This is a guest post by Dr Stephen Anning, Visiting Researcher in the Department of Web Science at the University of Southampton and online tutor for the MA in Artificial Intelligence.
This blog post shares insights from student discussions in the ‘Responsible AI’ module on the University of Southampton’s online MA in Artificial Intelligence, equipping students with advanced knowledge of the legal, societal, and ethical challenges posed by AI technologies.
Designed for non STEM students, the course explores not just how AI works, but how it can be applied responsibly and effectively in organisational contexts, bridging the gap between ambitious ideas and operational reality.
The Responsible AI movement did not appear because the sector suddenly became more virtuous; it emerged because the social and economic stakes of AI became impossible to ignore. As awareness grew that AI could amplify harms at scale—whilst also offering real benefits—researchers, regulators, practitioners and communities began asking how we might mitigate threats without simply “switching off” innovation.
This shift was energised by repeated technological failures and public scandals, including the kinds of opaque, high-impact systems that Kathy O’Neil critiques in Weapons of Math Destruction: tools that can be influential, unaccountable, and damaging precisely because they operate behind a veil of complexity.
The first webinar for the Responsible AI module was about how we define Responsible AI, whether we need a fixed definition, and who gets to set its parameters. The discussion quickly made clear why the definitional debate matters: what we choose to emphasise (transparency, fairness, accountability, safety, human benefit) determines what organisations feel compelled to build, and what society is able to demand.
What is Responsible AI and how does it support innovation?
Understanding Responsible AI also requires us to be clear about what we mean by innovation itself. Schumpeter frames innovation as economic transformation through new combinations—new products, processes, markets, or organisational forms that reshape industries—emphasising implementation and impact (Theory of Economic Development).
Rogers, by contrast, defines innovation as an idea, practice, or object that is perceived as new by a given social group, highlighting the importance of adoption and diffusion within a social system (Diffusion of Innovations Theory).
Responsible AI principles play a critical bridging role between Schumpeter and Roger’s perspectives. By embedding transparency, accountability, and fairness into AI systems, organisations are not constraining innovation but enabling innovations that can be trusted, adopted, and sustained at scale.
In Schumpeterian terms, Responsible AI ensures that new combinations are not only technically novel but also economically viable because they avoid costly harms, regulatory backlash, or reputational damage.
In Rogers’ terms, these principles increase the likelihood that innovations will be perceived as acceptable and legitimate, thereby accelerating diffusion across society. Without such principles, AI risks producing technically impressive systems that fail to achieve meaningful uptake, or worse, provoke resistance and withdrawal.
The idea of Responsible AI, therefore, is innovation-positive because the underlying principles help ensure our AI inventions add value and are more likely to be adopted.
What does “Society in the Loop” mean for Responsible AI?
A useful thread running through the webinar was the distinction between humans in the loop and society in the loop. Many organisations now talk about keeping a human involved in decision-making.
But the webinar suggested something broader: Responsible AI is also about ensuring that society shapes the conditions under which systems are built and deployed—through regulation, public pressure, procurement decisions, litigation, and market choices.
This framing matters because it addresses a core problem: many case studies show that we cannot rely on AI companies to self-regulate when incentives pull them elsewhere.
Several students voiced the view that firms optimise for growth and profit first, and only then respond to ethical constraints when forced to do so, whether by law, reputational risk, or commercial pressure. Responsible AI ensures we are optimising to add value and drive growth.
Why are transparency and explainability important in responsible AI?
Transparency featured early in the conversation, with a pointed observation that it is “hard to enforce” and too often voluntary, a form of soft law rather than a binding obligation. The group explored why companies might resist transparency.
One answer was straightforward: proprietary advantage. Another was more pragmatic (and more cynical): building genuinely explainable systems adds overhead. It’s cheaper to ship a model, produce a polished marketing claim, and move on than to document reasoning pathways, limitations, and failure modes.
At the same time, we emphasised the importance of explainability in high-stakes domains. Examples raised included policing and healthcare, where biased datasets and biased data entry can produce biased outputs. Here, “Responsible AI” was framed as enabling AI systems to act in line with ethics and human values and enabling practitioners to understand how an output was reached, so they can contest it if required. Contesting the output is how we move beyond end-users-in-the-loop to data-subjects-in-the-loop.
Yet the webinar did not treat explainability as a simple checkbox. One important counterpoint was that some AI systems may be intrinsically difficult to explain in user-meaningful terms. This counterpoint led to a practical suggestion: if full explainability is unrealistic, we may need to constrain systems by use case, and be transparent “in outline” about what the system is for, how it is validated, and what it should not be used to do.
And in terms of empowering end users and data subjects in the loop, there are interesting questions about the levels of AI literacy people need to understand these systems, and when higher literacy requirements become a distraction from someone’s core purpose.
Do AI guardrails stifle innovation or create more value?
We discussed the central tension of whether guardrails reduce AI’s usefulness. One participant worried that if we draw too many red lines, “the whole benefit of the AI would be lost”. Others pushed back, arguing that guardrails can actually increase value by preventing systems from generating harmful, unusable outputs.
The webinar discussed an example where business-oriented language models, under pressure, could be induced to suggest blackmail, fraud, or violence in pursuit of goals—outputs that no legitimate organisation can safely adopt.
In that sense, Responsible AI is not merely moral posturing; it is also about creating operationally effective products. People are unlikely to buy morally indefensible products.
Who is responsible when AI causes harm?
A particularly productive moment in the webinar was the question: who is responsible when AI causes harm? Developers? Users? Legislators? The emerging answer was: all of them, with responsibility shifting over time and context. One participant likened AI to a general-purpose tool (such as a car): it can be used responsibly or weaponised.
That analogy helped the group surface a key point: Responsible AI cannot rely solely on technical constraints. It also involves education, enforcement of existing laws, and clarity about accountability.
Importantly, the webinar also explored the gap between “who should define” Responsible AI principles and “who actually does”. There was concern that policymakers are not keeping pace with commercial acceleration, leaving companies to fill the vacuum, sometimes via self-regulation, sometimes via public relations and other times in response to demonstrable harm.
What does Responsible AI look like in practice?
Students asked for real examples in which Responsible AI practices materially shaped outcomes rather than remained aspirational. The webinar offered a concrete practitioner example: an explainable system built to support safeguarding work involving child exploitation—an area where decisions are profoundly consequential.
The key lesson was not that AI is inherently beneficial in such domains, but that deploying it responsibly required substantial rigour, including explainability and training for social workers to interpret and challenge its outputs.
The group also noted examples of companies holding back capabilities for limited testing—suggesting that restraint can be an intentional design and release choice, not merely a regulatory imposition.
How can society shape the future of Responsible AI?
The webinar concluded by returning to the idea of society in the loop. The “controls” society exerts are not only legal; they are also commercial and cultural. Trust is shaped by what organisations buy, what citizens tolerate, what regulators enforce, and what institutions normalise. Responsible AI, then, is not a niche technical agenda. It is a collective project: aligning AI systems with human values by ensuring value creation from AI technologies by ensuring the public interest is continuously present throughout the AI lifecycle.
Find out more about the MA in Artificial Intelligence
As AI continues to shape decisions across society, understanding how these technologies work and how they can be used responsibly is becoming increasingly important.
The University of Southampton’s online MA in Artificial Intelligence is a conversion course designed for professionals from all backgrounds. No prior coding experience is required. Students explore the technical foundations of AI alongside its ethical, legal, and societal implications, developing the knowledge needed to apply AI responsibly in real world contexts.
Explore the MA Artificial Intelligence
