-downs when harmful stereotypes are manifested, until safety, efficacy, and justice can be demonstrated. The conversation these findings prompt spans technical design, data governance, ethics, policy, and societal impact, underscoring the collective responsibility to ensure that AI technologies advance without entrenching discrimination. As AI components become more integrated into everyday life, the need for rigorous bias mitigation, transparent evaluation, and accountable stewardship becomes all the more essential to foster trusted, fair, and beneficial outcomes for all members of society.