DAPHNE
What happens to human judgement, expertise and purpose when AI becomes embedded in how we work, create, remember and decide?
A living research stream about the relationship between artificial intelligence and the human capacity to think, choose, question, create, change and take responsibility. AI does not simply increase or decrease human agency; it redistributes it.
A signature research stream connecting professional practice, workplace transformation, digital ethics, human rights, AI-assisted creativity and the ongoing Digital Twin experiment. It sits above companion work including Fear, Leadership and AI and Love, Leadership and AI.
Role:Rol: Researcher / Practitioner
Format:Formaat: Living research collection and interactive essay
Brings questions about AI work redesign, expertise, human oversight, memory, identity, creativity, distribution and responsibility into one research frame: what must remain possible for humans because of the way we choose to build and use AI?
Most discussions about AI begin with capability: what the model can do, what tasks agents can execute, how much faster work can be completed and how many jobs are exposed. These are important questions. But they start halfway through the problem.
The deeper question is: who becomes more able to act, and who becomes less able to shape what happens?
Philosophically, agency is commonly associated with the capacity to act intentionally; autonomy adds the stronger idea of self-government.1 I use human agency here more practically: the ability to form and revise intentions, exercise judgement, act on them, challenge what influences us and remain responsible for what follows.
A system can expand my options without giving me control over the conditions under which those options are created. It can reduce effort while making me less able to perform the task without it. It can recommend a decision while leaving me formally responsible for an outcome I no longer understand well enough to contest.
Capability is part of agency. It is not the whole of it.
It would be easy to begin this essay with a warning. That would also be incomplete.
AI has already enabled people to perform work that previously required more time, specialist knowledge or access to other people. The ILO's 2025 global index estimates that one in four workers worldwide is in an occupation with some generative-AI exposure. Its central conclusion is not mass replacement but job transformation, because most occupations still contain tasks requiring human input.2
The 2026 Microsoft Work Trend Index goes further and explicitly describes AI through the language of agency. In a survey of 20,000 AI users across ten countries, 58% said AI enabled them to produce work they could not have produced a year earlier. Microsoft also reports that nearly half of analysed Copilot conversations supported cognitive work such as analysis, problem-solving and evaluation.3
That matters. A junior employee can interrogate a dataset without waiting for an analyst. A founder can prototype something she could not afford to commission. Someone working across physical, cognitive or time limitations may be able to participate in work that was previously inaccessible.
I know this effect personally. AI has expanded the amount of ground I can cover. It helps me connect ideas across technology, law, identity, writing and projects that might otherwise have remained in separate notebooks. That is agency of a kind.
But there is a question hidden inside the success: if I can now do more, am I necessarily more in control of what I am doing?
One of the most useful findings in current human-AI research is that AI does not have a clean boundary between things it can do and things it cannot. The boundary is jagged.
In a field experiment involving 758 BCG consultants, people using GPT-4 on tasks within the model's capability frontier completed 12.2% more tasks, worked 25.1% faster and produced higher-quality work. But on a task deliberately placed outside that frontier, participants using AI were 19 percentage points less likely to reach the correct answer.4
The unsettling part is not merely that the machine can be wrong. Humans are wrong constantly. The unsettling part is that fluency can conceal the boundary between assistance and misdirection.
A 2025 CHI study of 319 knowledge workers, based on 936 reported real-world examples of generative-AI use, found that higher confidence in AI was associated with less reported critical-thinking effort, while higher confidence in one's own abilities was associated with more. The authors also found that critical thinking was changing form: toward verification, integration and stewardship of AI-generated material.5
Perhaps expertise is not disappearing. Perhaps part of it is moving: from producing the first answer to deciding whether the answer deserves to survive.
We often assume that removing a task frees the human for something higher value. Sometimes it does. But competence is not a cupboard from which a skill can be retrieved unchanged after years of non-use.
Judgement is partly built through exposure. You notice the strange contract clause because you have read hundreds of normal ones. You know a customer explanation feels wrong because you have spent years listening to customers. You recognise that a number cannot be right because you understand how the system producing it behaves.
The irony of advanced automation is therefore uncomfortable: the better the machine becomes at routine execution, the fewer routine situations humans may encounter through which they develop the judgement required for exceptional situations.
The 2026 Microsoft Work Trend Index reflects one side of this transition. Among surveyed AI users, quality control of AI output and critical thinking were the two most frequently named human skills becoming more important, and 86% reported treating AI output as a starting point rather than a final answer.3 Those responses are encouraging, but they are self-reported behaviours in vendor research, not evidence that skill retention will happen automatically.
Which activities can we stop doing safely, and which must we continue practising because they are how we become capable of intervening when the system is wrong? Efficiency answers the first half. Agency requires answering the second.
For a long time, professional expertise benefited from scarcity. I knew something you did not. I had access to information you could not easily retrieve. I knew which document mattered. I understood the language of a domain.
AI weakens some of those advantages. That can be profoundly democratising. But it also raises a difficult question for every profession built partly around informational scarcity: what remains valuable when knowledge becomes easier to reach?
Expertise increasingly includes knowing what question has not been asked, what evidence should change the answer, when context invalidates a sensible recommendation, where a model's confidence should not become your confidence, which trade-off has been hidden by optimisation, whose knowledge is absent and who will have to live with the consequences.
AI may reduce the premium on remembering some answers while increasing the premium on epistemic judgement. But that outcome is not guaranteed.
We could build organisations in which people become better directors of intelligent systems. Or we could create organisations full of people who know how to prompt systems they no longer know how to challenge.
I do not approach this question from outside AI. I work with it. I use it. I teach around it. I build with it.
AI has made me more capable. I do not want that capability to make me less able. I use AI in writing, yet I care intensely about authorship and about whether an idea still sounds like it has passed through a particular human life.
I believe good personalisation can create relevance. I also know that enough context, joined together, can become surveillance. I am building a Digital Twin because there is knowledge scattered across conversations, documents, experiences and decisions that I do not want to lose.
And yet that project has produced one of the questions I now care about most: can an AI remember me without deciding who I am?
Historical frequency should provide context. It should not become authority over future identity. A machine that knows me only by what I have done before could become exceptionally good at helping me remain the person I used to be. That is not the same as helping me grow.
Persistent AI changes the question from what does the model know to what does the system know about me?
There is enormous value in context. An AI that remembers a project does not need the project explained again. An AI that knows how I reason can challenge me more effectively. An AI that understands my previous decisions can surface inconsistencies.
But memory is not neutral infrastructure. Memory selects. Memory interprets. Memory creates patterns. And patterns can quietly become identity claims.
UNESCO's Recommendation on the Ethics of Artificial Intelligence places privacy, human dignity, accountability and human oversight among its core principles and states that AI systems should not displace ultimate human responsibility.6
For personal AI, I think the principle has to go further. A person should be able to say: that was true of me, but it is no longer authoritative. Remember the event, but not your inference about what it says about me. This part of my life does not belong in the system.
The right to memory may become important in personal AI. So may the right to revision. And the right to be forgotten by your own machine.
Creativity provides another version of the agency paradox.
A 2024 experiment published in Science Advances found that access to generative-AI ideas improved evaluations of short stories, particularly for people who initially scored lower on a creativity measure. But the AI-assisted stories were also more similar to one another.7
That finding should not be stretched into a universal law of creativity. The study involved one specific creative task. But the question it exposes is larger: what if AI raises the floor of individual creativity while lowering the variance of collective imagination?
An individual may write better. A team may brainstorm faster. A marketing department may produce more polished material. A professional who struggled to express an idea may finally be heard. These are real gains.
But if millions of people begin from related models, related defaults, related patterns of fluency and related ideas of what good looks like, everything becomes competent, readable and recognisable. Gradually, something strange disappears.
The human edge may not always be superior performance. Sometimes it is difference.
Governance language often reassures us that a human remains in the loop. I increasingly think that phrase is insufficient.
A person can technically remain in the loop while having almost no meaningful agency. The system has already gathered the data, selected the features, ranked the options, produced the recommendation and determined the pace of work. The employee has thirty seconds to click approve. A human is present. But what exactly is human about the control?
The EU AI Act's Article 14 requires high-risk AI systems to be designed so that they can be effectively overseen by natural persons, with oversight intended to reduce risks to health, safety and fundamental rights. Employment and worker-management applications can fall into the high-risk category; under the current post-Omnibus implementation timetable, those Annex III high-risk rules are scheduled to apply from 2 December 2027.8
NIST's AI Risk Management Framework similarly treats human oversight as part of broader risk governance, including impacts on individuals, groups, organisations and society.10
Ethical human oversight must ask an additional question: is the person actually equipped to disagree? Do they understand enough? Do they have time? Do they have authority? Will dissent damage their career? Can they inspect the evidence? Can they stop the process? Can the affected person appeal?
If the answer to those questions is no, human oversight risks becoming ceremonial.
One weakness in popular AI discussion is that we locate agency almost entirely inside the individual: learn AI, prompt better, become a frontier professional, adapt, experiment, stay relevant. There is truth in all of that. There is also power missing from the picture.
An employee cannot personally prompt their way out of a workplace that measures every efficiency gain as additional capacity. A worker cannot preserve autonomy if the algorithm determines pace, ranking and access to opportunities. A manager cannot exercise meaningful judgement if the organisation punishes any deviation from an automated recommendation.
The OECD tested worker consultation around algorithmic management in a laboratory experiment involving workers from three German manufacturing firms. The study found that consultation among workers, managers and works-council representatives could produce technology designs that participants considered capable of preserving productivity gains while improving job quality. The researchers explicitly caution that the limited number of firms, sectors and national setting require broader study.9
Participation is not merely change management. The ability to shape the system that will shape your work is itself a form of agency.
There is another danger in talking about the human as if everybody stands in the same relationship to AI. They do not.
The ILO's 2025 exposure estimates show significant differences across countries, occupational groups and gender. Globally, 3.3% of employment fell into its highest exposure gradient, but the estimate was 4.7% for female employment and 2.4% for male employment; in high-income countries, the gap was larger.2 Exposure is not displacement, but it tells us whose work may face more intense redesign pressure.
The person deploying the system and the person being evaluated by it have different agency. The executive who chooses an AI performance-management system and the employee ranked by it have different agency. The professional with paid access to advanced tools and the applicant screened by those tools have different agency.
Every claim that AI empowers people needs a second sentence: which people? Every discussion of risk needs the same discipline. Technology can widen access. It can also scale an existing asymmetry. Sometimes it does both at once.
When an organisation introduces AI, I do not think the most useful first question is what can we automate. I would begin with: what is moving between human and system?
The Human Agency Ledger is my working framework for making that movement visible. It is not a validated psychometric instrument and should not become another maturity score. It is a conversation tool by Daphne Iris van Vliet, 2026.
For each dimension, ask: what does AI expand, what does AI remove, what becomes newly possible, what becomes harder to recover and who has the power to decide?
The Ledger should not produce one score. The pattern is more important than the total. A system might dramatically increase capability while reducing voice. It might expand access to knowledge while creating dangerous dependence in judgement. It might improve individual agency while weakening the bargaining power of a group.
Agency has more than one account.
Technology design teaches us to dislike friction: fewer clicks, faster answers, shorter workflows, automatic completion, seamless experiences. Often that is excellent design. But not every form of friction is waste.
Sometimes the pause before a consequential decision is where judgement happens. Sometimes writing the argument rather than selecting one forces us to discover that we do not understand it. Sometimes verification feels inefficient because being careful is inefficient.
This does not mean we should romanticise difficulty. People have spent enough of human history wasting life on badly designed processes. But we should distinguish friction that obstructs agency from friction through which agency is exercised.
A future of perfect convenience could still be a future in which people increasingly accept decisions they did not meaningfully make.
Human agency should not become a new excuse for human exceptionalism. Humans are inconsistent. We forget. We discriminate. We rationalise. We miss patterns. We become tired. We protect our status.
There will be circumstances in which an automated recommendation is more accurate, more consistent or fairer than the human process it replaces. There may even be situations in which requiring human intervention makes the outcome worse.
Human agency therefore cannot mean preserving human control everywhere merely because the control is human. The objective is not maximum human involvement. It is meaningful human authorship of the systems, purposes, boundaries and responsibilities that matter.
Sometimes responsible agency will mean delegating. But delegation should itself remain deliberate.
We need clearer language for different relationships with AI: AI as tool helps me perform an action; AI as adviser shapes what I consider; AI as delegate performs an action I authorised; AI as representative lets other people experience its actions as mine.
These transitions are ethically significant. If an AI summarises a document for me, the primary risk concerns my understanding. If it recommends whom I should hire, another person's opportunity is involved. If it negotiates with a supplier, it is exercising delegated authority. If a digital version of me answers someone's question using my history, tone and views, it begins representing me.
At that point, accuracy is no longer enough. Representation requires permission, provenance, boundaries and the ability to say: this system may resemble my thinking. It does not possess my authority.
That is one of the principles guiding my Digital Twin experiment. The twin can remember. The human must still be allowed to change.
To value human agency is not to imagine the autonomous person as a perfectly independent individual who needs nobody. That person has never existed.
We learn language from others. We inherit culture. We depend on institutions. We build judgement through relationships. We borrow knowledge constantly. Agency is therefore relational.
The question is not whether we are influenced. We always are. The question is whether the relationships and systems around us leave room to examine that influence, respond to it, negotiate it and sometimes reject it.
AI will become one of those relationships, not because machines become people, but because systems will increasingly mediate our relationship with information, institutions and one another. That makes the design of those systems a question about human freedom.
Microsoft's 2026 Work Trend Index argues that as agents take on more execution, humans can gain more room to direct work, make decisions and own outcomes.3 I think that is one plausible future. It is not automatic.
Another future is possible: agents execute, humans approve, metrics accelerate, skills narrow and decisions become difficult to contest because nobody can quite explain where they came from.
Workers become more productive but less influential over the definition of good work. Executives receive more information but encounter fewer dissenting interpretations. Individuals produce more while trusting their own unaided judgement less.
Both futures can use exactly the same technology. That is why AI transformation is not fundamentally a deployment question. It is an institutional design question.
I do not think we need a definitive list of things machines should never do. Technology will keep making fools of lists like that. I am more interested in the conditions we should hesitate to surrender.
The ability to determine purpose. The ability to question. The ability to appeal. The ability to learn through doing. The ability to change our minds. The ability to distinguish our memory from someone else's interpretation of it. The ability to know when something is acting in our name. The ability to enter parts of life that are not continuously optimised, measured or predicted.
And responsibility. Especially responsibility.
UNESCO's AI ethics framework makes the point institutionally: ultimate human responsibility and accountability should not be displaced by AI systems.6 I would make the personal version even simpler: do not give a system authority merely because you gave it work.
I am not trying to preserve a world before AI. I do not want it. There are things this technology allows me to do that I would not willingly give back.
I want AI that remembers enough context to help without claiming ownership of identity; AI that makes expertise more accessible without pretending expertise no longer matters; AI that removes drudgery without removing every opportunity to practise judgement; AI that helps people express themselves without quietly teaching everyone the same voice.
I want AI that can act on our behalf while remaining visibly subordinate to human responsibility, and AI that gives someone with less power more ability to participate—not merely someone with more power a cheaper way to manage them.
The goal is not human control over every machine action. Nor is it maximum automation. It is technology that expands what humans can do without quietly shrinking their ability to decide who they become.
Before asking where can we deploy AI, ask where agency should sit when we do.
Who defines the purpose? Who exercises judgement? Who still develops expertise? Who can challenge? Who remembers? Who benefits? Who is represented? Who can refuse? Who repairs the consequences?
And then one question I think will become increasingly important: if this system succeeds exactly as designed, what kind of human does the surrounding organisation require people to become?
More capable? More curious? More responsible? More dependent? More measurable? More interchangeable? More free?
Technology rarely answers that question for us. But it can make our answer difficult to reverse.
This is a practitioner-led research essay and evolving research stream, not a systematic literature review, philosophical theory of agency or legal opinion.
I combine evidence from labour economics, organisational research, human-computer interaction, AI governance and professional practice with original conceptual work. The empirical literature is developing quickly. Results from one occupation, organisation, model generation or experimental task should not automatically be generalised to other settings.
Corporate research is useful but has incentives and methodological constraints that should remain visible. Self-reported behaviour is not the same as observed behaviour. Exposure to AI is not job displacement. Productivity is not automatically job quality. Human participation is not automatically meaningful agency.
Human judgement should not be presumed superior merely because it is human.
The Human Agency Ledger is an original conceptual tool by Daphne Iris van Vliet. It has not been psychometrically validated and should not be used to score individuals or make consequential employment decisions. Its purpose is to make questions visible before efficiency makes them disappear.
A research-led essay and practical framework on fear, identity, responsibility and human agency during AI transformation.
Find out more →Meer weten →A research-led essay arguing that AI makes love more structural in leadership: a disciplined commitment to the dignity, agency and flourishing of the people affected by technological decisions.
Find out more →Meer weten →Take one real AI-enabled workflow and map it through the Human Agency Ledger: Intent, Judgement, Knowledge, Skill, Voice, Memory, Responsibility and Distribution.