• Last time, we looked at AI and the labour market
• AI is different from earlier waves: it targets cognitive tasks, not just physical ones
• Automation displaces tasks, not whole jobs (the Acemoglu & Restrepo framework)
• AI raises worker productivity but may compress wages for some and widen inequality
• The graduate job market is rough right now, and AI is part of the reason
• Today: a different angle. What does AI do to people, not jobs?
• Three themes: the attention economy, mental health, and the environment

Part 1: The attention economy

• “Your attention is the product”
• How recommendation algorithms work
• Filter bubbles: what the evidence says
• What companies are actually doing

Part 2: AI and mental health

• The treatment gap
• Chatbots in therapy: what the evidence says
• Risks and hard limits
• Augmentation, not replacement

Part 3: AI and the environment

• Energy and water costs
• Footprint in context
• AI as a climate tool

We will pause for discussion after each part

There are no right answers to the questions I put on screen. I’m curious what you think!

• Herbert Simon (1971): information-rich environments create attention-poor ones
• You have limited hours and limited focus. That’s the real scarce resource (Becker, 1965 - “A theory of the allocation of time”)
• The business model is to capture attention and sell it to advertisers, just like TV, radio, and newspapers before
• Tristan Harris (ex-Google design ethicist): called it “the race to the bottom of the brainstem”
• Algorithms are designed to maximise engagement, not truthfulness or wellbeing
• Tim Wu’s The Attention Merchants (2016) traces this logic from 19th-century newspapers to social media

• Eli Pariser coined “filter bubble” in 2011
• His argument (which many people believe) is that algorithms seal you in a cocoon of confirming content
• However, two large experiments say that the effect is probably overstated
• Nyhan et al. (2023, Nature): reduced like-minded content on Facebook during the 2020 US election, polarisation barely changed
• Guess et al. (2023, Science): algorithmic vs. chronological feeds → little difference in political attitudes
• So filter bubbles may not drive polarisation the way we feared (phew! 😉)
• But that’s one outcome. Effects on anxiety, body image, attention spans, and sleep are separate questions, with worse answers. Let’s talk about those now!

• YouTube has gradually shifted away from watch time as a sole metric, adding satisfaction surveys and other signals
• Took over a decade to move beyond pure engagement metrics
• The “outrage drives clicks” problem (Brady et al., 2017, PNAS): content that makes you angry, anxious, or fearful gets shared more and watched longer
• The algorithm learns this. Surfaces more of it
• The Facebook Papers (2021, WSJ): internal research showed Instagram made body image worse for teenage girls
• Facebook had the data, but kept the feature because fixing it would have cost engagement
• The problem is hard to solve, as company revenue and user wellbeing point in opposite directions

• Jonathan Haidt’s The Anxious Generation (2024): smartphones and social media are rewiring childhood and driving a mental health crisis
• However, Orben & Przybylski (2019, Nature Human Behaviour) say that screen time explains at most 0.4% of the variance in adolescent wellbeing
• This doesn’t mean individual experiences aren’t real. It means the population-level effect is small
• Haidt’s evidence is mostly correlational: teen depression rose around the same time as smartphone adoption
• Candice Odgers (2024, Nature): reviewed Haidt’s claims and found the evidence does not support his causal story
• Many studies (Godard and Holtzman, 2024, Nagata et al, 2025, Shannon et al, 2021) also find that the association between screen time and wellbeing is small and inconsistent

• Most are opt-in or reach a minority of users. The ad model is unchanged
• Also handy to point at when regulators come knocking
• US Surgeon General (2023): issued an advisory on social media and youth mental health; called for warning labels in 2024
• Amy Orben (Cambridge): transparency about why content is recommended changes behaviour more than usage caps
• Knowing you’re being nudged turns out to matter

• Since March 2022, Chinese users can legally turn off recommendation algorithms on all platforms
• The Internet Information Service Algorithmic Recommendation Management Provisions requires every app to offer a non-personalised feed
• Douyin (Chinese TikTok), Baidu, Taobao, and WeChat all added one-tap opt-out buttons
• Early research suggests users appreciated the transparency, though long-term effects on usage and wellbeing are still being studied
• The EU Digital Services Act (Article 38) now requires large platforms to offer a non-profiling feed option in the EU too
• China requires algorithmic transparency for users while maintaining state censorship of content itself. Both things are true at the same time

• WHO estimate: 75% of people with mental health conditions in low- and middle-income countries receive no treatment at all
• Even in wealthy countries:
  • Average US wait for a first psychiatric appointment: 25 days nationally; median 67 days in recent data, with severe shortages in rural areas
  • A therapy session costs $150-250 out of pocket
  • ~0.1 psychiatrists per 100,000 people in low-income countries
  • Stigma stops many from seeking help where care does exist
• This is the context for AI mental health tools
• If the alternative is literally nothing, the calculation looks different than if you’re imagining replacing well-funded human care

• Woebot (2017): first major CBT-based chatbot, 2M+ users. Follows clinical CBT protocols, not a free-form LLM
• Wysa: similar model, piloted by the UK NHS in 31 trusts
• Replika: not therapy, but an “AI companion.” People form real emotional attachments (more on the risks of this shortly)
• Best recent evidence: Habicht et al. (2025, JMIR): Generative AI used alongside group therapy for real patients
• Improvements in clinical outcomes and patient engagement
• But AI was supplementing human therapists, not replacing them

• Opel and Breakspear (Science, 2026), two clinical neuroscientists, say AI “may reduce care inequities when deployed responsibly”
• “May” and “responsibly” are doing a lot of work here 😅
• What the evidence contains:
  • Small RCTs: short-term symptom relief for mild-to-moderate anxiety and depression
  • Very few studies run past 8–12 weeks. Long-term effects? Unknown
  • High dropout: many people stop using these apps quickly
  • Publication bias: positive results get published; negative ones mostly don’t
• Real signals that AI can help some people with some conditions, short-term. But the evidence base is much thinner than the marketing suggests

These are documented, not hypothetical:

• Dependency. When Replika changed features in 2023, users reported grief comparable to losing a real relationship. What happens when an app you’re attached to gets discontinued?
• Data privacy. Therapy touches the most private parts of someone’s life. Chatbot data is rarely protected like clinical notes. Where does it go?
• Harmful responses. Moore et al. (2025): LLMs expressed stigmatising attitudes toward mental health conditions and gave harmful crisis responses
• No crisis escalation. A chatbot cannot call an ambulance, contact a GP, or do anything in the physical world
• Regulatory gap. Most mental health chatbots are classified as apps, not medical devices. No clinical regulation in the US, UK, or most of the EU

But also consider:

• Therapy with a human isn’t fully private either: notes, supervision, insurance coding
• “Harmful responses from AI” vs. zero access to care for 75% of the world
• AI tools may give governments and employers an excuse to avoid funding real care

LLMs cannot:

• Hold reliable long-term memory between sessions. Some chatbots (ChatGPT, Claude) now offer memory features, but these are limited and not designed for clinical continuity
• Verify what you tell them. A therapist builds context over months
• Provide legally enforceable confidentiality
• Diagnose or prescribe medication
• Read body language, tone, or facial expression

LLMs can:

• Be available at 3am
• Be patient. Not judge. Not get tired
• Deliver structured info on anxiety, depression, coping strategies
• Scale to millions at near-zero marginal cost

These are different capabilities, perhaps complementary, but not perfect substitutes

The most credible use cases right now:

• First contact: AI handles initial triage, reduces stigma, connects people to professionals
• Between-session support: coping strategies and mood tracking between weekly appointments (the Habicht et al. model)
• Stepped care: AI for mild symptoms, human professionals for moderate-to-severe
• Admin burden: AI writes session notes, handles scheduling, finds local services. Frees therapist time for actual care
• OpenAI updated ChatGPT in 2025: added safety messaging, crisis hotline signposting, trauma-informed language. An acknowledgement that people already use the product this way
• Who decides what “responsible” looks like?

• Training a large model costs serious energy. Companies rarely disclose the numbers
• Patterson et al. (2021): training GPT-3 produced ~552 tonnes of CO₂. That is equivalent to:
  • 368 return flights from London to New York
  • 120 cars driven for a full year
  • The annual carbon footprint of 61 Americans
• GPT-4? Nobody knows. Estimates range from 5x to 50x GPT-3
• But training is a one-off cost
• Running billions of queries (inference) is ongoing, and at scale it probably dwarfs training
• Microsoft’s carbon emissions rose 29% between 2020 and 2023, partly driven by AI
• Google’s 2024 Environmental Report: emissions up 48% since 2019, largely from data centres

• Energy gets the headlines. Water gets ignored
• Li et al. (2023), “Making AI Less Thirsty”: training GPT-3 used ~700,000 litres of freshwater for cooling
• A typical ChatGPT conversation: about 500ml, roughly a bottle of water
• Data centres get built where land and electricity are cheap, often in water-stressed regions: Phoenix, Las Vegas, northern Chile
• GPUs last 3-5 years. Manufacturing needs lithium, cobalt, and rare earth metals, each with its own environmental cost
• None of this shows up in the headline CO₂ figures

Current estimates (IEA, 2024):

• Right now, AI’s direct climate footprint is real but not huge compared to aviation or steel
• The worry is growth rate: AI electricity use grew ~60% year-on-year between 2022–2024
• IEA projections: data centre electricity demand could double to ~945 TWh by 2030, roughly equivalent to Japan’s total consumption
• The Jevons paradox: more efficient models → more usage → efficiency gains get cancelled out
• GPT-4 is far more efficient per query than GPT-3. But there are vastly more queries

The same technology that burns energy may also help cut emissions:

• GraphCast (DeepMind, 2023): 10-day weather forecasts, faster and more accurate than traditional models. Helps manage renewable grids
• Grid optimisation: AI predicts electricity demand, balances wind and solar in real time (DeepMind, 2019)
• Wildfire prediction: better detection and spread modelling (Jain et al., 2020, Environmental Reviews)
• Data centre cooling: DeepMind’s RL system cut cooling energy by 40%
• Rolnick et al. (2022, ACM Computing Surveys): identified 70+ high-potential ML applications for climate change, from agriculture to carbon capture
  • But: these are potential benefits, not guaranteed ones