In November 2022, OpenAI released ChatGPT to the public. Within five days it had one million users. Within two months, one hundred million. The product team had expected a modest research preview. Instead, an unprecedented wave of ordinary people began stress-testing the system — finding failure modes, surfacing biases, and flooding social media with screenshots of problematic outputs. That public pressure directly influenced OpenAI's subsequent content policy revisions and the speed of its safety mitigations. No single researcher caused this; millions of everyday users collectively shaped the trajectory of the most widely-deployed AI system in history.
Earlier, in 2018, Google employees — not executives, not regulators — circulated an internal petition signed by roughly 3,700 workers demanding the company withdraw from Project Maven, a Pentagon contract to use AI for drone imagery analysis. Google announced it would not renew the contract that June. Individual engineers, exercising voice inside an institution, altered a major AI deployment decision.
A common assumption is that AI safety is a closed technical problem, solvable only by people with machine learning PhDs working inside frontier labs. This assumption is wrong in at least three important ways.
First, many of the hardest questions in AI alignment are not purely technical — they concern values, priorities, and acceptable trade-offs. Should an AI system prioritize individual privacy or aggregate safety? Should it defer to the user or to the developer? These are questions democratic societies are equipped to answer, not just engineers.
Second, AI systems learn from human-generated content and human feedback. Every time a person rates an AI response, corrects a model's error, flags a harmful output, or writes publicly about their experience with an AI system, they contribute signal that shapes future behavior.
Third, frontier AI development happens inside companies that respond to reputational pressure, regulatory threat, talent recruitment, and consumer choice. Each of these levers is operated, at least in part, by ordinary people.
Reinforcement Learning from Human Feedback (RLHF), the technique central to making ChatGPT helpful and less harmful, was refined substantially using ratings from contractors — many of them non-specialists — who judged which AI responses were better. Individual human judgments, aggregated, become the training signal. Your evaluation of AI outputs is literally part of how the technology gets built.
Researchers at the Center for Human-Compatible AI (CHAI) and the Future of Life Institute have informally categorized the ways non-specialist individuals influence AI development. Four categories emerge consistently:
The AI systems deployed over the next decade will reflect the values of the society that produced them. Whether that society is attentive or passive, informed or ignorant, organized or fragmented — matters enormously. Individual choices about attention, voice, and civic participation aggregate into collective outcomes.
In this lab you will discuss with an AI advisor how the four categories of individual leverage apply to your own situation — your profession, your civic context, and your daily AI use. The goal is to identify at least two concrete, realistic actions you could take.
Before GPT-4 launched in March 2023, OpenAI engaged hundreds of external "red teamers" — including domain experts in biosecurity, law, medicine, and education — to probe the model for dangerous capabilities and failure modes. The red-teamers were not all AI researchers. A significant portion were subject-matter experts who understood the domains where AI errors could be most harmful. Their reports directly shaped the safety mitigations included in the public launch. OpenAI published a technical report documenting this process, noting that red-teamers identified failure modes the internal team had not anticipated.
Separately, in 2023 Anthropic published its Constitutional AI methodology in detail, partly to allow outside scrutiny. The paper explicitly acknowledged that the choice of which principles to include in the "constitution" is a normative question that the company alone should not decide — inviting broader societal input into what values AI systems should embody.
Every major AI lab now maintains some form of public feedback channel. The quality and impact of these channels varies enormously, but they exist — and using them is categorically different from saying nothing.
The National Institute of Standards and Technology's AI Risk Management Framework (AI RMF 1.0), published in January 2023, incorporated feedback from over 240 organizations and hundreds of individuals during its drafting process. The public comment process ran for months and substantively shaped the document. NIST guidance is not legally binding, but it is widely adopted as a de facto standard by both industry and state-level regulators.
Generic complaints rarely change anything. Useful feedback shares specific characteristics that make it actionable for developers and policymakers.
Feedback channels are only as good as the institutional will to act on them. Several documented problems constrain their effectiveness: companies are not legally obligated to share what they do with submitted feedback; red-team findings are sometimes overridden by product timelines; and public comment processes can be dominated by well-resourced industry actors who submit voluminous technical comments.
Knowing these limits is not a reason to disengage. It is a reason to combine feedback with other levers — civic action, professional organizing, and support for mandatory disclosure requirements — that create accountability structures feedback alone cannot provide.
Recall an AI interaction you have had that produced an output you found problematic — misleading, harmful, biased, or simply wrong in an important way. If you cannot recall one, the advisor will provide a scenario. Then work with the advisor to draft feedback that meets the standards covered in Lesson 2.
The European Union AI Act, finalized in 2024, is the world's first comprehensive AI regulation. Its drafting took years and involved thousands of stakeholder submissions, parliamentary hearings, and public consultations. Several provisions — including the ban on real-time biometric surveillance in public spaces and requirements for transparency about AI-generated content — were substantially shaped by civil society organizations representing ordinary citizens. The algorithmic accountability nonprofit AlgorithmWatch, the digital rights group EDRi, and others filed detailed technical submissions that were cited in committee reports.
In the United States, when the FTC opened comment periods on AI and data practices in 2022 and 2023, individual citizens submitted tens of thousands of comments. While industry comments dominated in technical detail, the sheer volume of public comments documenting real harms from AI systems — discriminatory hiring algorithms, manipulative recommendation systems, predatory credit scoring — was cited in subsequent agency guidance documents.
AI policy is being made continuously, not in a single dramatic moment. The relevant venues include regulatory agencies, legislative committees, international standards bodies, and court decisions. Each has different access points for public participation.
Several major AI policy questions are being actively debated in 2024–2025, meaning public input is most valuable right now — before positions calcify into law:
California's Senate Bill 1047, which would have required safety evaluations of large AI models, passed the legislature in 2024 before being vetoed by Governor Gavin Newsom. The bill's drafting, debate, and veto were all substantially influenced by organized advocacy — from AI safety researchers supporting the bill to AI industry groups opposing it. Individual constituent calls and emails to the Governor's office were a documented part of the advocacy effort on both sides. The episode illustrates that state-level legislation is a real arena where organized public voice matters.
A single letter to a legislator rarely changes a vote. Organized, sustained engagement — especially coordinated through civil society organizations with policy expertise — is more effective. The most realistic individual contribution is: stay informed, support organizations doing this work, contact representatives consistently on specific bills, and participate in public comment periods with substantive rather than boilerplate submissions.
Choose one of the open AI policy questions from Lesson 3 — mandatory safety evaluations, algorithmic transparency, AI in hiring, deepfake disclosure, or liability for AI harm. Work with the advisor to draft a short but substantive public comment on that question, as if submitting to a regulatory body.
The AI safety information landscape in 2024 is genuinely difficult to navigate. On one side: breathless hype about AGI arriving next year and existential risk requiring immediate radical action. On the other: dismissive claims that alignment concerns are science fiction and anyone worried is naive. Both extremes distort the real picture. The Overton window of mainstream AI commentary has expanded dramatically since 2022, but quality varies enormously — peer-reviewed research, well-reasoned blog posts, journalistic investigation, advocacy content, and pure speculation all circulate on the same platforms and are difficult for newcomers to distinguish.
In 2023, a group of researchers at MIT published a study examining how AI literacy affected people's responses to AI-generated misinformation. The finding: people with moderate AI knowledge were sometimes more susceptible to confident-sounding AI misinformation than those with very low knowledge — because they trusted their own ability to evaluate it. Calibrated skepticism, not just knowledge accumulation, is the actual goal.
The following sources are maintained by organizations with documented track records of technical accuracy and intellectual honesty. This is not a complete list, and no source is infallible — but these are reasonable starting points for building an informed picture.
Individual engagement with AI safety does not have to be all-or-nothing. The following represents a realistic spectrum from minimal to substantial commitment, each level building on the last:
The career advice organization 80,000 Hours, which focuses on high-impact career paths including AI safety, reports that its career guide and one-on-one advising have influenced thousands of professionals to shift toward AI safety roles or to incorporate safety considerations into existing careers. Several researchers now at Anthropic, DeepMind, and independent AI safety organizations cite 80,000 Hours resources in their career narratives. Individuals deciding where to direct professional effort is a significant, compounding influence on the field's talent composition.
Several patterns undermine otherwise well-intentioned individual engagement with AI safety:
The trajectory of AI development is not predetermined. It will reflect the aggregate of many individual choices — about what to build, what to deploy, what to tolerate, what to demand, and what to support. Your choices are part of that aggregate. The question is not whether individuals matter. The question is what you will do with the fact that they do.
Using the commitment spectrum as a framework, work with the advisor to build a personal AI safety plan that is realistic for your schedule and circumstances. The plan should include at least one action from three different leverage categories and a realistic time estimate for each.