AI and Productivity – Evidence, Recommendations and Ethics

There’s a word around AI you rarely hear at conferences – PRODUCTIVITY. If we fail to understand what is happening here, all the discussions around what we do ethically, politically and economically will misfire. Whether it is in education, training, healthcare or any other sector, there has been a plateauing of productivity. AI promises to solve this problem.

Everyone has views on ethics, often expressed in phrases such as ‘humans-in-the-loop’ and ‘AI won’t take your job but someone with AI skills will’. In truth, humans are already being taken out of the loop. They have been for hundreds of years, as automation moved work from fields to factories, then factories to offices. These massive shifts were caused by increases in productivity. We all became better off, more comfortable, able to achieve more in our lives and live longer.

After 20-odd years of technological escapism, where technology has thrown ad-driven services, fractious social media, hailing Ubers, fast food delivery, gaming, gambling and porn at us, we at last have a dramatic shift into the real economy with intelligence on tap, productivity tools, uses in education and health, accelerated research and agentic productivity. This should surely be welcomed?

EVIDENCE FOR PRODUCTIVITY

The evidence coming through now is that entry-level jobs in white collar and graduate work are drying up. Lichtinger (Aug 2025), in Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data used a huge dataset, from 285,000 US firms, with over 150 million jobs, 62 million workers and over 245 million job postings from 2015 to 2025 to analyse the shift. The findings were clear.

The first to be hit were coding jobs. Senior developers continue to find jobs but junior developer hires are on the decline, as their jobs are being automated. A more serious conclusion, which the above study showed, is that this applies across many sectors, such as manufacturing, retail, finance, professional services, healthcare, and education.

Of particular interest is the University sector, where there is a U-curve trend. Low tier institutions, that tend to offer more vocational options are fine, as are a few at the very top (Stanford, Harvard, MIT). The vast majority in the middle are in trouble. Several crises are hotting at once, on increased costs, increased administrators, crisis of relevance, changes in revenues from foreign students and so on.

Having spent the last year looking specifically at the evidence for a book on ‘AI and Productivity’ (published 3 Nov), it has become clear that this technology has surpassed forecast expectations. Few expected such rapid progress on the outputs’; on released models in general, reasoning models, research models, agentic capability and coding proficiency. Multimodality in producing images, now video has exploded. Costs for AI services have plummeted and quality has increased. On energy efficiency, we’ve seen a vast improvement at around 0.0003 kWh per standard prompt (8-10 secs of watching Netflix). Getting more specific, in just one domain mathematics. Forecasts in 2022 of winning Gold in the Mathematics Olympiad were 2.6 and 8.6%; several models have just won Gold.

How has this impacted productivity? A phrase often used in AI is ‘the jagged frontier’, from a paper by Dell’Acqua in 2023, one of the earliest papers showing significant increased productivity in high-end cognitive tasks. They pointed out that the capabilities of AI were still variable or ‘jagged’ across tasks. Now that AI’s capabilities are growing exponentially, productivity gains are being seen in almost all sectors.

Yet it is still jagged, as there are many obstacles and barriers in specific sectors that act as brakes on productivity. These are human, institutional and economic, which show paradoxically, that even when the productivity gains are clear, we still adhere to old behaviours, processes and institutional resistance.

PRODUCTIVITY PARADOXES

Since 2008, the productivity paradox has lingered, as technology races ahead while measured output per worker stalls, because adoption is filtered through human, institutional and economic contradictions.

At the personal level, behavioural paradoxes mean we resist change, cling to rituals and misread what actually creates value; the bias paradox shows cognitive biases (confirmation, status-quo, loss aversion, anthropocentrism, negativity) make AI feel threatening even as it helps; the procrastination paradox captures our tendency to delay valuable work despite knowing better; and the perfection paradox shows that chasing flawless outcomes cripples throughput. In organisations, the busyness paradox confuses activity with results, while Parkinson’s paradox (Parkinson’s Law and the law of triviality) explains how work inflates to fill time and attention fixates on the insignificant; the boiled frog paradox warns that gradual inefficiencies become invisible; and the Clark paradox notes that AI boosts individual productivity even as institutions suppress its use.

On the technology side, technology paradoxes (including capability overhang) remind us that even obvious productivity tech diffuses slowly; the Solow paradox observes that we “see computers everywhere but not in the stats” until complements and measurement catch up; and the legacy paradox shows old platforms ossify process and smother gains. In capability terms, the Moravec paradox points out that AI aces abstract cognition yet struggles with sensorimotor tasks, slowing gains in the physical world.

Economically, the Jevons paradox predicts efficiency can raise total consumption and workload; the Friedman paradox (“spoons not bulldozers”) exposes the fallacy that more labour equals more value; the Easterlin paradox shows higher income doesn’t guarantee greater wellbeing; Turchin’s paradox warns that rising productivity can widen inequality and fuel instability; and the Pollyanna paradox cautions that techno-optimism blinds us to frictions, lags and unintended effects.

AI can help resolve these tensions by automating routine work to puncture busyness and procrastination, using copilots and agents to set priorities and enforce “good-enough” outputs against perfectionism, embedding bias-aware workflows that counter cognitive traps, refactoring or front-ending legacy systems to unlock process flow, improving measurement so Solow-style lags shrink, extending from bits to atoms to chip away at Moravec constraints, and designing policy and deployment that blunt Jevons, Turchin and Easterlin effects e.g., pricing energy and compute, sharing productivity dividends, and optimizing for wellbeing, so that adoption becomes faster, fairer and visibly felt in the numbers.

JAGGED FRONTIER

Sectors differ in their uptake of productivity. Across this jagged AI frontier, progress advances unevenly: in resisted zones, deeply embedded habits keep the line fixed (typical of creative media and the arts); static fronts stay put for lack of senior buy-in, even as covert use spreads (notably in education where learners adopt AI while teachers hold the line); spiky fronts move in fits and starts across functions, with broad gains but irregular across consultancies, marketing, coding, law and policing; broad fronts see widespread, steady adoption with large gains (as in healthcare); breakouts deliver dramatic wins on narrow research problems (e.g., AlphaFold’s leap to 200 million protein structures); and surges open entirely new domains as embodied AI pushes into robots, self-driving, drones, satellites and space.

AI has matured into a serious productivity stack; poems to business plans, to research and full-stack coding, to agents that run workflows with memory and bigger context—getting better, faster and cheaper; it began by helping us do our jobs and is now reshaping the jobs themselves, with clandestine, bottom-up adoption often outpacing policy because it works. Evidence shows the gains are real but uneven: Hampole et al. (2025) found that a one-standard-deviation rise in AI use in established firms links to +12% productivity, +28% sales, +38% profits and higher wages, with AI exposure explaining roughly +14% employment growth in public companies as tasks are automated but roles shift; Noy & Zhang (2023, MIT) show professionals using GPT cut task time by 0.8 SD while quality rose 0.4 SD, indicating faster and better work and freeing attention for higher-value problem-solving and innovation, an inflection point where concrete wins are on the table for those who integrate, upskill and redeploy.

So how do we improve productivity in teachers, learners, managers and employees? We also have evidence on what leads to both success and failure. It turns out that policies, frameworks and courses on AI are not the answer. Key factors in success include top-down support from leadership, as well as strong communications and a focus on ‘doing’, using AI to get things done in your own teaching, study and work. Also, clarity on deployment. Rather than humans-in-the-loop, it turns out that experts-in-the-loop, once they understand the nature of AI, its capabilities, not search but dialogue, data analysis and research, really do get significant increases in productivity.

ETHICS

A strong focus on productivity cuts through utopian–dystopian hype and grounds AI ethics in real people, real organisations and real trade-offs, distinguishing noisy worries from serious consequences. AI is already reshaping work, institutions and geopolitics as cognitive systems and embodied robotics scale to billions of uses; in practice, the world is likely to be reshaped by AI more than AI is reshaped by politics. Today’s models largely encode Enlightenment ideals, such as rationality, open discourse and the scientific method, because their data and interfaces privilege those norms; that brings intellectual as well as economic leadership but also raises questions of cultural pluralism.

Fears that AI entrenches privilege are countered by its extraordinary diffusion via phones, falling costs, open models and rapid gains in non-Western languages, alongside community efforts to capture oral traditions, build low-resource datasets and localise tools; inclusion is imperfect but accelerating. The ethical crux is distribution: productivity booms can raise living standards yet still create unemployment, turbulence and sharper inequalities if gains are hoarded. Democracy therefore matters as the moat that keeps decisions about sharing the bounty, reskilling and safety in public hands.

Planetary ethics demand energy realism: training and inference consume power, but AI can out-produce humans with far lower marginal emissions for many knowledge tasks and can optimise buildings, data centres, batteries and grids, so the right metric is net productivity-per-joule across full lifecycles.

The gravest risks cluster where productivity meets force: militarised autonomy, swarming drones, cyber and bio enable ‘productive’ violence; beyond that lie loss-of-control scenarios where agentic systems pursue misaligned goals.

The answer is proactive control and alignment: tested guardrails, monitoring of internal reasoning, human approval points, graduated control levels, rigorous evaluation and institutional oversight that evolve as capabilities do. Moral panics will ebb as we adapt, by measuring, governing and redesigning work, but the choice remains stark: stagnation with mounting social costs, or growth with the ethical work of broadening access, cushioning transitions and constraining misuse. Treated as a practical project, where we deploy, measure, mitigate, and redistribute, AI can be a humane amplifier of shared prosperity rather than a driver of division.

I shall be delivering a session on AI and productivity, to pass on both what the research shows and practical advice on how to implement and get productivity gains in your institution and organisation at OEB this year, with a large discount on the book, and even some free copies. Look forward to seeing you there.

Written for OEB 2025 by Donald Clark.


Join Donald for his Learning Café “AI and Productivity – From Attention Economy to Productive Economy” at OEB25.