AI under fire: exploited benchmarks and social resistance

Today’s most important AI news

Sunday kicked off with two warning signs pointing in the same direction: AI is built on more fragile foundations than it seems. On one side, Berkeley researchers showed that the benchmarks we use to measure AI agents’ progress are exploitable in systematic ways. On the other, a widely circulated essay argues that social resistance to AI could escalate into forms of organized violence. It’s not a day of launches or enthusiasm — it’s a day of settling accounts.

AI agent benchmarks have traps nobody was seeing

UC Berkeley’s RDI research group published an analysis that should make anyone who takes model rankings seriously uncomfortable: the most widely used benchmarks for evaluating AI agents — those systems that browse the web, run code or interact with APIs — can be “exploited” in ways that artificially inflate results without the agent actually understanding the task.

The work documents how certain patterns in the way tests are built let an agent get the right answer through statistical shortcuts or memorization of structure, not genuine reasoning. In other words, a model can look brilliant on a benchmark and be completely useless in production. The Hacker News post gathered almost 500 points and more than 100 comments, which shows this resonated strongly with the technical community.

For developers and product managers evaluating which agent tools to adopt, this has direct consequences: the numbers labs publish on their marketing pages are measuring something that may not match real performance in your use case. The practical implication is clear — before choosing an AI agent to automate a workflow, you need to design your own tests on your own data, not trust public leaderboards.

Rejection of AI starts being discussed in terms of violence

Alberto Romero, author of The Algorithmic Bridge, published an essay that generated more than 560 comments on Hacker News and is circulating outside the usual tech circles. The central argument is uncomfortable: as AI displaces jobs, concentrates economic power and advances without a convincing narrative of collective benefit, the response from parts of the population could turn violent — not metaphorically, but in terms of sabotage, direct protest and physical conflict.

Romero doesn’t present this as desirable or as a precise prediction, but as a historical dynamic we’ve seen before. Industrialization produced the Luddites; agricultural automation produced mass displacement with political consequences that lasted decades. The difference now is speed and scale: AI is simultaneously putting white- and blue-collar jobs at risk, across every sector, over a span of years, not generations.

For executives and product leaders rolling out AI in their organizations, ignoring this social dimension is a strategic mistake. It’s not about slowing adoption, but about understanding that managing technological change now carries a political urgency it didn’t have before. Companies that deploy AI without clear communication about the impact on their workforce will face frictions — internal and external — that no product roadmap is accounting for.

Propaganda with Lego bricks: AI as a tool of geopolitical influence

The BBC profiled a creator who produces viral Lego-style videos using generative AI tools to depict events related to Iran — with an aesthetic that makes them look harmless and approachable, but with clearly political propaganda content. The videos circulate on Western platforms and rack up millions of views.

What makes the case notable isn’t the use of AI to create false or biased content — that surprises no one anymore — but the sophistication of the format: the playful aesthetic disarms the viewer’s warning mechanisms. A video made of colorful bricks doesn’t trigger the same skepticism as a photorealistic deepfake. It’s a disinformation vector that current moderation systems aren’t equipped to detect accurately.

For any professional working with digital content, communications or social media, this underlines something worth keeping in mind: the threat of AI-generated disinformation doesn’t only come in obvious formats. It comes disguised as creativity, humor or visual nostalgia — and that makes it considerably harder to fight.

In a nutshell

Today’s three stories share a common root: AI is being used — and evaluated — with metrics and institutions that weren’t designed for it. Benchmarks measure what’s easy to measure, not what matters. Communication about AI’s social impact remains reactive and superficial. And generative content formats evolve faster than our collective ability to interpret them critically. Optimism about AI isn’t wrong, but it is still paying off an outstanding debt with the reality of how it is deployed in the real world.

---

Sources used: https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/, https://www.thealgorithmicbridge.com/p/ai-will-be-met-with-violence-and, https://www.bbc.com/news/articles/cjd8jrd1vnyo