AI, being a GPT (General Purpose Technology) like electricity, has an increasingly important role to play in all three – hence it needs to be managed well – hence it needs to be measured well.
The DIGIT article was entitled:
‘AI is Rendering The Way We Measure Productivity Obsolete’
It covers a new study from Microsoft which posits that AI systems, capable of scaling, learning, and producing economic value, should be recognised as a distinct factor of production.
They call that factor ‘digital labour’.
INTRODUCTION:
- For decades, economists have described productivity with a simple equation – inputs like human labour and physical capital create outputs like goods and services.
- A new study argues that this framework is missing a critical variable – one that’s already transforming industries, yet remains invisible in most economic metrics.
- That variable is digital labour – the autonomous, cognitive work performed by AI systems – evermore prevalent in tech stacks, from writing code to diagnosing diseases.
- The paper, Evolving the Productivity Equation by Microsoft’s Alex Farach, Alexia Cambon, and Jared Spataro, posits that treating AI as mere software (or worse, burying its contributions in statistical “black boxes”) obscures its true economic role.
- Unlike traditional tools, AI doesn’t just augment labour – it behaves like labour, at least in terms of its measurable output. It scales exponentially, learns from experience, and deteriorates rapidly (based on data that can quickly become unfit for purpose.
- Just as economists once expanded their models to account for “human capital,” the study suggests it’s time to formally recognise digital labour as a new factor of production.
- Gross Domestic Product (GDP), designed, broadly speaking, for an era of factories and tangible goods, struggles to quantify AI’s intangible outputs.
- When an algorithm optimises a supply chain or a generative AI drafts legal contracts, the economic gain often vanishes into a statistical residual called Total Factor Productivity (TFP) – a catch-all category for things we struggle to group into measurable value – (some of us always struggle with it).
- This creates a paradox: even as AI accelerates innovation, official productivity statistics may stagnate.
- The study points to examples like healthcare, where AI systems now match or exceed human performance in tasks like medical imaging analysis. Yet these advances rarely appear in national accounts, because their outputs – faster diagnoses, reduced errors – don’t translate neatly into traditional metrics like units produced or hours worked.
- Corporate accounting faces similar blind spots.
- AI models trained on proprietary data can be worth billions, yet rarely appear on balance sheets because accounting standards classify them as expenses rather than assets.
- The result, as the study notes, is a systematic undervaluation of the digital economy’s engine.
What Makes Digital Labour Unique?
The paper identifies five traits that set AI apart from traditional inputs:
1 – Intangible & Non-Physical:
- AI isn’t something you can see or touch.
- It exists as code, data, and mathematical weights – an intangible asset with no physical form.
- A cutting-edge AI model might contain hundreds of billions of parameters, spread across global cloud infrastructure and refined through constant updates and user interaction.
- Unlike traditional assets, AI produces insights – decisions, forecasts, language, and designs – rather than physical goods.
- This makes it difficult for conventional accounting and national statistics, which often undervalue or ignore intangible assets.
2 – Unprecedented Scalability:
- Traditional human labour and physical capital have natural limits.
- A person can only work so many hours or serve so many customers, and a machine can only perform one task at a time in one place.
- Digital labour, by contrast, is highly scalable.
- Once developed, an AI model can be used by one person or a million with little extra cost or loss in performance.
- An AI customer service agent, for example, can handle 1,000 queries just as easily as one – something no human agent could match.
- Economically, digital labour is non-rivalrous – one AI model can be used by many people or for many tasks at once without losing effectiveness – however, it is still excludable – access can be limited through licences, subscriptions, or other controls.
- This combination – non-rivalry and practical excludability – gives digital labour its unique scalability, driving strong network effects and the potential for rapid economic growth.
3 – Autonomous Learning & Self-Improvement:
- Perhaps the most distinctive feature of AI as a factor of production is its ability to improve through use.
- Human labour also improves with training and feedback, but learning is slow, biologically limited, and costly to scale.
- Physical capital, on the other hand, wears out over time.
- With the right data, usage context, and human guidance, modern AI systems can refine their models, adapt strategies, and produce increasingly accurate or insightful results.
- In other words, the productivity of digital labour can grow as it’s used.
- For example, an AI code assistant might learn from every snippet of code it helps write, gradually offering better suggestions as it learns from feedback.
- This self-improving quality gives digital labour the potential for increasing returns — the opposite of what we expect from traditional labour or capital.
- It also creates a network effect – the more an organisation uses and refines its AI, the more valuable it becomes, setting off a virtuous cycle of improvement.
- But this same adaptability can also be a source of risk.
4 – Rapid Depreciation and Obsolescence:
- Unlike traditional capital, which tends to depreciate at a steady and predictable rate, AI can degrade rapidly – and often invisibly – through two key mechanisms:
- First, the quality of digital labour can decline as data becomes outdated, training signals weaken, or models are retrained on their own synthetic outputs, leading to performance drop-offs akin to mechanical wear and tear.
- Second, AI systems are prone to becoming obsolete – often suddenly. Advances in algorithms, such as model distillation, reinforcement learning techniques, or improved prompt engineering, can render existing models outdated almost overnight.
- Treating digital labour simply as another form of capital overlooks these dynamics.
- The same AI systems that improve through use can just as easily become obsolete without ongoing maintenance and adaptation.
- This volatility is what sets digital labour apart from both physical capital and traditional labour, making it a uniquely complex and strategically demanding production factor.
5 – Substitutability with Human Labour:
- AI’s relationship with human labour is highly elastic and deeply dependent on context.
- In some cases, AI can fully replace people by acting as digital labour; in others, it enhances and supports human work.
- One system might automate production scheduling entirely, while another might help a designer with ideas—boosting output without replacing the person.
- Economists refer to this variation as the elasticity of substitution, which differs widely by task:
- In structured, rule-based environments, digital labour can be a near-perfect substitute – often faster, cheaper, and more reliable. For instance, large language models now include the correct diagnosis in 88% of differential diagnoses – on par with clinicians, while achieving 80% accuracy in identifying skin conditions when combining text and image inputs (the paper cites from other research for these stats).
- In more creative, interpersonal, or judgement-based roles, AI is more likely to act as a collaborator.
- This pattern differs from traditional capital and labour dynamics.
- While machines have replaced repetitive manual work, they’ve also long served as complements – amplifying human capabilities in communication, computing, and transport.
- Digital labour takes this further: it scales instantly, performs cognitive tasks, and works across many processes at once.
- An AI system might replace the output of many workers – or, conversely, a single worker might manage several AI agents. AI doesn’t simply replace one job with one model. It can dramatically increase productivity in some areas, leave others unchanged, and even generate entirely new kinds of work.
- Offloading certain tasks can make remaining work more valuable – especially when it involves scarce expertise.
- But there’s no guarantee those remaining tasks will be better paid or more secure.
- If they become commodified or undervalued, wages and employment may fall even as productivity increases.
- In this sense, AI is not just a tool – it’s a force for reshaping how work is allocated, which skills matter most, and how expertise is rewarded.
- Digital labour comes with its own costs and trade-offs.
- But unlike traditional inputs, its substitutability with human labour is more elastic, more uneven, and far more dynamic.
The Stakes:
- The study warns that clinging to outdated models risks two traps:
- Misinvestment – Companies may underfund AI initiatives because their returns are invisible in traditional accounting.
- Policy lag – Governments relying on flawed productivity data could misdiagnose economic health, delaying investments in AI infrastructure or education.
- But for firms that adapt, the upside can be transformative.
- Explicitly valuing digital labour lets businesses allocate it strategically, measure its ROI and, critically, pair it with human labour where both thrive.
- The goal isn’t replacement, the researchers stress, but recombination
- Using AI to handle scale and speed, while humans focus on judgement, ethics, and innovation.