Week 3: Societal Impact

<aside>

🧭 Dashboard · Week 1 · Week 2 · Week 3 · Week 4 · Resource Library

</aside>

Systems Thinking + Capitalism and Power

AI doesn't exist in a vacuum. Every deployment decision is a political and economic act. This week shifts from individual ethics to structural analysis: who benefits, who bears the cost, and what responsibilities come with having the power to deploy.

Deployment Ethics: When is AI ready? When should it never be used?
Labor & Work: Displacement, augmentation, worker surveillance
Environmental Impact: The carbon cost of computation
AI Governance: Ongoing responsibility and walking the walk
AI Literacy & Equal Access: The human-in-the-loop fallacy

Pre-reading

Cory Doctorow on Reverse Centaurs - On acceleration and the human attention gap. Essential framing for this week's systems thinking lens. Full interview is 51 minutes, but link below is to a specific (relevant) topic.

Cory Doctorow on Ensh*tification, The AI Bubble, Reverse Centaurs, and The Post-American Internet

Watch →

</aside>

Policy Design in the AI Era - Sarah Downey, 2025 - AI policy is a change leadership problem disguised as a technology problem. This article introduces the governance gap, the two-layer policy structure, and why most AI policies fail before they're implemented. Essential pre-reading for the AI and Change: Governance section.

AI Policy Design for Nonprofits: A Guide by Sarah Downey Consulting

Read →

</aside>

How Long Are Privacy Policies? - A data visualization on the impossible length of policy documents - and what it means for meaningful consent and the human-in-the-loop fallacy.

A Policy Length Analysis for 70 Digital Services - The Biggest Lie on the Internet

Read →

</aside>

Foundation Model Transparency Index - December 2025 - Stanford CRFM - The current state of transparency across major foundation model providers. Useful framing for the deployment ethics discussion.

Foundation Model Transparency Index

Explore →

</aside>

How much energy does Google’s AI use? - Google Cloud - The current state of transparency across major foundation model providers. Useful framing for the deployment ethics discussion.

Measuring the environmental impact of AI inference | Google Cloud Blog

Explore →

</aside>

Course Content

Deployment Ethics: When Is AI Ready?

<aside>

01-deployment-readiness-when-is-ai-ready-for-the-world.png

Building on the risk-based frameworks covered in previous weeks, this section asks what responsibility looks like in practice — for those designing, deploying, or integrating AI systems that affect real users and real outcomes.

At minimum, three evaluation criteria should be applied to any AI system you engage with, deploy, integrate into a product, or develop yourself.

1. Privacy and Information Flow

Rigorous management of how data moves through AI systems — understanding training data composition where available, how it shapes output, and how bias propagates and compounds through reasoning chains. This is especially critical when the output of one automated system feeds as input into a downstream process, where the risk of amplification intensifies. Understanding system-level impacts arising from information flows is a foundational responsibility for every AI system you touch.

2. Counterfactual Evaluation

For any risk you can articulate, you should also be able to systematize its measurement. If you suspect a system treats individuals differently based on gender markers in textual references, you can test for that. Systematic monitoring of counterfactuals is how you move from intuition about risk to empirical evidence — and from anecdote to accountability.

3. Moral Responsibility to Understand Before You Build

Design systems from a position of understanding — whether academic or intuitive — of the full range of possible outcomes. What behaviours are possible? How do you constrain them? How do you test for and monitor them? Do not hand over the keys without understanding what you are handing over. This technology confers immense power, and with that, immense potential to cause harm.

When is AI ready? The question of readiness is not purely technical — it is organizational, political, and contextual.
When should it never be used? High-stakes domains where error is catastrophic, irreversible, or discriminatory.
Key considerations: consent, explainability, auditability, reversibility, and human oversight.

</aside>