Information, manipulation, livelihoods, and the planet

Lesson 5 of 6 in Why AI Needs Governance: A Taxonomy of Harms and Risks.

The harms in this lesson mostly target no one in particular — they target the commons: what societies can trust, how freely people choose, how work is valued, and what the physical infrastructure costs the planet.

Information integrity. Two days before Slovakia’s 2023 parliamentary election — during the legal pre-vote media blackout, when corrections were hardest — a faked audio recording of candidate Michal Šimečka “discussing” vote-buying spread across social media. Whether it changed the result is unknowable; that is precisely the problem. Synthetic media at scale produces three distinct harms: direct deception (deepfakes, fabricated events), non-consensual intimate imagery — now the most common malicious deepfake use, overwhelmingly targeting women, and the subject of dedicated criminal laws including the US TAKE IT DOWN Act — and the subtlest one, the liar’s dividend: once anything can be faked, anyone caught on tape can shout “deepfake!”, and real evidence loses its power. Add the flood of low-grade generated content (“AI slop”) displacing human writing in search results, and the harm is to the epistemic environment itself — which is why disclosure-and-labeling rules (EU AI Act Article 50, China’s labeling measures, election-deepfake statutes across US states) became the fastest-moving corner of AI law.

Manipulation and autonomy. AI can optimise persuasion the way it optimises anything else: dark patterns tuned per user, hyper-personalised political messaging, recommender systems that discover outrage retains attention, and AI companions engineered — or merely optimised — for emotional dependence. The EU AI Act’s prohibited-practices list reads as a canonical inventory of manipulation harms: subliminal or purposefully manipulative techniques causing significant harm, exploitation of vulnerabilities due to age or disability, social scoring, and emotion recognition in workplaces and schools, all banned outright since February 2025.

Economic and labour harms. Three layers, often conflated:

• Displacement and transformation — tasks automated, occupations reshaped. Honest forecasting is hard, but governance questions arrive regardless: who is consulted, retrained, compensated?
• Algorithmic management — gig workers assigned, scored, and deactivated by systems they cannot see or appeal; warehouse pace-setting; screen-monitoring of remote workers. Here AI is the boss, and labour law is racing to catch up (the EU’s Platform Work Directive is the flagship response).
• The hidden workforce and the creative economy — the annotation workers who label training data (including traumatic content, sometimes for under $2/hour, as TIME’s 2023 reporting documented), and creators whose life’s work trained the generators now competing with them — the core of the copyright litigation wave from NYT v. OpenAI to the artists’ suits against image-model makers.

Environmental harms. Training a single frontier model consumes gigawatt-hours; data-centre water cooling collides with local droughts; hardware turnover creates e-waste and supply-chain impacts. The honest current state is a measurement-transparency problem: companies disclose little, estimates vary wildly, and you cannot govern what nobody measures — which is why disclosure requirements (energy reporting for GPAI training under the EU AI Act’s documentation duties) are the current regulatory front line.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.