This is an important and genuinely difficult issue, and I appreciate the seriousness of the concern.
That said, I think it’s also crucial to be careful about where we focus the discussion, especially in open technical forums.
From my perspective, harm involving minors and synthetic imagery is not primarily a detection problem, even though detection often becomes the most discussed technical angle.
On your first question:
While there is ongoing work around classifiers, forensic signals, and content analysis, history suggests that single-layer detection approaches inevitably enter an escalation loop. As detection methods become more public and more precise, they also become easier to probe, benchmark against, and ultimately bypass. For that reason, many practitioners increasingly view detection as one component in a broader system, rather than a standalone safeguard.
What tends to matter more in practice are multi-layered measures that sit around the model and the content lifecycle: platform-level risk signals, abnormal generation or sharing patterns, fast takedown and reporting pipelines, human review for high-risk cases, and clear legal escalation paths when minors are involved. These measures are less about “catching perfect fakes” and more about reducing real-world harm and amplification.
On your second question regarding provenance and watermarking:
Source attribution and watermarking can help with post-hoc accountability and credibility assessment, particularly for journalists, platforms, and courts. However, they are unlikely to function as a primary protective barrier for minors. Even strong provenance systems do not prevent generation or initial misuse; at best, they help establish context and responsibility after the fact. Treating them as a silver bullet risks overestimating what technical markers alone can achieve.
More broadly, I worry that discussions framed too narrowly around how to detect or how to verify synthetic images can unintentionally miss the larger issue:
Why these systems are accessible in harmful ways, why harmful outputs are able to circulate socially, and why institutional responses (schools, platforms, legal systems) are often slower than the damage itself.
Protecting children and adolescents here likely depends less on ever-more clever technical tricks, and more on clear red lines, strong platform responsibility, rapid response mechanisms, and education that addresses misuse as a form of abuse rather than as a technical curiosity.
In other words, this problem sits at the intersection of technology, governance, and human systems — and over-optimizing the technical layer alone may give us the illusion of progress while leaving the underlying harm pathways intact.