AI's Impact on Society

Learning Objectives

• Analyze AI's economic impact through the lens of GDP growth, job displacement, and wealth distribution
• Explain how AI-generated disinformation and synthetic media threaten democratic processes
• Articulate the privacy risks of AI surveillance systems and the intellectual property debates reshaping creative industries

The Scale of What We Are Navigating

The previous modules examined AI from the perspective of the learner — what tools to use, which models to understand, how to future-proof your career. This module steps back to examine AI's impact at the level of society itself.

The patterns we are observing with AI do not fit neatly into "good" or "bad" categories. They are complex, contested, and unfolding in real time. The goal here is not to make you alarmed or reassured — but to give you an honest, evidence-grounded understanding of the forces at work.

Economic Impact: Growth, Displacement, and Distribution

The Productivity Upside

The economic case for AI is substantial and empirically grounded:

• McKinsey Global Institute estimates AI could add $13–17 trillion to global GDP by 2030, driven primarily by productivity gains across professional services, manufacturing, and healthcare
• Accenture estimates AI could boost labor productivity by up to 40% in specific sectors by 2035
• OECD research shows early AI adopters are seeing 15–30% productivity gains in knowledge-intensive tasks like software development, legal research, and content creation

These are real economic gains — more value produced with the same or fewer inputs.

The Displacement Risk

The same technology that generates productivity gains creates displacement:

• Goldman Sachs (2023): Estimated 300 million jobs globally are "exposed to automation" — meaning AI can perform a significant portion of the tasks in those roles
• WEF Future of Jobs Report (2025): Projected 170 million new jobs created and 92 million displaced over 2025–2030 — a net gain of 78 million roles globally. However, the displacement is concentrated in administrative, clerical, and routine knowledge work, while new roles skew toward technology, data, and AI-adjacent functions
• The transition is not symmetric: new jobs created by AI (AI engineers, AI trainers, AI product managers) require different skills than jobs displaced by AI (data entry, basic content creation, routine document processing)

Public Sentiment Is Cooling — Especially Among Gen Z

The first cohort to enter the workforce in an AI-saturated world is also the most skeptical of it. The Walton Family Foundation's April 2026 Gallup Panel survey of 1,572 Americans aged 14 to 29 found that enthusiasm has stalled and concern has risen:

• 31 percent of Gen Z now report feeling outright anger toward AI, up from 22 percent a year earlier
• 48 percent believe the workforce risks of AI outweigh its benefits, an 11-point year-over-year increase
• Fewer than 20 percent of Gen Z would choose AI over a human for tutoring, financial advice, or customer service
• Roughly half still use AI weekly, but adoption growth has flattened — the trend line is sentiment shifting against the technology, not usage falling away
• 74 percent of schools now have AI policies, but students remain skeptical of classroom integration and report widespread perceptions of academic dishonesty among peers

The institutional response is now visible at the most selective US universities. In May 2026, Princeton's faculty voted — with one dissenting vote — to mandate proctoring for all in-person exams starting July 1, 2026, ending 133 years of unproctored testing under the student honor code. The proposal explicitly cited "AI and personal electronic devices as major catalysts," noting that AI tools on small devices make cheating much harder for peers to observe, which had hollowed out the peer-reporting model the honor system relied on. The Daily Princetonian's 2025 Senior Survey of more than 500 graduating seniors found 29.9 percent admitted cheating on an assignment or exam, 44.6 percent said they knew of violations they did not report, and only 0.4 percent said they had ever reported a peer to the Honor Committee. The Princeton case is the clearest signal yet that elite institutions are giving up on self-policing in an AI-saturated student environment.

The counter-case arrived weeks later, and it is the more instructive one. Mexico's National Autonomous University of Mexico, known as UNAM, ran its undergraduate entrance exam fully online for the first time in mid-2026, using a locked-down browser and behavior-monitoring software that watched webcam feeds for wandering eyes, extra people in the room, and background conversation. The results were unlike anything in the university's history: about 16 percent of candidates scored 100 or more correct answers out of 120, against roughly 3 percent across the preceding five years. A technical commission concluded the results could not be trusted, and roughly 58,000 applicants — about 36 percent of the field — were recalled to sit a control exam in person, under human supervision. Princeton and UNAM reached the same destination from opposite directions: one abandoned the honor code in favor of proctoring, the other tried algorithmic proctoring and fell back to a room with people in it. Note what actually caught the problem at UNAM — not the surveillance software, but the score distribution.

Both of those are universities responding to assessment failure. The third case is a school district acting before one. On September 2, 2026, New York City — the largest school system in the United States — imposed a one-year moratorium on classroom AI use through eighth grade, covering roughly 600,000 students, banning chatbots built for companionship in every grade, and barring teachers from using AI to grade work. Mayor Zohran Mamdani framed it as a developmental argument rather than an integrity one: "Children need teachers and human connection in order to learn and grow. They need to develop skills alongside their peers, build relationships with educators, and wrestle with tough problems on their own."

The design of the policy is more interesting than the headline. It is a pause with an expiry, not a prohibition, and it is deliberately uneven: high schoolers keep limited access plus AI literacy classes twice a year, teachers may still use AI for lesson planning, and exceptions cover students with disabilities, multilingual learners and computer-science courses. A pilot in up to five classrooms will test vetted tools. Read together with Princeton and UNAM, the pattern is that institutions are not choosing between adoption and rejection — they are drawing lines by age, by subject, and by who is using the tool for what. That is a more durable framework than a blanket answer, and it is the one a parent or an educator is most likely to encounter in practice.

This matters for two reasons. First, sentiment among the youngest cohort tends to lead broader public attitudes by 5 to 10 years — Gen Z attitudes today are a leading indicator of where the median voter and median consumer will be by the early 2030s. Second, professionals entering the workforce who view AI as adversarial are unlikely to become enthusiastic adopters of it inside their employers. The companies betting on AI-driven productivity gains over the next five years are betting partly on a cohort that is increasingly resistant to the technology.

The US-China Geopolitical Dimension

AI is not just an economic technology — it is a geopolitical competition:

• The US and China are competing directly for AI leadership in semiconductors (NVIDIA vs. Huawei Ascend), model capabilities (GPT/Claude vs. DeepSeek/Qwen), and talent
• US export controls on advanced AI chips have been a key lever — though the policy direction has shifted. Biden-era restrictions attempted to limit China's access to cutting-edge chips; the Trump administration loosened some controls in late 2025, approving H200 chip exports to select Chinese customers and shifting licensing from "presumption of denial" to case-by-case review. China continues investing heavily in domestic semiconductor alternatives
• Control of AI infrastructure — compute, data, models — is being treated as a national security asset by both governments
• Smaller countries and regions (EU, India, Southeast Asia) are navigating between these poles, trying to build domestic AI capacity while avoiding dependence on either superpower

Information & Democracy: Synthetic Media at Scale

Deepfakes and Synthetic Media

Deepfake technology — AI-generated video, audio, and images indistinguishable from real content — has advanced dramatically:

• Voice cloning at high quality requires only a few seconds of audio and is available through commercial services
• Video deepfakes of public figures can be generated without specialized equipment
• AI-generated photographs of people who do not exist are indistinguishable from real photographs to human observers

Documented harms: Deepfake audio of politicians has been used in election contexts across multiple countries — including Romania (fabricated candidate investment scams), South Korea (AI-generated candidate smears and fake news anchors), and Canada (a deepfake of PM Carney viewed over a million times). Deepfake pornography of non-consenting individuals is a growing harm predominantly targeting women. Financial fraud using voice cloning of executives has resulted in hundreds of millions in losses.

Legislative response: The TAKE IT DOWN Act (signed May 2025) is the first US federal law targeting deepfakes. It criminalizes non-consensual intimate deepfake imagery and requires platforms to remove non-consensual intimate images and known identical copies within 48 hours of a valid request. The compliance deadline arrived on May 19, 2026 and FTC enforcement began immediately: on May 20, 2026 the Federal Trade Commission sent warning letters to twelve "nudify" websites accused of letting users generate non-consensual sexual images, alongside reminder letters to fifteen of the largest US platforms — Alphabet, Amazon, Apple, Automattic, Bumble, Discord, Match Group, Meta, Microsoft, Pinterest, Reddit, SmugMug, Snapchat, TikTok, and X — with civil penalties of up to $53,088 per violation on the table. As of early 2026, 48 of 50 US states have also enacted some form of deepfake legislation.

California went further on August 2, 2026, when its AI Transparency Act (Senate Bill 942, as amended by Assembly Bill 853) became operative. Where the TAKE IT DOWN Act polices a specific harm after the fact, the California law regulates the supply side in advance: any generative AI system with more than one million monthly users in the state must embed machine-readable provenance data in every image, video, and audio output, and must publish a free public tool that reads it back. Penalties run $5,000 per violation, with each day of non-compliance counted separately. The obligations then widen — large online platforms and model-hosting services must preserve and display that provenance from January 1, 2027, and camera makers must offer it at the point of capture from January 1, 2028.

How Much of the Web Is Machine-Written

Deepfakes are the vivid case, but the larger shift is quieter and made of ordinary text. In August 2026 Pew Research Center analyzed roughly 490,000 English-language pages drawn from the Common Crawl archive and found that more than a third of pages published since ChatGPT's late-2022 release carry significant signs of machine authorship.

That number needs one qualification, because it is the one most often dropped in the retelling. Measured against the web as a whole, the figure is about 10 percent — most of the internet was written before the tools existed. The one-third figure describes new pages only. Both numbers are correct; they answer different questions, and quoting the larger one as though it described the whole web overstates the change considerably.

The distribution is uneven in a way that maps onto incentives:

Commercial pages are roughly ten times more likely to show machine authorship than educational or government ones — consistent with AI writing being adopted fastest where volume is directly monetized. Pew also tracked stylistic markers, finding em dashes roughly doubled in frequency since 2023 and AI-favored vocabulary such as "delve" more than doubling.

AI-Generated Disinformation Campaigns

The cost of generating persuasive disinformation has collapsed:

• Before 2022: Running a large-scale disinformation campaign required significant human labor — writers, translators, social media managers
• After 2023: AI can generate thousands of unique, contextually appropriate disinformation variants, in multiple languages, adapted to local political contexts, at near-zero marginal cost

Examples from documented research:

• AI-generated voter suppression messages targeting specific demographics with personalized false information about voting procedures
• AI-generated astroturf campaigns creating the appearance of organic public opinion
• AI-synthesized "news articles" with realistic formatting distributed on social media

Recommendation Systems and the Information Environment

AI-powered recommendation systems — the algorithms that decide what content you see on YouTube, TikTok, Twitter, and Facebook — optimize primarily for engagement. Engagement and accurate information are not the same thing, and in many cases they are inversely related.

This creates structural pressure toward outrage, confirmation bias, and radicalization — not because anyone designed these systems to cause harm, but because the optimization target (time on platform) rewards emotionally engaging content disproportionately.

Detection technology: Watermarking (Google's SynthID), AI content detection tools, and cryptographic content provenance — the Coalition for Content Provenance and Authenticity, or C2PA, standard — allow authentic content to be verified. As of August 2026 this is no longer purely voluntary: California's AI Transparency Act makes C2PA-compatible provenance a legal requirement for large providers, which removes the adoption problem that had been the approach's main obstacle. A change on August 14, 2026 shows how the regime actually works. Google made the visible watermark on its AI images, video and music optional — a settings toggle across Gemini, Nano Banana, Omni, Lyria and the Flow editor — while leaving the invisible SynthID watermark and C2PA metadata attached and not user-removable. Read quickly that looks like a provider retreating from labeling. It is closer to the opposite: the visible badge was always a courtesy, and the machine-readable layer underneath is the part the law reaches. Expect more of this split, because a provider can satisfy a disclosure rule without putting a mark on the pixels a viewer sees.

The harder limit is structural and remains unsolved. Provenance proves what was labeled; it says nothing about unlabeled content from a provider under the user threshold, a model hosted outside the reach of the rule, or an open-weights model running on someone's own hardware. And a provenance signal only helps if something in the chain actually checks it — a machine-readable credential nobody reads is not a safeguard.

Privacy & Surveillance

The Facial Recognition Landscape

Facial recognition is deployed in at least 60 countries, including by governments with limited democratic oversight. Applications include:

• Law enforcement: identifying suspects in crowds, matching arrested individuals to databases
• Border control: automated passport verification, traveler tracking (US CBP implemented facial recognition for all foreign travelers in December 2025)
• Commercial: retail loss prevention, employee monitoring, customer identification
• Mass surveillance: China's Social Credit System as the most extensive example, integrating facial recognition, purchase data, social media, and behavioral tracking into a citizen scoring system

The accuracy and bias problem: Multiple studies, including audits of commercial facial recognition systems by NIST (National Institute of Standards and Technology), have documented higher error rates for women, darker-skinned individuals, and younger and older age groups. Misidentification in law enforcement contexts has resulted in wrongful arrests.

The consumer-grade tier: Facial recognition is no longer only a government or enterprise capability — reverse-image services sell it to anyone with a browser, and they inherit ordinary startup security practices. In August 2026 security researcher Jeremiah Fowler found that ClarityCheck, a service that lets a user upload a photo and try to identify the person in it, had left roughly 9 million image files (about 450 gigabytes) in an unsecured cloud storage bucket, reachable through a link sitting in the site's own public page code. The exposed pictures included adults, teenagers and children, many apparently taken from private profiles; a second misconfiguration exposed email addresses and phone numbers. The site displayed a notice describing its image search as private and secure throughout. The company has since secured the data.

The incident is worth holding onto for a reason beyond the breach itself. The subjects of those photographs were never customers of ClarityCheck, never agreed to anything, and had no way to know they were in the database — so there was no one with both the standing and the knowledge to object. That gap, rather than the misconfiguration, is the structural problem: privacy frameworks built around notice and consent do not reach a system whose inputs are scraped and whose subjects are strangers.

Faces Are Not the Main Event

Facial recognition dominates the public conversation about AI surveillance, but it is not the form most Americans actually encounter. The widest-deployed police AI in the United States reads license plates, not faces, and it is worth separating three different things that get discussed as one:

• Identification — working out who someone is from a biometric signal. This is facial recognition, and it is the one with the well-documented accuracy and bias problems above.
• Location tracking — recording where a vehicle was, and when, without identifying anyone directly. Automated license plate readers photograph passing cars and log plate, timestamp and location. No face is involved, and in most states no warrant is required.
• Fusion — joining those records to other databases (registrations, permits, case files, other jurisdictions' cameras) until the combined picture identifies a person and reconstructs their movements over months.

Fusion is where the civil-liberties question actually lives. A single plate photograph is close to meaningless; a searchable national history of every road a car has travelled is a different kind of object, and it is assembled from individually unremarkable records.

The most transferable lesson in this whole debate is about safeguards rather than cameras: a control that an operator can switch off is a default, not a limit. After sustained criticism in 2026, one major plate-reader vendor cut default data retention from thirty days to seven and required officers to enter a case number before searching. Both were real improvements, and both can be overridden by the operating agency, with an "Evidence Mode" that preserves records past the retention window. So when you evaluate any surveillance system, ask two questions: who can turn the safeguard off, and is that action logged? A protection the watching party controls unilaterally constrains accidents, not intent.

The accountability failure this produces is usually not a hack. It is authorized misuse — a lawfully procured system, used by credentialed staff, for purposes nobody approved. In 2026 dozens of US officers were accused of running plate searches to track partners and former partners. Nothing was breached and nothing leaked; the system worked exactly as designed, for the wrong reason. That is a governance and audit problem rather than a security one, and it is invisible to the technical safeguards that dominate procurement discussions.

Regulatory response: The EU AI Act banned real-time biometric surveillance in public spaces (enforceable from February 2025). The May 2026 Omnibus deal postponed broader high-risk AI obligations to December 2, 2027 for standalone systems and August 2, 2028 for AI embedded in regulated products, but the biometric surveillance ban itself remains in force. China enacted its first dedicated facial recognition regulation (effective June 2025). In the US, 23 states now restrict biometric data scraping, though there is no federal facial recognition law.

Behavioral Prediction and Surveillance Capitalism

Beyond facial recognition, AI systems construct detailed behavioral models of individuals:

• Purchases, search history, location data, app usage, and social connections are combined to predict political affiliation, health status, sexual orientation, financial vulnerability, and psychological traits
• These models are used for advertising targeting, credit decisions, insurance pricing, and — increasingly — hiring
• The economic model of internet platforms is built on the value of these behavioral predictions, creating structural incentives to maximize data collection

Creativity & Intellectual Property

The Copyright Questions

AI training on human-created content has generated significant legal disputes, and the pace is accelerating. The AI Copyright Case Tracker counted 131 US copyright lawsuits against AI companies as of August 10, 2026 — up from about 100 in April — with at least 30 more filed outside the US. It has been counting since January 2023. The legal landscape is evolving rapidly:

• Key question: Does training an AI model on copyrighted content constitute copyright infringement?
• Landmark rulings and settlements (2025):
  • Thomson Reuters v. ROSS (Feb 2025): The first federal ruling to reject an AI fair use defense. The court found that ROSS's use of Westlaw headnotes to train a competing AI search tool was not fair use. On appeal to the Third Circuit.
  • Bartz v. Anthropic ($1.5 billion settlement, final approval July 2026): The largest copyright settlement in US history. Anthropic admitted downloading over 7 million books from pirate sites to train Claude; the deal averages roughly $3,000 per work across an estimated 500,000 books and required Anthropic to destroy the pirated libraries. A federal judge granted the settlement final approval in July 2026, over objections from some authors. Because Anthropic chose to settle rather than appeal, the original judge's split ruling — that training on lawfully acquired books is fair use, but storing pirated copies is not — stays a single district-court decision and never becomes binding precedent for the rest of the industry. The settlement grants no future license. Watch what happened when the money actually moved, because it is the part of a headline settlement figure that nobody reports: payment is split evenly between author and publisher where a book is still in print with a traditional publisher, and paid in full to the author where the book was self-published or the rights have reverted — and in September 2026 writers began reporting that publishers had claimed books whose rights returned years ago, and had claimed the whole amount on books where they were owed half. Literary agencies filed claims too, though an agent holds no rights in a book they sold. The Authors Guild attributed the pattern to poor record-keeping and a confusing claims process rather than to bad faith, and some publishers said the claims were errors they had asked Anthropic to correct. The mechanism that decides most disputes is a single date: to claim the full amount an author must show the rights reverted before August 10, 2022, the download date named in the settlement. A class settlement is an administrative system, not a payment, and the distribution rules determine who is actually compensated.
  • Sony Music Publishing and Warner Chappell v. Anthropic (filed August 28, 2026): the gap above being tested. Two of the three largest music publishers sued in the Northern District of California, naming co-founders Dario Amodei and Benjamin Mann personally alongside the company. The complaint covers tens of thousands of musical compositions and alleges the lyrics reached Claude through torrented book libraries, scraping of the licensed lyric sites MusixMatch and LyricFind, and the Common Crawl and Books3 datasets — seeking up to $150,000 per work willfully infringed. Anthropic says it disagrees and will defend itself. It is the fourth publishing suit against the company, and it matters structurally: because Bartz settled rather than reaching appeal, the acquisition-versus-training distinction was never made binding, so the same question is being litigated again against the same defendant with a different rightsholder class
  • Getty Images v. Stability AI (UK, Nov 2025): The UK High Court largely rejected Getty's copyright claims, ruling that AI model weights are not "copies" under UK law. Getty won only limited trademark claims — a significant divergence from the US legal trajectory.
  • Warner Music v. Suno (Nov 2025): Settled. Suno launched a new model trained on licensed data.
  • NYT v. OpenAI (ongoing): Main claims proceeding. The court ordered preservation of all ChatGPT output logs. Centers on "regurgitation" of memorized copyrighted content. In September 2026 the US government intervened on OpenAI's side, filing a statement of interest arguing that training a large language model on copyrighted text is fair use. Government attorneys wrote that narrowing fair use to exclude such training would be inconsistent with basic copyright principles and would hamper "the Progress of Science and useful Arts," and framed the stakes in economic terms — that constraining model development "would thwart such creative and scientific progress while hindering American prosperity and economic mobility." The filing does not decide anything: a statement of interest is an argument the court may weigh, not a ruling, and the government explicitly conceded that "the fair-use inquiry hinges on the specific facts and uses at issue in each case." What it does signal is that the executive branch has taken a public position on the central legal question facing the industry, in the case most likely to set precedent for every other publisher's claim
  • The regurgitation evidence (September 2026): because the case turns on reproduction, the measured rate of it is the argument, and Microsoft — a co-defendant — put numbers on the record for the first time. In discovery it handed the publishers' expert 8.2 million Copilot chat logs, selected, in its own description, because they hit keywords making them the most likely of any conversations to contain the plaintiffs' work. Of that set, 59,545 shared at least sixteen words with news content used to ground the model. In the authors' parallel suit, 24 responses out of the 8.2 million contained thirty or more matching words, and only ten of 212 books evaluated produced any match at all. An expert for the Center for Investigative Reporting found 51 instances of substantial overlap. The New York Times rejects Microsoft's conclusions. Read carefully, these are a defendant's figures on a defendant's framing: the sample was deliberately enriched for matches, which cuts both ways — it makes a low hit rate more meaningful, and it means the percentages describe an unrepresentative slice rather than typical use. What the numbers cannot settle is the legal question of whether rare-but-real reproduction of a protected work is infringement, which does not obviously turn on frequency
• International variation: The EU AI Act requires foundation model developers to disclose training data; Japan has taken a more permissive approach to AI training on copyrighted content; the UK ruling suggests a narrower view of copyright in the AI context than US courts

Artists and the Displacement Question

• Music: AI can generate songs in the style of any artist without licensing or compensation. Universal Music Group and others have sued AI music companies.
• Visual art: AI image generators (Midjourney, Stable Diffusion) can generate work in the style of living artists, raising economic displacement concerns for commercial illustrators and concept artists.
• Film and TV: The 2023 WGA and SAG-AFTRA strikes explicitly addressed AI usage in Hollywood, resulting in negotiated protections around AI-generated scripts and likeness rights.
• Writers: The flood of AI-generated content has depressed rates for commercial writing, SEO content, and certain categories of journalism

The Cultural Homogenization Risk

A subtler concern: AI trained on the most common patterns in human-generated content may produce output that averages and blends existing styles. If AI-generated content displaces human-created content at scale, we may end up with a cultural environment that is technically competent but increasingly samey — optimized for average preferences rather than the edges where culture actually evolves.

Key Takeaways

• AI's economic benefits are real ($13–17 trillion GDP addition estimated by 2030, net +78 million jobs per WEF) but are distributed unevenly — productivity gains tend to concentrate among capital owners and highly skilled workers
• Synthetic media and AI-generated disinformation represent genuine threats to democratic information environments at a scale and cost-effectiveness not previously possible
• AI surveillance systems — facial recognition, behavioral prediction — are deployed globally with limited oversight, raising serious civil liberties questions
• Copyright law, artist compensation, and the nature of creative originality are all in active legal and social contestation as AI-generated content proliferates
• Being an informed citizen in the AI era requires engaging with these questions, not just the productivity opportunities