> ## Content Index > Fetch the complete content index at: https://www.aigl.blog/llms.txt > Use this file to discover other available public pages before exploring further. # AI Governance Library Newsletter #3: Show, don't tell - URL: https://www.aigl.blog/ai-governance-library-newsletter-3-show-dont-tell/ - Published: 2025-04-24T12:02:33.000Z - Updated: 2025-05-01T06:00:50.000Z - Description: This week we look at how visual design can make or break AI governance, review the first real playbook for agent oversight, and highlight why Tom Scott’s YouTube channel might teach you more about systems failure than most compliance workshops ever could. - Author: Jakub Szarmach - Tags: Newsletter There’s a strange thing happening in AI governance. We talk about transparency, clarity, and accountability—then publish 50-page PDFs with unreadable diagrams and footnotes that feel like riddles. Visuals shouldn’t be decoration. In governance work, they’re **core infrastructure**. The best visual communication doesn’t just look clean—it **helps people think**. It gives structure to ambiguity. It shows what words can’t. It earns its place on the page. Here’s what that looks like in practice—and where we’ve seen it done well: 🔹 **It makes roles legible.** The [**Bird & Bird Guide to the EU AI Act**](https://www.aigl.blog/european-union-artificial-intelligence-act-a-guide/) does this well—its chapter structure mirrors regulatory roles. You don’t get lost in definitions of “provider” vs. “deployer” because the document *shows* where responsibilities live, how they move, and who holds them. 🔹 **It invites re-use.** A great table or matrix shouldn’t be something you scroll past—it should be something you screenshot. The [**MIT AI Risk Framework (AI-RAF)**](https://www.aigl.blog/mit-ai-risk-analysis-framework-ai-raf/) does this perfectly. It’s not just readable—it’s interactive. You can filter by sector, risk type, uncertainty. It’s built to be borrowed, cited, and embedded in actual workflows. 🔹 **It shows control in motion.** The [**AI Governance Controls Mega-map**](https://www.aigl.blog/ai-governance-controls-mega-map-feb-2025/) aligns over 60 controls from global frameworks. But it’s not just a list. It groups, compares, and tags them across lifecycle stages. You can trace a single principle—like transparency—across ISO, NIST, the EU AI Act, and UNESCO in a single glance. 🔹 **It surfaces real distinctions.** The [**AI Auditing Checklist**](https://www.aigl.blog/ai-auditing-checklist-for-ai-auditing/) doesn’t just list tasks—it lays out when they apply and who’s responsible. The visual framing of “System vs. Organisation vs. Context” helps readers **think like an auditor**, not just read like one. 🔹 **It simplifies layered processes.** The [**AI Impact Assessment Template**](https://www.aigl.blog/ai-impact-assessment-the-tool-for-a-responsible-ai-project/) is one of the few that *actually* helps teams start from scratch. It breaks down steps visually, shows input/output flow, and connects stakeholder mapping to risk severity without losing the thread. If you’ve worked in policy, you’ve probably read five frameworks this year that sounded great—but couldn’t show their own structure. That’s a red flag. Governance is communication. And good visual design isn’t optional. It’s how we make accountability traceable, decision rights visible, and obligations real. --- ## **🔍 Spotlight Review:** [**AI Agent Governance – A Field Guide**](https://www.aigl.blog/ai-agents-governance-a-field-guide/) **By Jam Kraprayoon, Zoe Williams, Rida Fayyaz (April 2025)** ### **Quick Summary** One of the first governance resources built specifically for autonomous AI agents. This field guide doesn’t treat agents as a technical feature—it treats them as a distinct policy challenge. It lays out the risks, benchmarks current performance, and proposes a five-part governance strategy. ### **What’s Covered** This guide takes AI agents seriously—as systems that act, plan, and adapt with minimal human input. The framing is clear: these aren’t theoretical. Agents are already in use at Google, Salesforce, Klarna and others, handling everything from customer support to software development to autonomous web interaction. It opens with two futures: One where agents boost productivity across sectors. And one where they replicate unpredictably, accumulate power, and outpace oversight. From there, the report gets concrete. It walks through agent benchmarks across six domains, showing where they succeed (short tasks) and where they still fail (open-ended goals, persistent context). Still, their performance is accelerating—task duration capabilities are doubling roughly every seven months. At its core is a five-part governance taxonomy: - **Alignment** – tuning behavior, modeling risk attitudes, paraphrasing outputs - **Control** – rollback tools, interrupt switches, fine-grained shutdown mechanisms - **Visibility** – agent IDs, persistent activity logs, observable internal states - **Security & Robustness** – sandboxing, attack surface reduction, adversarial testing - **Societal Integration** – equitable access, liability schemes, structural power limits This isn’t just about compliance. It’s a guide for anyone designing systems that act *on their own*—and a warning that governance tooling isn’t keeping up. ### **💡 Why It Matters** Too many current frameworks still treat agents as just another use case for large models. This guide flips that. It shows why autonomy, memory, and planning demand new thinking. Without dedicated governance, agents risk becoming the most opaque, misaligned, and unsupervised actors in digital infrastructure. ### **What’s Missing** Implementation remains thin. The report is strong on framing, but light on how current laws like the EU AI Act might adapt. There’s little guidance for procurement, enforcement, or what readiness looks like in the public sector. Societal risk is acknowledged—but left unresolved. ### **Best For** - Governance professionals exploring agent oversight - Policymakers drafting future agent-specific regulation - Technical teams designing agents with real-world goals - Researchers modeling human-agent interaction risks - Funders looking to define the agent governance research agenda ### **Source Details** **Title:** AI Agent Governance: A Field Guide **Author:** Jam Kraprayoon, Zoe Williams, Rida Fayyaz **Publisher:** Frontier Security **Published:** April 2025 ![](https://storage.ghost.io/c/44/95/449506ca-034e-480f-9725-fcde08ef1cc1/content/images/2025/04/image-2-1.png) ## **🌙 After Hours** Quiet shifts, long-range signals, and one YouTube channel that teaches systems thinking better than most frameworks. 📺 **The Most Unexpectedly Relevant Channel in Governance: Tom Scott** [Watch on YouTube](https://www.youtube.com/@TomScottGo/featured?ref=aigl.blog) Tom Scott is wrapping up his time on YouTube after over a decade of weekly videos—and if you’ve never watched his channel, now’s the time. He doesn’t talk about AI. He doesn’t talk about law. What he *does* talk about: - Why bureaucratic systems fail in subtle ways - Why infrastructure gets abandoned - What happens when no one owns the problem - How constraints shape decisions in code, cities, and communication Each video is 3–6 minutes. In a field obsessed with what’s cutting-edge, Tom Scott explains the **mundane things that quietly shape everything else**—and he makes it stick. If you work in policy, risk, safety, or oversight, this channel is a reminder: Most real problems start small, seem boring, and get ignored until they’re irreversible. Start anywhere. You’ll learn something. And you’ll probably steal his framing in your next report. --- 📉 **The Slow Drift Begins: US Science Brain Drain** [Read in Nature](https://www.nature.com/articles/d41586-025-01216-7?ref=aigl.blog) Here is a sentence I would never expect to write. US is showing signs of a Brain Drain. It hasn’t made headlines yet, but it’s starting to show. Researchers are leaving the US at a slow but steady pace—pushed by political interference, immigration limits, and uneven funding. The consequences won’t be obvious right away. But if you’re tracking where the next generation of AI governance, safety, or standards research will happen—this matters.