Two years ago I had the pleasure of working alongside my supervisor and a fellow PhD student to develop and publish a typology of risk regarding generative text to image models. I think it is time to reflect on the changes that have occurred since we wrote the paper 🙂 (https://arxiv.org/pdf/2307.05543)

Perhaps the most significant development since our paper has been the regulatory response. When we wrote about these risks, policymakers were largely still scrambling to understand what DALL-E and Midjourney even were. Fast forward
to 2024, and the various comprehensive legal frameworks have been developed across both the EU and the US. The EU response explicitly addresses generative AI,
requiring that AI-generated content be identifiable and mandating transparency obligations. The fact that “deepfakes” now has its own regulatory category speaks volumes about how quickly the discourse has matured.

In the US, the response has been more fragmented but still substantial. The FCC ruled that AI-generated robocalls fall under existing regulations, and 19 states have now enacted laws restricting deepfakes in political campaigns. The
Copyright Office clarified in January 2025 that AI-generated outputs only qualify for copyright if humans provide sufficient creative input, mere prompts don’t count.

One area where our concerns proved both validated and overblown was electoral misinformation. We worried about synthetic media being weaponised to undermine democratic processes. The 2024 global election cycle was widely predicted to be the “deepfake election.”

The reality was more nuanced. Yes, there were incidents, but analysis
found that “cheap fakes” without AI were used seven times more often than AI-generated content for election misinformation. The apocalypse didn’t quite arrive—though whether this reflects effective countermeasures or simply
attackers still learning the tools remains unclear.

Yet, what troubles me is how many of the gaps we identified persist. Bias in image generation systems remains pervasive. The environmental costs of training these models continue to climb. The labour exploitation underlying training
data, from uncredited artists to traumatised content moderators, hasn’t been meaningfully addressed. The International AI Safety Report 2025, produced by 96 experts across nations, notes that general-purpose AI capabilities have increased markedly since 2023, with models showing dramatically better performance in programming and scientific reasoning. Yet our frameworks for governing these systems remain reactive rather than anticipatory.

Two years on, I guess I am cautiously optimistic. The conversation has shifted from “should we regulate?” to “how do we regulate effectively?” Industry is engaging with safety research in ways that felt aspirational in 2023. But the gap
between identifying risks and mitigating them remains vast. Our typology was meant to be a starting point, not a destination. The work continues…

Sources:

• https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
• https://ash.harvard.edu/articles/the-apocalypse-that-wasnt-ai-was-everywhere-in-2024s-elections-but-deepfakes-and-misinformation-were-only-part-of-the-picture/
• https://www.npr.org/2024/02/08/1229641751/ai-deepfakes-election-risks-lawmakers-tech-companies-artificial-intelligence
• https://knightcolumbia.org/blog/we-looked-at-78-election-deepfakes-political-misinformation-is-not-an-ai-problem
• https://internationalaisafetyreport.org/publication/international-ai-safety-report-2025/
• https://research.aimultiple.com/generative-ai-copyright/