Legibility and AI: Restructuring of Knowledge Work

This is a follow-up to The Legibility Trap.

In The Legibility Trap, I made the argument that that AI does not set new standards for all. It sets them for some particular tasks that are observable and assessable. This deserves more attention and analysis. What does it mean for a task to be "AI-legible"? What are the consequences of unexpected changes in standards? And who suffers from the moving benchmark faster than workers can follow?

In this piece, I take such issues seriously not as a game but as problems with concrete evidence and policy implications.

The Nature of AI-Legibility

Before talking about changes in standards for "AI-legible tasks," we need to define the term. Political scientist and anthropologist James C. Scott gives us the concept in his book Seeing Like a State: How Certain Schemes to Improve the Human Condition Have Failed (1998). Scott says that "legibility" means a transformation of complicated reality into something simple, standard, and visible by a centralized authority. The idea was developed in terms of states, but is easily transferable to organizations using automated systems of evaluation.

We commonly refer to AI-legibility in situations when the output of a task is (1) clear and text-based, (2) comparable to a set of examples or template, (3) scalable using language model and (4) evaluatable without contextual information that the system cannot get. These outputs would include reports, emails, code, slide decks, and literature reviews. Strategic negotiation, mentoring, conflict resolution, and institutional trust-building activities are not included (yet).

The main point is that AI-legibility is not a fixed characteristic of the nature of work. It depends on technological development. As the AI develops, there are more and more legible tasks. Those jobs that before needed judgment by a human being – medical coding, legal document review, financial analysis – become legible.

It is not happening evenly. However, it is happening in one direction. It is important for understanding of productivity statistics, performance reviews and labor-market outcomes in the next decade.

The Attribution Gap Between Effort and Output

Standard economic theory says productivity is the function of effort, skill and capital invested. Wages reflect marginal product of the worker – the addition of the worker to production. It assumes the employer's ability to observe effort and connect it to output.

With generative AI the connection of effort and output is lost in a very particular way. When the quality of output is high, there is no guarantee that it is generated due to the durability of skills of a person. It may happen due to the skill of using AI that is human skill as well but less stable. In turn, poor output may be generated by the person using their judgment to deal with ambiguity, maintaining relationships, and making decisions that cannot be found in the final product.

There is an attribution gap between AI-generated output and invisible judgments. Organizations that pursue the former will gradually lose sight of the latter. Even if that judgment is crucial for functioning of the organization.

This has implications not only for distribution but for structure. If the valued employees are those producing legible output, organizations tend to value and retain people suitable for such a world. They undervalue the judgments, adaptability, and social skills needed in future.

Unannounced Standards as Structural Change

"Nobody issued a memo" is not just a rhetorical device. It describes the reality of standards changing in complex organizations.

Researchers of work know for a long time that there is a gap between formal job descriptions and informal expectations. Formal standards change through the contracts and HR regulations. Informal standards change through accumulation of precedent – the expectations of managers, modeling of peers, what is rewarded, what is quietly penalized. The informal system changes faster than the formal one and it is influenced by technological development. However, workers learn about it only after the gap started hurting them.

What is new with AI is the speed and the process itself. While in previous technological transitions – spreadsheets, e-mail, enterprise software – informal standards shifted through years or decades and workers had time to adapt to it, however not necessarily equitably, the shifting of standards in AI-legible task work can happen in months. The worker that was doing well in 2022 is likely to be below the new, unspoken standard in 2 and half years because his output is compared to the output augmented by AI.

This is not the question of individual adaptation. It is the problem of the way how organizations communicate with workers the standards, their duties under changing the implicit contract and whether the labor regulations, trade unions, and regulators can react fast enough.

Costs Distribution

Costs are not evenly distributed across the population and knowing who is suffering is crucial for policy making.

The most vulnerable are those who do not have access to AI or who do not have the skills of working with it. If the standards assume augmented output, then those without the tools, skills and comfort working with AI do not meet the standard they were not supposed to meet. The group includes older workers, people who work in under-resourced organizations, workers with no AI access in their roles, and people living in regions where AI tools are not widely used.

People in mid-legibility roles face another risk: the legible part becomes commodity and invisible part stays hidden. This reduces the value of the work while total contribution remains the same or increases.

Those workers that adopt the AI successfully become more productive and faster. However, the increase of supply decreases wages for the task-type. The early adopters benefit individually while average standard for performance declines without higher wages.

All these patterns are familiar from the previous technological wave but they happen faster and affect larger part of the population. The workers most exposed to changes usually lack the power to push for changing the standards and the unexpected movement of the benchmark prevents such actions.

Policy and Institutional Implications

Careful consideration of the issue suggests a few policy and institutional recommendations.

Make standards explicit. Organisations using AI-augmented output should declare these expectations. It is not only ethical but required by law in the jurisdictions where the labor legislation is well-developed. Some scholars argue that algorithmic management triggers certain duty of disclosing that is not covered currently by existing regulations.

Treat AI-fluency as a public good. If AI-legible tasks become the core skill in knowledge work, the access to AI tools and the training should not be left to the employers. The dissemination of the computing literacy happened thanks to public investment in education. AI-fluency is disseminating unevenly. Therefore, public policy can help to correct it.

Re-evaluate the measures of productivity. With the appearance of AI the connection between effort and output starts deforming. Traditional measures – GDP per hour worked and total factor productivity may not capture the underlying reality of AI-augmented work. New measurement techniques are needed to distinguish between AI-augmented output and growth of human skill, and to measure the value of invisible part of the work.

Re-think the purpose of work. If AI can generate legible output quickly, human workers should concentrate on doing the invisible part – judgment, trust and social aspects of work. It is not new but becoming urgent. If organizations will persistently pursue the legible output they will mischaracterize human value creation and its conditions.

Conclusion

The benchmark moved. This is a real thing. However, the question arises: moved for whom, by which mechanism, and at whose costs?

The notion of AI-legibility allows to investigate these questions carefully. The expanding of the legible domain is not neutral technology. It changes the way how organizations observe and assess the work and it has major impacts on labor markets, organizations and regulation of work.

The workers that are the most exposed to changes usually have the least power to push for the explicit and clear contract. This is a policy problem. Unnoticed moving of the benchmark is not an inevitable consequence of AI usage. It is an organizational choice with predictable consequences.

The question is whether the institutions will react quickly enough to the technology. There is no sufficient evidence on that.