The Structural Failures Thwarting Responsible AI Development

By Deb Donig, All Tech Is Human’s Siegel Research Fellow

Over the past four years, I have been tracking the evolution of what we might call the "Responsible Tech workforce"—the people hired to make technology serve human values rather than just maximize engagement and profit. The data reveals a field that has largely evaporated even as public concerns about AI bias have grown more urgent.

In 2021, positions focused on ethical technology—roles like AI ethicists, algorithmic auditors, and diversity specialists—represented 58% of all jobs in the Responsible Tech ecosystem. By 2025, that number had collapsed to just 8%. Meanwhile, technical implementation roles focused on compliance and regulatory requirements exploded to dominate 66% of the market.

What This Shift Represents

This shift represents more than changing job titles. It signals a fundamental retreat from asking "what kind of AI systems should we build and why?" toward simply implementing predetermined requirements. What we are witnessing is the premature ossification of certain assumptions and frameworks into technical systems that will operate at massive scale across social, political, cultural, and economic life, all of it without the institutional capacity to deliberate about whether those embedded assumptions serve any coherent vision of human flourishing. 

Most tellingly, the remaining ethics positions are increasingly structured as contract work (typically temporary advisory roles without institutional power to influence how systems actually get built). This means that organizations can point to their PhD-holding ethics consultants to demonstrate they've "checked the ethics box" while ensuring those consultants can't challenge fundamental business practices or design decisions.

Deeper structural failures make accountable AI development nearly impossible. The infrastructure capable of building genuinely ethical AI systems has been systematically dismantled, even as political battles over AI bias have intensified.  Many critiques of AI bias perpetuate the fiction that biased AI systems result from individual programmers inserting their personal politics into code. This fundamentally misunderstands how AI systems work.  

Consider how large language models actually function: they're trained on massive datasets scraped from the internet, then fine-tuned using human feedback to generate responses that users find helpful and engaging. Any bias in outputs reflects some combination of what's in the training data (which mirrors existing biases in human-created text) and how the system has been optimized based on what companies believe users want or find valuable. 

The biases embedded in AI systems emerge from structural decisions about training data, optimization targets, and business models—decisions made at organizational levels far above any individual engineer. An engineer implementing bias mitigation techniques might reduce certain discriminatory outputs, but that person cannot fundamentally alter business models that prioritize engagement over accuracy, or data collection practices that determine what the system learns in the first place.

The Challenge of Building More Ethical AI Systems

This structural analysis reveals the fundamental challenge of attempting to build more ethical AI systems. The solution requires similar institutional capabilities: organizations where people with expertise in both technical systems and democratic values have real decision-making authority. But the current employment structure in tech systematically prevents this integration. 

Computer science programs teach machine learning optimization but rarely engage with frameworks for democratic governance. Policy programs teach regulatory frameworks but often lack technical literacy to understand how algorithmic systems actually function. Ethics programs develop moral frameworks but may not provide implementation skills necessary to embed insights into working systems. 

The kinds of programs we need are those that create pathways to "boundary-spanning" roles that operate at the intersection of technical implementation, ethical reasoning, and policy development. These positions require both quantitative research skills and normative reasoning capabilities. 

One such position I recently saw, an advertised position by the ACLU for an Algorithmic Justice Fellow position, for example, requires both quantitative research skills and legal literacy, combining technical auditing capabilities with civil rights advocacy. These "ethics translator" roles represent a potential path forward—but they remain rare and often misunderstood by traditional hiring practices. 

And yet, while the role is profound, the kind of training that would provide someone the means to accomplish this work is missing; this kind of professional is either a unicorn (and with those skills may want to move to positions that offer more funding than that provided by a nonprofit fellowship), or a very senior specialist. In either case, this professional would be somewhat of a unicorn, and neither of these two possibilities is scalable, at present. We need new, and better, infrastructure to train the workforce this work requires.

The broader failure here extends beyond workforce development into organizational logics and culture. Most current approaches to AI ethics focus on principle documents, advisory boards, and post-hoc auditing. Sometimes they rely on the operational and institutional capacities of non-profits to provide unpaid thought leadership and correctives, rather than companies themselves building organizational structures that can actually integrate ethical reasoning into technical decision-making. 

What Building More Ethical AI Systems Require

Building AI systems that serve democratic values, regardless of whether those values lean left or right, requires more than principles, boards, post-hoc auditing, and taking the occasional meeting with a concerned nonprofit advocacy group proposing correctives to tech products that are causing harm. 

It requires acknowledging uncomfortable truths about how power operates in technology companies. Individual engineers, regardless of their intentions or training, cannot solve problems created by organizational structures and business models that systematically prioritize short-term engagement over long-term social outcomes.

It requires building technology that actually serves human values and the  harder institutional work of implementing them, including means of evaluating success that assess AI systems based on their contributions to democratic flourishing rather than just their technical performance or user satisfaction scores.

Values won't be embedded in AI systems through executive orders or corporate diversity statements  or principles documents; they'll be embedded through business models and corporate leadership. Aspirationally, they should not just be embedded as orders but deliberated by people thoughtful and skilled on these topics, and sustained investment in building institutions, educational pathways, and professional practices capable of cultivating workers equipped to consider and implement democratic technological governance. 

Building Institutions, Educational Pathways, and Professional Practices

For that to happen, we need to start doing the patient work of building institutions, educational pathways, and professional practices that can actually hold technological power accountable to democratic communities. This includes developing integrated educational pathways that train technologists to think about social impact from the beginning, not as an afterthought, educational structures that engage seriously with questions of power, governance, and social responsibility alongside technical training, alongside employment structures that give people working on ethical AI real institutional power to influence outcomes, not just advisory roles that can be ignored when convenient. 

In the long run, this includes the re-imagining of business models that can survive prioritizing fairness and accountability over pure engagement metrics. And, finally, at some point, all of this likely requires regulatory frameworks that change the competitive landscape, not just voluntary  corporate initiatives.

Resolving these disagreements requires exactly the kind of sustained institutional development that our current approach to AI ethics has systematically failed to create. Until we build organizations capable of translating democratic values into technical systems — whatever those values turn out to be — we'll continue having symbolic battles over AI bias while the underlying problems remain unaddressed.