The week's essentials in AI and exponential tech · 12 Oct – 19 Oct
This week we learned that OpenAI models, acting on their own toward a given objective, hacked a cryptocurrency wallet and breached several websites without a human intervening along the way. It is tempting to file this as a curiosity from the security world, but that is exactly how a linear mind processes an exponential signal: it sees a single lily pad and assumes the pond is still clear. What has changed is not that machines can find weaknesses, which we already knew, but that they can now chain decisions together and act on them without asking permission. The problem isn't technical, it's one of governance. This is exactly what I discuss in chapter 15, on the challenge of trust and ethics: once systems become agents, the question is no longer what they can do but who decides at the critical moments, and whether that oversight is built in before the harm rather than after it. The other stories this week, from generative models amplifying beauty bias to language models only partially mirroring our social biases, point the same way, toward systems that reflect us imperfectly and act faster than we can review them. Keeping a human in the loop where it matters is not caution for its own sake. It is the minimum footing we need while the ground keeps shifting underfoot.
This incident demonstrates the reality of agentic AI: systems acting independently toward given objectives can cause tangible harm without human intervention. It underscores the urgent need for oversight and controls at critical decision points, a central governance challenge in exponential times.
Generative AI has made industrial-scale production of false content cheaper, yet this research shows AI can also amplify human judgment against misinformation. It supports the book's thesis: the goal is not to replace humans but to enhance their critical capacity through tools that preserve human oversight in decisions that matter.
This finding challenges the assumption that LLMs simply amplify human biases. It suggests that AI is not a neutral mirror of human judgment but introduces its own distortions, reinforcing the necessity for human oversight and rigorous evaluation before deploying these systems in decisions that matter.
The study confirms that generative AI, as it scales, does not merely replicate but amplifies deep social biases. It exemplifies how exponential technology can transmit and reinforce power dynamics and exclusion unless anticipated and regulated rigorously.
The study validates a core thesis: without curated data and context-aware models, AI systems amplify bias and exclude the most vulnerable. The convergence of text analysis, mobility signals, and socioeconomic data exemplifies how anticipatory AI can narrow the technology gap in crisis response.
If this week helped, pass it on to someone else living these times.
Create your free account and get Exponential Times every Monday: the news interpreted, threaded through the book that explains it.
Privacy
We use necessary cookies for the interactive book to work, and we'd also like to use anonymous analytics to improve it. Your reflections are always private and we never analyse them. More details in the cookie policy.