Three Detectors, Three Failure Profiles: Detector-Specific Demographic Bias in Public Deepfake Detection and Its Implications for Content Moderation in Latin America

This study tackles an urgent social challenge for content moderation in Latin America and convincingly demonstrates how aggregated metrics can mask important demographic disparities. Despite computational limitations, the project establishes a strong foundation for future work. Replicating the analysis using techniques that more closely resemble real-world threat scenarios would further enhance the practical relevance of the findings.

This paper aims to contribute to the prevention and response to one of the AI-related harms that is disproportionately affecting the most marginalized women: deepfakes of non consensual sexual imagery. While this is very important work, as you move forward in your research, my main suggestion is to incorporate in your team social scientists, human rights experts/defenders, and/or link with Mexican organizations already working on response to this type of gender-based violence. This will both strengthen your analytical work, and help you make sure that your research is useful for those already working in this area.

While the methodology of the paper is technically solid, I have two concerns:

* First, the need to problematize the concept of "race" and differentiate it from ethnicity,

* Second, the fact that the categories that are used as classifications (Middle Eastern, Indian, Latino...) were not developed in Mexico or Latin America, and therefore do not reflect the diversity of our countries in a way that is significant for us.

For instance in Table 1: "Latino" may be an ethnic group in the US, but what does it mean in Latin America where we have people of many different characteristics and identities? Which of the categories included there are supposed to represent the Indigenous populations from our countries? Even for the US, the category "Latino" may not be the most appropriate for the purpose of face identification. Similarly, what do you mean by Western features? Probably "whiteness". While the work already presented in your paper shows important results, this theoretical framework would much benefit from being modified as you move forward, including by reviewing research in this area and census data done in our own countries.

I also suggest you review the work that has been carried out by agencies such as UNFPA to incorporate international standards and concepts, for instance those of "technology-facilitated gender based violence" and "non-consensual sexual imagery using deepfakes".

I congratulate you for investing your time and effort in this topic that affects so many girls and women around the world. Working alongside experts in gender-based violence and social studies will make your research more solid and with more potential to have an impact that is much needed in Mexico and elsewhere.

What makes this research so crucial is how it dismantles the myth of a "universal" safety standard. It proves that we cannot blindly import black-box moderation models and expect them to protect all demographic groups equally. When a victim of non-consensual intimate imagery seeks recourse, their fundamental right to protection should not depend on the specific, un-audited biases of the detector a platform happens to use. The main findings solidify a critical governance mandate: algorithmic fairness cannot be assumed or inherited. For platforms operating in the Global South, relying on standardized, Western-centric safety benchmarks is insufficient and negligent. We need mandatory, deployment-specific auditing for any system acting as a gatekeeper for content moderation and victim protection. This is exactly the kind of concrete, disaggregated evidence required to shift AI policy from theoretical ethics to enforceable liability.

On impact, I would say that, although the problem is well-defined, the scope is quite limited, considering the grand scheme of AI Safety related challenges, although a fair one nonetheless. The technique the article use to address the problem, which is an evaluation of false negative rates of some public deepfake detectors, is not innovative enough so that one could consider how it transfer to these larger problems.

On execution itself, considering the paper on its own terms, I would say that the finding that they consider to not be the central one is the central and more reliable one, which is the unreliability on distributional shift. First, the models advertise general performance metrics that do not hold in the specific application of detecting fake faces, without disclosing exactly what they mean by detecting fake faces. They could defend themselves by saying that they are actually detecting in-the-wild face forgeries, but they are not specifying this claim, instead making a general one that do not hold in a specific application. So, in this sense, the paper finds something that is true. However, what the paper claims to be central I would argue is less central, given the finding that the detectors are already generally unreliable (as deployed at the default threshold) in their false negative rates. As in, conditioned on this finding, the main conclusion is that one should take care to use these in production at all, in particular in fake face detection. It would be different if, say, the detector did show a reliable FNR in general and for a privileged group, that instead does not generalize to other groups.

On the presentation, I felt some things were missing in context, that would strengthen or weaken the claim of relevance: (i) some research in the real world deployment of these deepfake detectors. One should consider that these should be part of a whole pipeline, ideally, and it would be interesting to know how this is applied in practice. I know this may be a bit out of the scope of the actual presentation, but it helps a lot to know the context in which these are applied in the real world. On the technical part, I felt strongly that what was missing was the per-group confidence intervals, not only the global ones. This is because a key proposal of the paper is to rank, for each specific group, their FNR, but they themselves admit that some group counts are too small to be reliable. For instance, for prithiv v1, the highest reported FNR group (black) has only n=12 samples, which the authors themselves flag as unreliable; a per-group confidence interval here would likely overlap most other groups and undercut the ranking the paper builds on.

I do not mean this to discourage the applicants, when saying that overall the problem is small in scope, but instead to help fuel their ambitions to tackle grander problems that may loom their entire country and the Global South, such as disempowerment, loss of control, catastrophic misuse, and so on. Instead, I raise the scope point precisely because I take this team seriously, and would hold them to the same bar as any researchers working on frontier problems. That being said, I will understand if the authors do not agree to my critique, and consider that instead these "small" problems are important to work on either way.