Generative AI and Everyday Life: Wellbeing, Work, and Skills Before and After ChatGPT (2015-2026)

1. Introduction and Motivation

On 30 November 2022, OpenAI released ChatGPT as a free research preview. Within five days it had passed one million users, and within two months an estimated 100 million people had tried it, a rate of adoption that UBS analysts described as the fastest for any consumer application in history. The product was not a new technology in the strict sense. It rested on the transformer architecture introduced in 2017 and on language models that OpenAI, Google, and others had been scaling for years. What changed in late 2022 was access. A chatbot that anyone could open in a browser turned generative AI from a specialist tool into a mass-market utility, and the pace of diffusion since then has been extraordinary: 100 million weekly users by November 2023, 400 million by February 2025, and about 800 million by October 2025.

This report asks what that diffusion has meant for ordinary people. The question is harder than it looks. Everyday life is not one variable but many, and the effects of a general-purpose technology are expected to show up differently across evaluative judgements (how people rate their lives), emotional experience (how people feel day to day), perceived control (whether people feel freer or more managed), health (stress and mental strain), and economic position (job security and earnings). The report therefore separates wellbeing into the two constructs that survey research has learned to keep apart: life evaluation and affective experience (Kahneman and Deaton 2010; Diener et al. 2010). It then examines perceived freedom, stress, job security, the distribution of gains and losses, and the emerging claim that AI lets non-experts do work once reserved for specialists.

The analytical window is 2015 to 2026, split at the ChatGPT launch. The pre-AI era (2015 to 2022) is long enough to establish baselines that include the pandemic; the post-AI era (2023 to 2026) is short but already rich in survey and experimental data. A pre/post comparison of this kind is suggestive, not causal. Three confounders sit squarely in the window: the post-pandemic recovery in social contact and mental health, the 2022 to 2023 inflation and cost-of-living shock, and pre-existing trends in happiness, stress, and automation. Where a change could equally reflect these forces, the report says so rather than crediting AI.

The evidence base combines large international surveys (Gallup World Poll, World Happiness Report, Pew Research Center, Microsoft Work Trend Index), labour market analyses (ILO, IMF, OECD, World Economic Forum), and a fast-growing experimental literature on generative AI and productivity. The first five dimensions are comparatively well measured; the last two (distribution and non-expert empowerment) are more conceptual, and the report operationalises each explicitly and flags where the data are thin.

2. Data and Method

Table 1 lists the indicator used for each dimension, the primary source, and the main limitation. The guiding rule is that every dimension is measured by a named indicator, reported separately from the others, and discussed with its measurement problems visible.

Two aggregation choices need stating. First, the global life evaluation mean is unweighted across countries, so small countries count as much as large ones; this is the standard convention in World Happiness Report summaries and is noted wherever a global number appears. Second, Gallup publishes some global figures as averages and others as medians across countries. The freedom series in Section 5 keeps those two formats labelled rather than merging them.

The report treats November 2022 as the inflection point in every chart, but it does not assume that the boundary is clean. Section 10 returns to the confounders and asks what a causal research design would need. Throughout, "exposed" means that a share of an occupation's tasks overlaps with what generative AI can do, not that jobs are disappearing, and the report is careful to distinguish exposure from displacement, and perceived risk from measured risk.

3. Quality of Life and Life Satisfaction

Life satisfaction is the most direct evaluative measure of whether lives are going well, and it has been collected consistently by the Gallup World Poll through the Cantril ladder since 2005. The World Happiness Report publishes country-level values as three-year rolling averages, and the unweighted world mean of those values is the closest thing to a global happiness trend. It rose from about 5.38 in 2015 to 5.65 in 2025, an increase of roughly a quarter of a point on a 0 to 10 scale (Figure 2).

The timing matters. The world mean moved from 5.38 in 2015 to 5.55 in 2021, before ChatGPT existed, and the 2021 to 2022 period saw a slight decline (5.55 to 5.54). The post-launch path is a recovery that had already begun: 5.53 in 2023, 5.58 in 2024, 5.65 in 2025. A simple pre/post comparison gives a rise from an average of 5.45 across the 2015 to 2022 rows to 5.59 across the 2023 to 2025 rows, but attributing that rise to AI would ignore the pandemic recovery, which plausibly explains a large share of it. The honest reading is that global life evaluation is slightly higher in the post-AI era than in the pre-AI era, that the improvement started before ChatGPT, and that no credible design yet separates the AI contribution from the recovery contribution.

The country-level pattern is more instructive than the global mean, because it shows how little the ChatGPT boundary explains. In the United States, life evaluation fell from 7.10 in 2015 to 6.72 in 2024 before a partial recovery to 6.82 in 2025, with most of the decline concentrated between 2019 and 2023, a period dominated by the pandemic, political polarisation, and inflation. The United Kingdom shows a similar hump: 6.73 in 2015, 7.17 in 2019, and back to 6.69 by 2025. Australia drifted down from 7.31 to 6.92. Japan moved the other way, from 5.92 to 6.13, continuing a slow improvement that predates generative AI. India's ladder rose from 4.40 to 4.54 over the full window despite a pandemic dip. No pattern in these series lines up with the November 2022 boundary, which is itself the key finding: at the level of national aggregates, generative AI has not yet moved the life satisfaction needle in any detectable direction.

4. Happiness and Affective Wellbeing

Life evaluation answers a reflective question about life as a whole. Affective wellbeing answers a different question about yesterday: did you feel enjoyment, laughter, worry, stress, sadness, or anger during a lot of the day? Gallup aggregates the five positive experiences into a Positive Experience Index and the five negative experiences into a Negative Experience Index. Both are reported here as separate constructs, because the two measures can move in different directions and because conflating them is a common error in public discussion of AI and happiness.

The two indices trace a pandemic-shaped arc. The Positive Experience Index held at 71 through 2018 to 2020, dropped to 69 in 2021, recovered to 70 in 2022 and 71 in 2023, and reached 72 in 2024, its best reading of the past decade. The Negative Experience Index rose from 28 in 2016 to a record 33 in 2021 and 2022, then fell to 31 in 2023 and 2024. In other words, the emotional dip and the recovery both happened before and during the pandemic, and by 2024 global affect had returned to roughly its 2019 level. Generative AI entered daily life in this window, but the direction of the affect data is opposite to what a simple "AI makes us miserable" narrative would predict, and the timing fits the pandemic story far better.

The caveat is measurement. Global indices average across more than 140 countries, which hides large regional and demographic variation, and they measure the prevalence of emotions, not their intensity. A technology that improves life for some groups while worsening it for others could leave a global index flat. The distributional analysis in Section 8 therefore matters as much as the averages. What the affect data can support is a bounded claim: there is no visible global deterioration in daily emotional experience in the post-launch years, and the widely repeated claim that AI has made the world unhappier is not supported by the survey record to date.

5. Perceived Freedom and Autonomy

Perceived freedom is the felt sense of control over one's own life, and it is harder to measure than life satisfaction. The best available global indicator is the Gallup World Poll question on satisfaction with the freedom to choose what to do with one's life. Gallup reports that satisfaction with this freedom is higher now than two decades ago: a worldwide average of 80 percent in 2017, a median of 81 percent across 142 countries in 2024, and a median of 82 percent across 138 countries in 2025. The United States is a striking exception. Satisfaction among US adults fell from a 2007 to 2021 average of 83 percent to 71 percent in 2023 and 72 percent in 2024, and the decline is concentrated among women, whose satisfaction dropped from 81 percent in 2021 to 66 percent in 2024, a period that coincides with the loss of federal abortion protections and with deepening political polarisation (Figure 4).

Two features of this chart deserve emphasis. First, the global trend runs against the claim that AI is eroding felt freedom: on the one measure that has been tracked for two decades, the world reports more satisfaction with personal freedom, not less. Second, the American decline is a reminder that perceived freedom responds to institutions and politics more than to technology. The US drop began in 2022, at the same time as the ChatGPT launch, but it coincides precisely with the Dobbs decision, and Gallup's own analysis links the timing to that event. Generative AI is not the obvious culprit in this series.

The workplace is where the freedom question gets sharper, and where the data are thinner. The optimistic mechanism is that AI removes drudgery and gives workers more time for meaningful tasks. The pessimistic mechanism is algorithmic management: monitoring, pacing, and decision support that reduces worker discretion. Gallup's State of the Global Workplace data show that only about one in five employees worldwide were engaged at work in 2024, with 62 percent not engaged, but that series has been flat or declining since long before ChatGPT, so it cannot be attributed to AI. Research on algorithmic management and worker autonomy is emerging but has not yet produced the large, comparable time series that the freedom dimension needs. This is flagged in Section 11 as one of the clearest measurement gaps in the field.

6. Stress and Mental Health

Stress is the best-measured emotional indicator in the Gallup World Poll, and it has its own trajectory that does not map onto the AI timeline. Among all adults worldwide, the share reporting a lot of stress on the previous day rose from 37 percent in 2017 to a record 40 percent in 2020 and 41 percent in 2021, then fell back to 37 percent in 2023 and 2024 (Figure 5). The pandemic year 2020 also produced a 25 percent increase in the global prevalence of anxiety and depression, according to the World Health Organization, and those elevated levels have not fully normalised. Whatever generative AI has done to mental health, it entered a world already recovering from a historic stress shock, and the aggregate stress series shows improvement, not deterioration, in the post-launch years.

The workplace series tells a less comforting story. Employee stress hit 43 percent in 2020, 44 percent in 2021 and 2022, and remained at 41 percent in 2023 and 2024, still above its 2019 level of 38 percent. Microsoft's 2024 Work Trend Index, which surveys knowledge workers in 31 countries, found that 75 percent used generative AI at work and 46 percent reported burnout, with 68 percent saying they struggled with the pace and volume of work. These numbers sit side by side without a clean causal story: AI use and burnout rose together, but so did workload, restructuring, and the lingering effects of the pandemic. The most defensible statement is that generative AI is entering workplaces that were already stressed, that the always-on tools add a new cognitive load for some workers, and that no existing study cleanly separates the AI contribution from the workload contribution.

Two further notes. First, the WHO's 25 percent increase refers to the first pandemic year (2020), not to AI, and citing it as an AI effect would be an error. Second, the most plausible AI-specific mental health channel is workplace: fear of displacement, skill obsolescence, and monitoring. The survey evidence on worry is consistent with that channel being real but limited, which is the subject of the next section.

7. Job Security

Job security is the dimension where the economics literature is deepest, and where perceived and measured risk diverge most sharply. The perceived risk is real and rising. In September 2023, 22 percent of US workers told Gallup they worried technology would make their job obsolete, up from 15 percent in 2021. Microsoft's 2023 Work Trend Index found 49 percent of global knowledge workers worried AI would replace their job, and 70 percent said they would delegate as much work as possible to AI. By February 2025, Pew found 52 percent of US workers worried about AI use in the workplace, with 32 percent saying AI would lead to fewer job opportunities in their field (Figure 6). These are substantial numbers, and they moved up almost exactly with the ChatGPT timeline.

The measured risk is far smaller, at least so far. The ILO's 2023 analysis of generative AI and jobs found that only 5.5 percent of total employment in high-income countries sits in occupations potentially exposed to the automating effects of the technology, and 0.4 percent in low-income countries. The larger effect is augmentation: 13.4 percent of employment in high-income countries and 10.4 percent in low-income countries could be enhanced rather than replaced, which is why the ILO concluded that the technology is more likely to change the quality of jobs than to destroy them. The IMF's 2024 analysis uses a wider definition of exposure and finds that almost 40 percent of global employment is exposed to AI, about 60 percent in advanced economies, 40 percent in emerging markets, and 26 percent in low-income countries. The OECD's 2024 regional analysis estimates that about 26 percent of OECD workers are currently exposed to generative AI, with a potential for up to 70 percent as the technology spreads, and 39 percent of the potentially exposed workers considered highly exposed (Figure 7).

Projections of future churn are larger but still net positive in the latest round. The World Economic Forum's 2023 Future of Jobs Report expected 23 percent of jobs to change between 2023 and 2027, with 69 million jobs created and 83 million eliminated, a net reduction of 14 million. Its 2025 report, covering 2025 to 2030, projects 170 million jobs created and 92 million displaced, a net gain of 78 million, with 39 percent of workers' core skills expected to change (Figure 8). The swing from net negative to net positive between the two reports is itself a reminder that these are employer expectations, not outcomes, and that they respond to the economic cycle as much as to technology.

Three further observations tie the perception gap to the economics. First, exposure is not destiny: the occupations with the highest exposure, such as clerical work where the ILO estimates nearly a quarter of tasks are highly exposed, are the same occupations where augmentation may raise productivity and where wage effects will depend on how firms share the gains. Second, the historical record of automation scares is not reassuring about the fear itself: Frey and Osborne's 2013 estimate that 47 percent of US employment was at high risk of automation did not translate into mass displacement in the following decade, and task-level studies have consistently produced lower displacement estimates than occupation-level ones (Autor 2015). Third, wage data through 2025 do not yet show a generative-AI-driven displacement wave in national statistics, though the window is short and the effects are expected to show up in task content before they show up in employment counts.

8. Distribution of Effects

Who gains and who loses depends on which side of three gaps a person sits: the usage gap, the exposure gap, and the skill gap. The usage gap is the largest and the best measured. Pew's 2025 survey found 58 percent of US adults aged 18 to 29 had used ChatGPT, versus 41 percent of those aged 30 to 49, 25 percent of those aged 50 to 64, and 10 percent of those 65 and over (Figure 9a). An earlier Pew survey found the same pattern: 62 percent of Americans thought AI would have a major impact on workers generally, but only 28 percent thought it would have a major impact on them personally, a gap between generalised anxiety and personal expectation that is itself a distributional fact: those who expect to be affected are not evenly spread across the population.

The exposure gap runs the other way from the usage gap. The ILO finds that women's employment is more than twice as exposed to generative AI automation as men's: 3.7 percent of female employment worldwide versus 1.4 percent of male employment, driven by the concentration of women in clerical work, and the gap is larger in high-income countries. An ILO analysis published in 2026 finds the same pattern at the occupation level: about 29 percent of female-dominated occupations are exposed to generative AI, versus 16 percent of male-dominated occupations. The IMF finds exposure rising with development: about 60 percent of jobs in advanced economies are exposed versus 26 percent in low-income countries, because richer economies have more cognitive, white-collar work. The OECD finds a spatial version of the same pattern: 32 percent of workers in urban regions are currently exposed versus 21 percent in rural regions, and the OECD warns that generative AI could widen regional divides rather than narrow them. Education lines up with all of these: highly exposed occupations are disproportionately held by degree holders, and Gallup's 2023 finding that job obsolescence worry rose fastest among college-educated US workers is consistent with that.

Pulling the three gaps together produces a clear but uncomfortable picture. The people best placed to benefit, young and highly educated workers in rich countries, are also the most exposed, because exposure mostly means augmentation, and augmentation is where the productivity gains are. The people most at risk of displacement are those in clerical and administrative roles, disproportionately women, where tasks are more substitutable. And the people least exposed, such as workers in low-income countries and in physical, in-person occupations, are also the least likely to capture productivity gains, because they lack the complementary digital infrastructure and skills. Generative AI thus has the profile of a technology that could either compress or widen global inequality, depending on how its gains are distributed through training, wages, and social protection. The evidence so far is not sufficient to say which outcome dominates.

9. Non-Expert Empowerment versus Deskilling

The most consequential claim about generative AI is that it lets non-experts perform work once reserved for specialists: a person who cannot code can generate working scripts, a person who cannot draft legal language can produce a first-pass contract clause, and a student can receive instant explanations. The claim deserves to be tested, not asserted, because the evidence cuts both ways. The experimental literature on productivity is large enough to support real conclusions about where empowerment happens, and the learning literature is clear enough to show where it fails.

The consistent finding across tasks is that generative AI raises measured output, often dramatically. Noy and Zhang's randomised experiment with 453 college-educated professionals found ChatGPT reduced the time to complete writing tasks by 40 percent while improving judged quality by 18 percent. Peng et al. found developers with GitHub Copilot completed a coding task 55.8 percent faster. In a Fortune 500 customer support setting, Brynjolfsson, Li, and Raymond found the AI assistant raised issues resolved per hour by 14 percent on average, with a 34 percent gain for novice and low-skilled workers and near zero for experienced agents. Cui et al., running three field experiments across Microsoft, Accenture, and a Fortune 100 firm, found a 26 percent increase in completed software tasks. Dell'Acqua et al.'s experiment with management consultants found 12.2 percent more tasks completed, 25.1 percent faster work, and more than 40 percent higher output quality, with below-average consultants improving 43 percent versus 17 percent for above-average performers. This is the strongest version of the empowerment story: on well-defined tasks, the least skilled gain the most, which is what a general-purpose technology that compresses the cost of expertise should look like.

The counter-evidence comes from two places. The first is task complexity. Otis et al. ran a five-month field experiment in which 640 Kenyan small business owners received a GPT-4 business mentor. High-performing entrepreneurs increased profits by about 15 percent, but low performers did nearly 10 percent worse, because they lacked the complementary skills to act on the AI's advice. Empowerment is conditional on the ability to evaluate and execute, and where that ability is absent, AI can widen the gap rather than close it. The second is learning. Bastani et al. studied Turkish high school students in mathematics: those with unstructured ChatGPT access solved 48 percent more practice problems, but scored 17 percent worse on an exam taken without the tool, while students using a structured tutor prompt improved 127 percent on practice problems (Figure 11). The crutch effect is real: doing the task with AI does not equal learning to do the task, and unsupervised use can substitute for skill formation.

The honest synthesis is that generative AI is a genuine democratising technology for tasks with clear evaluation criteria, that its benefits on such tasks are largest for novices, and that it is a deskilling technology when it substitutes for the practice that builds expertise. These are not contradictory. The same tool that lets a junior developer ship code can prevent a student from learning to code. The empirical frontier is now about context: for which tasks, under which feedback and guardrail conditions, does AI-assisted performance translate into capability? The research opportunities in Section 11 are built around that question. What the current evidence cannot support is either of the extreme claims, that AI has levelled expertise everywhere or that it is deskilling an entire generation.

10. Discussion: What the Evidence Supports

Reviewed together, the seven dimensions support four solid conclusions and three honest non-conclusions.

First, generative AI has diffused faster than any comparable consumer technology, and its adoption is real and measurable. ChatGPT went from launch to 100 million weekly users in a year and to about 800 million by October 2025. McKinsey surveys show the share of organisations using AI in at least one function rising from 55 percent in 2023 to 72 percent in 2024 and 88 percent in 2025, with regular generative AI use up from 33 percent in 2023 to 65 percent in 2024 and 79 percent in 2025 (Figure 13). Pew's 44 percent ChatGPT usage among US adults in 2026 makes generative AI a mainstream consumer utility. This is not hype; it is the factual base on which the wellbeing questions sit.

Second, the productivity evidence for AI at work is unusually strong for a technology this young. Six independent randomised experiments, spanning writing, coding, customer support, consulting, and entrepreneurship, find gains of 12 to 56 percent, with the largest gains among less experienced workers in most settings. The OECD, Stanford, and the ILO all reach the same conclusion from different angles: the near-term effect of generative AI is augmentation, not mass substitution. This does not mean displacement will not come, but it means the reasonable baseline expectation is task reallocation and skill change, not a wave of unemployment.

Third, the wellbeing aggregates do not show an AI effect, and the burden of proof is on those who claim one. Global life evaluation, positive and negative affect, and population-wide stress all improved after their pandemic lows, and their movements align more with the pandemic recovery than with the ChatGPT boundary. The freedom data show the world feeling more, not less, free, with the notable exception of the United States, where the decline coincides with a political and legal event, not a technology. A responsible economist reading these series should conclude that generative AI has not yet produced a detectable aggregate effect on measured wellbeing in either direction.

Fourth, the risks are real and concentrated where the aggregates cannot see them. Workplace stress is high and flat, job worry rose from 15 to 22 percent between 2021 and 2023, clerical workers and women face the highest automation exposure, and the learning experiments show that unsupervised AI use can reduce skill formation. The distributional profile is the opposite of reassuring: usage is highest among the young and educated, exposure is highest among the same groups through augmentation, and displacement risk is highest among clerical workers who are disproportionately women.

The three non-conclusions follow directly. It cannot be claimed that AI has improved global wellbeing, because the confounders (pandemic recovery, cost-of-living shock, pre-existing trends) are too entangled and no credible causal design exists yet. It cannot be claimed that AI has harmed wellbeing, for the same reason. And it cannot be claimed that the non-expert empowerment question is settled, because the same body of experiments shows gains for novices on well-defined tasks and losses for low performers on complex ones. The defensible position is narrower and more useful: generative AI changes the value of skills, and the direction of that change depends on task structure, feedback, and the surrounding institutions.

11. Research Opportunities

The window 2023 to 2026 is short, which makes the existing evidence remarkable but also means the field is wide open. The highest-value opportunities combine the natural experiments created by AI rollout with the wellbeing and labour outcomes that the surveys track (Figure 14).

Eight topics stand out. (1) Causal effects of AI on wellbeing: staggered firm-level and school-level AI rollouts matched to employee and student outcomes would identify effects that the aggregate series cannot. (2) Algorithmic management and perceived freedom: linked employer-employee data with task-level monitoring could finally measure whether AI expands or erodes autonomy. (3) AI and job quality: repeated cross-sections of task content, work intensity, and working conditions would show whether augmentation changes the quality of jobs before it changes their quantity. (4) Deskilling versus empowerment: longitudinal measurement of skills before and after AI adoption, building on the Bastani et al. design, is the single most policy-relevant gap. (5) Distributional effects: ILO and IMF exposure measures matched to labour force surveys would allow the usage, exposure, and skill gaps to be tracked over time. (6) AI in education: school-level panels with variation in AI policy would test which guardrails preserve learning. (7) Mental health: administrative health records linked to AI exposure would test the stress channel with outcomes that self-reports miss. (8) Labour policy and social protection: cross-country variation in training, wage, and social protection responses would show which policies shape the distribution of gains.

For each topic, the report's datasets provide the starting point: the exposure measures, the survey series, and the experimental results are all open and downloadable, and the replication script reproduces every chart. The methodological message is the same as the substantive one: measure the constructs separately, respect the confounders, and let the evidence accumulate.

References

Data Availability

All data underlying this research brief are open access and downloadable from this page. Each CSV begins with comment lines documenting variable definitions and sources. A complete methodology document (README_methodology.txt) describes sources, compilation methods, and known limitations.

The replication script reproduces every chart and statistic in this report. It is written in Python with pinned dependency versions, sets an explicit random seed (42), auto-downloads the publicly available World Happiness Report data file with local caching, and reads the compiled datasets from the data folder. Generated charts are saved to the charts/ directory.

These data are provided for academic research use with appropriate citation. If you use them, please cite: Zhang, Y. (2026). Generative AI and Everyday Life: Wellbeing, Work, and Skills Before and After ChatGPT (2015-2026). Research Brief, Auckland University of Technology. Available at: https://zhangyuqian.com/ai-era-daily-life/.