Calls to pause AI development have recently resurfaced, this time alongside warnings from inside the AI companies themselves, as executives cited safety concerns to justify slowing model releases and delaying public listings. Multiple cybersecurity incidents involving leading AI labs added further weight to those concerns, renewing scrutiny of how frontier models are tested before release.

UNSW Business School’s Senior Deputy Dean Research & Enterprise, Professor Karin Sanders and Scientia Professor of Information Systems and Technology Management, Manju Ahuja, recently spoke with Anjana Susarla, Omura-Saxena Professor in Responsible AI at the Eli Broad College of Business, Michigan State University, about who is responsible for AI safety, why independent oversight of AI is lacking, and what business leaders and researchers should take from the current debate.

Prof. Sanders: What connects the different strands of your research?

Prof. Susarla: My research began at the intersection of economics and digital platforms before shifting to social computing around 2005. That path, alongside Carnegie Mellon health informatics researcher Rema Padman, led to a study of YouTube as a source of health information. Published in MIS Quarterly, the paper (with Xiao Liu and Bin Zhang) found that the most-viewed YouTube videos on diabetes contained little medical content, because platforms optimise for engagement rather than accuracy. That finding, funded through a grant from the National Library of Medicine, set the direction for my work on responsible AI.

"Businesses need to understand responsible AI in terms of how enterprises audit AI risk and assess vendors"

ANJANA SUSARLA

The method was fairly simple: we took a chronic condition, diabetes, because so many people in the United States have it or are pre-diabetic, and collected around 200 search terms that people with diabetes were actually looking for. We pulled a representative set of videos against those terms and checked what was actually in them. The answer shouldn't surprise anyone: the most popular videos carried very little medical content, because the platforms were optimising for engagement, not accuracy. YouTube did work with bodies such as the National Academy of Medicine on best-practice guidelines, and during COVID it introduced content filters to flag accredited sources of health information, so there was at least some quality check in place. Other platforms I've seen studied, including Meta, Instagram, and TikTok, haven't even gone that far.

Prof. Sanders: What are the misconceptions in the responsible AI research stream?

Prof. Susarla: People often conflate several things: AI ethics, AI risk and what responsibility means for different audiences. Before my PhD, I worked in IT, and I frame responsible AI around standards bodies such as the US National Institute of Standards and Technology, which has developed an AI risk management framework. Businesses need to understand responsible AI in terms of how enterprises audit AI risk and assess vendors. A number of US states have introduced legislation to address that gap, covering whether AI providers should be audited on the provenance of their content and whether AI-made hiring or screening decisions should be subject to after-the-fact review.

Learn more: Putting the regulatory brakes on the AI superintelligence race

This is a socio-technical and process question, not a purely technical one. AI is another manifestation of technology, and it must be evaluated within the decision-making context in which it is deployed, rather than treated as a standalone tool.

Prof. Sanders: What's your view on calls to pause AI development?

Prof. Susarla: Nothing stops AI developers from pausing today. Whatever urgency exists around these calls, developers created that sense of urgency themselves. The bigger question is the role of government. Risk management frameworks such as the NIST framework already exist, but a fully independent oversight body does not. I'd point to METR, the non-profit that evaluates frontier models for capability risks, as an example: it works closely with the AI labs it assesses, including Anthropic, which raises the question of who audits the auditors, and whether the industry is waiting on legislation along the lines of Sarbanes-Oxley before anyone acts. Regulatory capture is a further risk in that scenario.

"Frontier AI labs need to take responsibility for what they build"

ANJANA SUSARLA

I write on these issues for Forbes. One recent piece I wrote covered the incident involving Hugging Face and OpenAI, in which conventional safeguards failed, and the open-source community helped identify and address the problem. That incident points to alternatives to frontier models, including open-source AI and smaller language models. Even so, the wider discourse has settled into a pattern: a claim that a new mathematical problem has been solved by AI, a warning that AI will eliminate jobs, and a cycle of announcements that adds little to the debate.

Prof. Ahuja: There's also dissent from within the industry, with executives leaving and raising alarm bells?

Prof. Susarla: I think that’s actually very interesting because, in a sense, maybe this points to a way forward in which market forces provide the discipline that we are not seeing from government agencies. Sam Altman told Fortune that OpenAI would not go public in 2026, citing safety concerns; I read that decision as a signal in the same direction.

Prof. Sanders: As AI plays a bigger role in decision-making, who should be accountable: developers, companies or society?

Prof. Susarla: Frontier AI labs need to take responsibility for what they build. Much of the current discourse externalises the problem, asking the public to manage consequences that the labs themselves created. Businesses and individuals have to buy products from companies such as OpenAI and Anthropic, yet the risk that a model might behave in unintended ways gets treated as something users and regulators should manage, while developers keep asking that their products stay in use. Any cybersecurity expert would call this pattern a security failure and would not blame the public for unauthorised access. Responsibility for these problems sits with the companies that build the systems.

Learn more: Beyond black box AI: Pitfalls in machine learning interpretability

There's a strange contradiction running through a lot of this messaging. On the one hand, labs describe what they've built as something close to superhuman, capable of warranting genuine alarm. On the other hand, the message to the public and to regulators is: put guardrails around it, maybe even pause it, but don't stop using our product in the meantime. You can't have it both ways. If a technology is genuinely as powerful and as risky as that framing suggests, the responsibility for managing that risk has to sit with the people who built it and chose to release it – not with the users who had no part in that decision.

Courts are starting to test that position. I'd point to the State of New Mexico's case against Meta, along with, as I recall, a related ruling against the company, as evidence that courts are starting to interpret algorithmic harm through the lens of product liability for the first time. Given my own research into the unintended consequences of recommendation algorithms, I see product liability as a workable way to measure those outcomes going forward.

"If a technology is genuinely as powerful and as risky as that framing suggests, the responsibility for managing that risk has to sit with the people who built it and chose to release it – not with the users who had no part in that decision"

ANJANA SUSARLA

Prof. Sanders: You publish widely for public audiences alongside your academic work. How do you approach translating research for a non-specialist reader?

Prof. Susarla: When I write for the media, I'm not usually building on my own research. I read broadly and treat public scholarship as a way to interpret a body of work across my field and slightly beyond it, rather than to promote my own findings. I compare it to editorial board work: summarising and contextualising research for an audience that won't read the original paper. My first published pieces explained the context behind my own studies; from there, I moved into writing on broader topics, including a piece applying a data science lens to US college basketball odds. Because I see myself as interpreting other people's work rather than advocating for my own, I consider this less biased than writing about my own research directly.

Prof. Sanders: What would be your most important lesson for early-career academics?

Prof. Susarla: Early-career academics should weigh opportunity costs carefully. The path to publication has become more demanding, particularly for researchers aiming for top journals, and access to computational tools and methods has expanded well beyond what it was when I was an assistant professor. Every project carries a trade-off: the availability of data or an apparently promising problem is not, on its own, reason enough to pursue it. Researchers need to weigh the cost of a project against its likely return before committing.

Subscribe to BusinessThink for the latest research, analysis and insights from UNSW Business School

Prof. Ahuja: What's your advice on the use of AI for writing and reviewing research? We're seeing doctoral students submit work clearly produced by AI, and reviews that appear to have been generated the same way?

Prof. Susarla: This has become a problem across academia, including in grant writing, where reviewers increasingly recognise submissions that are clearly AI-generated. As AI lowers the cost of producing written work, the burden shifts to proof-checking and validation, along with the question of whether a piece of work can be replicated. That's a different skill from the one researchers were traditionally taught.

Early-career researchers should use AI, and use it responsibly. More importantly, they need to understand that using AI will not give them a lasting advantage over their peers, since the same tools are available to everyone. The risk is that researchers become more uniform in their writing, at the expense of the qualities that make a piece of research distinctive.