This Month in Responsible AI with Margaret Mitchell
All Tech Is Human Executive Director Rebekah Tweed sat down with Dr. Margaret Mitchell, Researcher and Chief Ethics Scientist at Hugging Face for All Tech Is Human’s “This Month in Responsible AI,” a monthly livestream series hosted by Rebekah Tweed that features special guests invited to talk about emerging tech issues. The conversation centered on "Fully Autonomous AI Agents Should Not be Developed," the paper that Mitchell and her Hugging Face colleagues Avijit Ghosh, PhD, Dr. Sasha Luccioni, and Giada Pistilli released last month.
Rebekah and Margaret discussed the authors’ impetus for writing the paper, their definition of fully autonomous agents (“Computer software systems capable of creating context- specific plans in non-deterministic environments”) and their description of progressive levels of autonomy:
Simple processor → Router → Tool call → Multi-step agent → Fully Autonomous Agent.
The conversation turned to the additional ethical dimensions that are introduced with agents, like the delegation of decision-making authority, the role of human oversight, and potential for emergent behaviors, before turning to a discussion of some specific risks that increase with increased agentic levels of autonomy that are not necessarily inherent to the models themselves, like safety, security, and privacy.
Shortly after the paper was published, Chinese startup Butterfly Effect released Manus, arguably one of the buzziest agents to ever hit the market.
“Manus is super interesting to me because I had put out all this writing about stuff that was foreseeable to happen in the future that was a bit worrisome — and and then Manus just did all that! I was like — think twice about saying “fully autonomous”; Manus comes out: ‘We’re fully autonomous!’ Like, think twice about saying something is “general AI”; Manus comes out: ‘We’re general AI!’ Uh, so it was an interesting kind of like ‘Ah, yes, this is an example of the things I was worried about.’”
Margaret discussed several projects she is currently working on at Hugging Face, like Open R1:
“We’ve released Open R1 which is an open source version of DeepSeek. The goal was to try and see - can we create a large language model that is on par with the best proprietary technology, that’s on par with DeepSeek, which blew so many other technologies out of the water? And we’ve recently discovered that we can. ...It’s in one of our recent AI policy reports - open approaches to AI development are on par with closed approaches in terms of actual capability, in terms of the benchmarks that we have to measure them. So that’s been a pretty big deal.”
SmolAgents was the next topic of discussion:
“SmolAgents is an agentic framework that HuggingFace has been working on for a while - actually, for a couple years - but just recently, it’s reached a point that I think is pretty exciting for people who are interested in building their own AI agents. There’s a tutorial. There’s a Discord channel (...they thought it was a big spam attack because so many people were taking this course). ...[It] uses the sandbox approach so you know it’s also influenced by this ethical thinking where we care about security, privacy, [etc]; so, as part of the SmolAgent framework, it provides the information about how to make sure that this is a secure agent and talks about pros and cons, how to make them even more controllable, how to not cede human oversight, [etc].”
Finally, Margaret shared efforts she is involved with at Hugging Face to build up evaluation technology that is more informed by how the technology is likely to be used and its actual societal impact. She and her colleagues are trying to develop a more rigorous science of measurement in AI in terms of both data and evaluation, to overcome the disconnect between the way that AI systems are being evaluated and benchmarked and what they are actually useful for and supposed to be doing in real-world deployments.
Margaret took a few audience questions, on agents’ impact on the tension between convenience and privacy, on the role of AI standards in governing the development of autonomous agents, and one pointed question on whether or not the issue of bias has been resolved:
“No it hasn’t. There have only been a few years where people actually cared about biases...As someone who has been working on this for a long time - I was working on fairness and bias before people cared about it. Then people started caring about it, which was amazing - and this was in large part because of the work of Joy Buolamwini, who really shone a light, and Timnit Gebru as well, my former co-leader at Google. They really were amazing at shining a light on what the issue was. It was exciting to me that suddenly people cared - but now they don’t care as much anymore. [The] pendulum swings - like, it was in for a while and now maybe it’s not in. But, it has not been fixed.”
Watch the whole conversation below:
