On June 25, 2018, roughly 4,000 Google employees signed an open letter to CEO Sundar Pichai. The subject was Project Maven — a Pentagon contract to use Google's AI to analyze drone surveillance footage. The letter was blunt: "We believe Google should not be in the business of war." Within two months, Google announced it would not renew the contract. No law forced that outcome. No regulator compelled it. Employees did.
The dominant narrative around AI is one of inevitability — that a handful of labs and governments will determine everything, and everyone else will simply receive whatever emerges. That narrative is wrong, and it is wrong in measurable ways.
Consider the timeline of facial recognition. By 2019, Amazon, Microsoft, and IBM were all selling facial recognition services to law enforcement agencies. Then, in June 2020, following sustained pressure from civil liberties groups, researchers publishing accuracy audits, and internal employee campaigns, all three companies announced moratoriums on law enforcement sales. IBM exited the business entirely. The technology did not change. The human pressure around it did.
Or consider content moderation. When researchers at the University of Washington and Distributed AI Research Institute published audits showing that DALL-E 2 and Stable Diffusion generated sexualized images of real people, public pressure — amplified by journalists and policymakers — led to policy changes at OpenAI and Stability AI within weeks. Published evidence, shared publicly, forced platform-level changes.
AI companies are acutely sensitive to three things: talent (employee recruitment and retention), capital (investor confidence), and legitimacy (public trust). Individuals operate in all three arenas simultaneously — as workers, as shareholders, and as members of the public whose trust is being solicited.
Your leverage over AI's future depends on which roles you occupy. Most people occupy several simultaneously:
You do not need to be a technologist, regulator, or billionaire to influence AI's trajectory. You need clarity about which role you occupy, what leverage that role provides, and how to use it deliberately. This module is about developing that clarity.
Participation is not the same as protest, nor is it the same as passive acceptance. The most durable forms of individual influence on AI have been persistent, specific, and evidence-based.
When the Algorithmic Justice League's Joy Buolamwini published research in 2018 showing that commercial facial recognition systems had error rates of up to 34.7% for darker-skinned women versus 0.8% for lighter-skinned men, she did not simply object to bias in the abstract. She produced measurable evidence. That evidence was cited in congressional hearings, in news coverage that reached millions, and in the policy responses that followed. Specificity is power.
This course has equipped you to understand how AI systems work, what they can and cannot do, how they are governed, and where they are headed. That knowledge is not inert. It is the foundation of informed participation — and informed participation is what separates consequential engagement from noise.
In this lab, you'll work with an AI assistant to map your personal leverage over AI development. Think honestly about your situation — your profession, your consumer choices, your civic engagement. The goal is to move from abstract concern to concrete, role-specific action.
On March 22, 2023, the Future of Life Institute published an open letter calling for a six-month pause on training AI systems more powerful than GPT-4. Within days, over 27,000 people had signed — including Elon Musk, Yoshua Bengio, and Stuart Russell. The letter generated enormous media coverage. But it also revealed a problem: many signatories, and far more observers, did not understand the technical specifics well enough to evaluate what a "pause" would actually accomplish, what its risks were, or whether it was even enforceable. Being alarmed is not the same as being informed.
AI generates more media coverage than almost any other technology sector, and most of it falls into two failure modes. The first is hype — stories about capabilities that are either extrapolated far beyond what systems actually do or are simply fabricated by companies' PR operations. The second is panic — catastrophizing narratives that conflate current systems with speculative futures, often for engagement purposes.
Neither hype nor panic is useless — they signal where attention and investment are flowing, which matters. But neither is a substitute for understanding. The people who most effectively shaped AI policy in the past decade — Joy Buolamwini, Timnit Gebru, Yoshua Bengio, Gary Marcus — share a common trait: they read primary sources and published specific, evidence-based claims.
A primary source is a research paper, a company's published technical report, a regulatory filing, or a public dataset. A secondary source is a journalist's interpretation of those things. Both matter — but most people's AI knowledge consists entirely of secondary sources, often tertiary or worse.
Timnit Gebru, who co-led Google's ethical AI team before her controversial departure in 2020, has consistently emphasized that effective advocacy requires understanding technical constraints. Her 2021 paper "On the Dangers of Stochastic Parrots" (co-authored with Emily Bender, Margaret Mitchell, and Angelina McMillan-Major) was influential precisely because it made specific, defensible claims rooted in linguistic theory and empirical evidence — not vague concerns about AI being "dangerous."
You do not need to read every AI paper. But a sustainable information practice for a non-specialist might include:
When a claim about AI circulates — whether it is "AI will take all jobs in ten years" or "this new model achieves human-level reasoning" — the following questions cut through most noise:
Who is making this claim, and what incentives do they have? A company announcing its own model is a different kind of source than an independent researcher. A think tank funded by an AI company is a different kind of source than an academic institution.
What is the specific, measurable claim? "Human-level performance" on what benchmark? Designed by whom? Evaluated how? GPT-4 scored in the 90th percentile on the bar exam — but it was the multiple-choice section, not the essay section, and it was using the official study guides.
What does the primary source actually say? In 2023, a Goldman Sachs report claimed AI could automate 300 million jobs. The actual report said it could "expose" the equivalent of 300 million full-time jobs to automation — a very different claim, with significant uncertainty attached.
You don't need expertise in everything. You need enough fluency to know when you're being manipulated — by hype or by panic — and enough curiosity to check a primary source before passing a claim along. That combination, practiced consistently, makes you genuinely influential rather than merely loud.
In this lab, you'll practice the skill of source evaluation. Bring an AI claim you've seen recently — from a headline, social media, a conversation — and work with the assistant to break it down: Who made it? What exactly does it claim? What would you need to check?
In May 2023, a New York lawyer named Steven Schwartz submitted a legal brief citing six precedents — all fabricated by ChatGPT. When opposing counsel pointed out the citations didn't exist, Schwartz admitted he had not verified them. He was sanctioned by the court. His partner, Peter LoDuca, who signed the brief without reviewing the AI-generated content, was also sanctioned. Both had used a powerful tool without understanding what it was actually doing. The judge in the case called it "an unprecedented circumstance."
The Schwartz case is not primarily a story about AI's limitations — it is a story about a professional who did not maintain the oversight responsibilities that came with his role. AI language models are designed to produce fluent, confident-sounding text. They are not designed to be accurate. The gap between fluency and accuracy is where most AI-related errors happen.
This applies far beyond legal briefs. When a journalist at CNET used AI to draft financial explainer articles in late 2022 and early 2023, more than half contained factual errors. When the errors were discovered, CNET issued corrections — but the original articles had already been read by millions. When Stack Overflow temporarily banned AI-generated answers in 2022, it was because the answers looked correct while being wrong at a rate that degraded the platform's usefulness.
The consistent pattern is the same: AI produces output that looks authoritative but requires human verification to be reliable. This is not a temporary problem waiting to be solved in the next version. It is structural — it follows from how language models work.
Any AI-generated content that will influence a decision, be shared with others, or carry your professional reputation requires independent verification of specific factual claims. Not spot-checking. Verification.
The most effective pattern for using AI in professional and creative work treats the system as an amplifier of your existing knowledge and judgment — not a substitute for it. This framing has practical consequences.
When GitHub Copilot was studied in 2022 by researchers at Stanford, they found that developers who used it produced code with security vulnerabilities more often than those who did not — but this was specifically true for developers who used it passively, accepting suggestions without critical review. Developers who used it as a starting point for their own judgment did not show the same vulnerability rate.
The amplifier frame means: you bring the domain knowledge, the judgment, and the accountability. AI brings speed and scale. When those roles get reversed — when AI is author and you are the approver — errors compound and accountability disappears.
Using AI tools thoughtfully also means understanding what you're feeding into them. In 2023, Samsung engineers inadvertently leaked proprietary source code and internal meeting notes by submitting them to ChatGPT as prompts. OpenAI's terms of service at the time stated that inputs could be used to improve the model.
Before inputting information into any AI system, three questions are worth asking: Who owns this data? (Is it yours to share?) Where does it go? (What do the terms of service say about storage and use?) Could this harm anyone? (Does it contain information about people who haven't consented to it being shared with a third-party AI company?)
These are not exotic privacy concerns. They are the basic due diligence that professionals in law, medicine, and business routinely apply to other information-sharing decisions — and that AI tools now require as well.
Thoughtful AI use is not about using less AI or more AI. It is about maintaining clear accountability for outputs, verifying factual claims you act on, treating AI as an amplifier of your judgment rather than a replacement for it, and applying the same data-handling standards to AI tools that you apply to everything else. These habits are what separate effective AI users from those who create problems — for themselves and for others.
In this lab, you'll conduct a short audit of your own AI use with the assistant's help. Think about the AI tools you currently use — at work, at home, for creative or research purposes. Where are you applying verification? Where are you accepting output passively? Where might data-sharing issues exist?
Between March and June 2023, the US Copyright Office received more than 10,000 public comments on its inquiry into AI and copyright. The European AI Act had already received significant public input during its drafting process. In California, Senate Bill 1047 — a sweeping AI safety bill — was shaped substantially by public debate, testimony from civil society organizations, and eventually Governor Gavin Newsom's veto, itself citing concerns raised by researchers and smaller companies. None of these processes required professional credentials to participate in. They required only engagement.
AI governance is not a single system. It operates simultaneously at the level of corporate policy (what companies choose to do), industry standards (what bodies like IEEE or ISO specify), national regulation (laws passed by legislatures and rules issued by agencies), and international coordination (agreements between governments). Each level has different entry points for public influence.
The most accessible entry point for most people is the regulatory comment process. In the United States, when a federal agency proposes a rule — including rules touching on AI — there is a mandatory public comment period during which any person can submit views. These comments are legally required to be read and considered. The Federal Trade Commission, the Copyright Office, the Department of Commerce, and NIST have all issued AI-related requests for comment in recent years.
Comments do not need to be long or technically sophisticated. A specific, factual comment about how a proposed rule would affect a particular use case carries real weight. Comments that say "I am concerned about AI" are less useful than comments that say "I am a radiologist, and this rule as written would create the following specific problem for medical AI applications."
In the US, regulations.gov aggregates public comment opportunities across all federal agencies. State-level processes vary. In the EU, the European Commission's Have Your Say portal serves a similar function. Most require only an email address to participate.
For most people, the most efficient path to influence is through organizations that aggregate and amplify individual concerns. Several organizations have been consistently effective at translating public concern into policy change:
AI governance is not going to be resolved in a single election cycle or a single regulatory round. The EU AI Act, which represents the most comprehensive AI regulation in the world so far, took four years from initial proposal to passage — and its implementation will take several more years. The US federal approach remains fragmented across agencies. International coordination is nascent.
This means that meaningful engagement is necessarily sustained over time. The people who have had the most consistent influence on AI governance — Joy Buolamwini, Kate Crawford, Safiya Umoja Noble, Gary Marcus — have all maintained presence in the conversation for years, updating their arguments as evidence accumulated, building coalitions, and participating in multiple venues simultaneously.
You do not need to match that scale. But you can identify one area of AI governance where your specific knowledge, experience, or position gives you a distinctive contribution to make — and you can make it persistently, specifically, and in the venues where it will be read by the people making decisions.
This entire course has been preparation for a set of choices you will make over the coming years — about how you use AI, how you talk about it, how you vote on it, and what you demand from the institutions building it. None of those choices requires you to be an expert. All of them require you to be informed, specific, and engaged. You are now more equipped for that than most people in the world. What comes next is up to you.
In this final lab, you'll develop a concrete engagement plan — not abstract intentions, but specific actions tied to your actual roles, knowledge, and available time. The assistant will help you identify the most productive venues for your specific situation and draft a first action step you can take within the next two weeks.