Regulation-dodging companies push "responsible AI," but research shows otherwise

The big AI companies regularly tout the growing capability of their wares. But they also want the power to chart their own course free from regulation.

Look at the recent "manifesto" from Palantir CEO Alex Carp. He argues that the most "responsible" thing a tech company can do is ensure that Western liberal democracies maintain a decisive edge in AI over their adversaries. In other words, trust us and get out of the way.

It's an attitude that makes sense when you look at the latest version of Stanford's Foundation Model Transparency Index, which shows that disclosure about training data, compute usage, and real-world impacts fell from an average of 58 in 2024 to 40 in 2025.

This decline happened as corporate AI investment reached nearly $600bn, and the number of AI incidents kept climbing. While capability increased, it's clear that transparency went down. At the same time, so-called "responsible AI" continued to adorn strategy documents and conference agendas.

A broad church

The responsible AI movement was always a broad church. As with similar movements before, it was created with good intentions. The founders argued that AI systems should be fair, explainable, safe, and accountable. Frameworks multiplied. Principles proliferated. Ethics boards were convened, then quietly disbanded.

What never really happened was a clear definition of what “responsible” is and who gets to classify themselves as such.

A recent report from MIT Sloan Management Review and Boston Consulting Group clearly demonstrates the problem. Around 80% of respondents said that responsible AI practice should address workforce impact, not just AI system risk.

It's an uncontroversial finding, but in practice, even the sensible remains a minority position. The majority of AI governance remains focused on the models themselves. They cover things like accuracy, bias, and the tendency to produce confident nonsense.

In many ways, this narrowness is no accident. The vocabulary used to describe AI safety, such as bias and hallucination, was largely forged by an ideologically narrow group of Silicon Valley research labs. The concern was more about the behavior of the models than the effect on people who use them.

Labor displacement, discriminatory algorithms in criminal justice, and the concentration of economic power were all treated as downstream problems for policymakers rather than technologists. As a result, the industry ended up with a governance framework that it could largely accept, partly because of helping to draft it.

"Joan Tronto referred to it as care washing, and it's a really big concern as you end up using the concepts behind care to justify tools that may look caring but actually aren't at all," Caroline Green, developer of Civic AI and researcher at the University of Oxford, told me before her talk at SXSW.

Structural problems

The Stanford HAI AI Index clearly shows the structural problems with this approach. It highlights how the industry is keen to show off capability benchmarks, while at the same time ignoring benchmarks for safety, fairness, and factuality. The industry has, in effect, agreed on a shared standard for measuring what its systems can do, while maintaining the freedom to keep quiet about what they might cause.

Nowhere is this evasion more apparent than in workforce data. There's a clear trend towards entry-level jobs being gutted, especially in areas like software development. This is not, as some had hoped, a margin-repair story dressed up in AI language.

Something structural is happening to the youngest workers in the most AI-exposed occupations, and the governance apparatus built to manage "responsible AI" has essentially nothing to say about it.

The MIT researchers agree that the issue goes beyond merely those immediately affected. Entry-level roles are where tacit knowledge and professional networks get built. As soon as we erase those, the damage goes beyond today's workforce and into the mid-career workforce a decade down the road.

Such forward thinking is rare when companies focus on short-term productivity gains.

A fragmented response

Sadly, the governance response to this has been pretty fragmented. This was demonstrated clearly at the start of 2025, when the United States signed an executive order treating AI as a strategic asset to be supported in any way possible.

Across the Atlantic, the EU AI Act's first prohibitions took effect, treating it as a risk surface to be constrained. At the same time, China finalized mandatory content-labeling rules, treating it as a domain to be controlled.

Three jurisdictions, three completely different theories of what the problem even is. Even the European Union is struggling to fulfil its role as the world’s regulator. American companies serving European markets have shown a clear willingness to modify or even restrict products in Europe rather than change their underlying systems.

Ahead of her talk at SXSW London, Gartner’s Priscila Chaves explained that the AI industry is increasingly not just selling a product but engineering trust. For instance, agentic AI isn't just assisting us; it's the interface through which decisions are made, relationships are mediated, and emotional life is organized.

We're in a world in which hundreds of millions of people are already in regular, intimate interaction with AI companions and assistants whose terms of service explicitly assign ownership of those interactions to the platform.

Any talk of "responsible AI" has not really caught up with this reality. It is still largely a conversation about model outputs, not about the infrastructure of trust being built on top of them.

What we can do

Fixing this is by no means easy, but any serious attempt must try to achieve a number of things. Firstly, governance frameworks would be expanded to treat workforce impact as a core design parameter, not an external policy question.

We would also need to make responsible AI benchmarks as weighty and influential as the kind of performance benchmarks that AI companies strongly pursue. This pressure needs to be both social and regulatory.

Bolstering consumer rights would also be a worthwhile step, especially in areas like healthcare, education, and companion services, where the power asymmetry is greatest.

Last, but not least, labor protections need to adjust to a world in which roles are increasingly eliminated and restructured by AI and its management. We're much better at the former than the latter at the moment.

Inevitably, regulation will always move at a slower pace than technology. In a sense, regulation has always been reactive in nature, but AI is making the lag more painful. We need to be honest with ourselves, though.

"A really important idea is the boundedness of AI, which means that it's not created for maximisation and for continuing to grow more powerful through more data, but is instead created with the purpose of serving people, with people having the power to shut it down," says Green.

Responsible AI, as currently practiced, is a governance regime whose terms were largely set by the industry it is meant to govern. It is less a set of constraints than a very polite request. It's not a question of whether better approaches are feasible, as they clearly are, but whether we have the institutions with the appropriate authority and urgency to act before the ways of working become entrenched.

Despite promising frameworks, it's hard to be optimistic that the tech leviathan will change course, which may be telling in itself.

Unlock more exclusive Cybernews content on YouTube.