The realisation of intentional, values-aligned action through designing and building your own environment.
or, the human answer to AI-driven automation
"We don't need you anymore - we have a Chatbot!"
Translators are being replaced by machine translation. Programmers are being replaced by code generation. Lawyers are being replaced by semantic search engines. If someone can build an automation that does what you do, faster, cheaper and better they will. And your employer will buy that automation.
The impacts of this are not only felt by those who have already been made rudundant. Knowledge workers still in a job report deskilling2Ferdman (2025)p. concept noteCapacity-hostile environments are instances where AI mediation impedes human capacity cultivation.Supports the claim that AI 'quietly degrades the foundational competences they never fully develop'.Greengard (2025)p. full articleEndoscopists who routinely used AI performed worse when the technology was unavailable, with detection rates dropping from 28.4% to 22.4%.Concrete evidence for the 'illusion of expertise' and deskilling claims.Greengard (2025)p. full articleLaw students using AI chatbots showed higher rates of critical errors.Evidence that AI boosts apparent output while degrading foundational competences.: AI boosts their apparent output while quietly degrading the foundational competences and inhibits further development3Quattrociocchi et al. (2025)p. 7-8Despite these fault lines, LLM outputs often appear superficially fluent and confidently articulated, even when they are factually wrong. This creates substantial room for an 'illusion of veracity', a systematic divergence between the actual accuracy of an LLM output and the accuracy perceived by a human user.Supports the 'illusion of expertise' claim; fluency substitutes for epistemic reliability.Quattrociocchi et al. (2025)p. 8-9The core harm is not the production of falsehoods, but the structural bypassing of evaluation itself. When answers are delivered in finalized form, without visible traces of uncertainty, conflict, or evidential grounding, the user is placed in a position of epistemic passivity. Judgment is not exercised; it is consumed.Supports the claim that AI creates 'an illusion of expertise without the judgment that expertise requires'.. Making them even less competitive in the future.
The discourse is also commodifying expertise itself4Quattrociocchi et al. (2025)p. 2The result is an architecture that produces answers that look increasingly reliable without possessing the machinery that normally makes reliability possible. The system becomes more convincing rather than more knowing.Supports the commodification claim: outputs appear to capture domain knowledge when they only reproduce surface form.. When a chatbot can produce a passable first draft of a legal brief, a marketing strategy, or a research summary, the assumption follows that the domain knowledge behind those outputs has been captured by the model. The professional who spent years acquiring those competences watches their work reduced to a prompt.
Why study for years when a model already appears to know everything and can do anything? In the canton of Zurich, new ICT apprenticeships dropped 12% in 2025, with application development down 25%1Inside IT (2025)p. articleIm Kanton Zürich haben im August 679 junge Menschen eine Lehre in Informatikberufen begonnen. Einer Mitteilung des Verbandes ICT-Berufsbildung Zürich zufolge sei das ein Rückgang um 12% zum Vorjahr.Empirical evidence that the replacement narrative is reshaping educational decisions in Switzerland.Inside IT (2025)p. articleDie meisten neuen Lehrverträge wurden zwischen Lernenden in der Applikationsentwicklung und Ausbildungsbetrieben geschlossen. Die Zahl von 219 entspreche einem Rückgang von 25% im Vergleich zu 2024.The sharpest decline is in the discipline closest to what generative AI automates.. Universities face declining confidence in the value of credentials.
Is there any hope for the future? I would like to suggest that there is.
I would like to argue that the above framing is wrong. The error is not in overestimating what AI can do because it can be genuinely capable. The error is in misunderstanding what the expert does. The expert is not a entity who produces outputs at a given rate of productivity. The expert is a person whose knowledge, judgment (automated reactions to highly complex situations), values, and experience (highly inter-connected knowledge network and skills base) allow them to design the conditions under which good outputs are produced5OECD (2019b)p. 74Epistemic knowledge is the understanding of how expert practitioners of disciplines work and think. This knowledge helps students find the purpose of learning, understand the application of learning and extend their disciplinary knowledge.Supports the argument that expertise involves 'designing the conditions' rather than just producing outputs., to evaluate whether those outputs meet the standard6Quattrociocchi et al. (2025)p. 8The defining mark of Epistemia is the decoupling of content from evaluation. In human cognition, judgment is embedded in an epistemic loop: claims are checked against evidence, beliefs collide with counterexamples, and conclusions remain revisable in light of new information and social feedback.Supports the claim that experts 'evaluate whether those outputs meet the standard'., to plain future actions based multiple internally simulated future states, and to take accountability for the past, present and future.
The concept of augmented actualisation is my attempt at reframing the expert as the designer of an environment[here I could footnote the fact that, unlike in previous literature, I am not using the term system, or tool, or software - my consept takes a more wholistic approach] in which their intentional, values-aligned action is realised at a scale and speed that was previously impossible.
"My design augments my environment. My environment amplifies my potential."
The expert designs three dimensions of their environment to support their individual and their community's context-dependent needs.
1. Data
You decide what data enters the environment, how it is structured, and in what form it is stored. Format choices are design decisions: plain text (.md, .json, .csv, .xml) keeps data accessible, inspectable, and processable by any tool, whereas binary formats (.docx, .pptx) lock data inside opaque containers that resist integration. Internal structure matters as much as format: a markdown document with a clear heading hierarchy reflects the semantic structure of the content; a JSON file with well-chosen labels maps directly to categories in downstream use cases. The data structures are decided on with processes, relationships and possible automations in mind and are constantly re-worked as the environment evolves.
2. Processes
You decide what you do with your data and and how you do it. Finding the right data-structure-task-process match is a sophisticated design choice.
Processes differ by the source of variation in their output. Four levels:
- Deterministic: same input -> same output, every time. An if/else branch, a regex extraction, a database query. The output is fully determined by the input and the rules. It is transparent, auditable and replicable.
- Stochastic-constrained: the output varies within bounds defined by a known probability distribution. A Monte Carlo simulation7Notep. A technique that uses repeated random sampling to estimate the probability of different outcomes. Named after the Monte Carlo casino, it was developed during the Manhattan Project by Stanislaw Ulam and John von Neumann (1940s). Used in finance (risk modelling), physics (particle simulation), engineering (reliability analysis), and any domain where analytical solutions are intractable but outcomes can be sampled.Example of a stochastic-constrained process: each run differs, but the aggregate behaviour converges on a known distribution., a statistical sampling process. The individual result is unpredictable, but the aggregate behaviour is precisely characterised.
- Associative-stochastic: the output varies, shaped by learned associations between patterns8Quattrociocchi et al. (2025)p. 1LLMs are not epistemic agents but stochastic pattern-completion systems, formally describable as walks on high-dimensional graphs of linguistic transitions rather than as systems that form beliefs or models of the world.Supports the claim that associative processes differ categorically from intentional processes.Quattrociocchi et al. (2025)p. 3Text generation in large language models can be described as a stochastic process evolving on a discrete, high-dimensional state space... Each output is therefore the realization of a stochastic trajectory generated by local sampling in this state space.Provides formal characterisation of LLM output as stochastic, supporting the 'associative' process type.. This is where language models operate. On inference, the model's final hidden state, a high-dimensional vector encoding everything the model has attended to, is compared via dot product against the embedding vector of every token in the vocabulary9Vaswani et al. (2017)p. Section 3.4We use learned embeddings to convert the input tokens and output tokens to vectors of dimension d_model. We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation.Technical basis for the associative-stochastic characterisation: learned associations (embeddings) produce the probability landscape from which the next token is sampled.. The resulting probability distribution maps tokens whose embeddings are close to the hidden state to high probability, those that are far to low probability. This mapping from hidden state to probability is learned through training. Here that the associative and the stochastic converge. The associations (meaning) shape the landscape of probabilities through learning, and the output is sampled from it. The output is plausible, coherent, often useful, but not grounded in direct external verification10Quattrociocchi et al. (2025)p. 3In a generative system, producing content that is ungrounded with respect to external reality is not the exception; it is the default operational state.Supports the claim that associative processes are 'not grounded in any external verification'.. Reasoning distilled from the abstract mental modelling of authors and intention distilled from the description of people's actions and self-reflection are captured in the embeddings of our language. Embeddings are not actual abstract mental modelling or embodied intention11Quattrociocchi et al. (2025)p. 2What is being automated here is not cognition but language. Large language models operate on statistical regularities extracted from human-produced text, not on representations of the world.Supports the distinction between associative (pattern-based) and intentional (belief-based) processes.Grindrod (2024)p. Section 3The task that LLMs are set in pre-training is usually just one of missing word prediction... The thought is that the prediction task is merely intralinguistic, i.e. it is a task focused on probabilistic relations between words, and so is not focused on features of the world.Supports the claim that LLMs lack the grounding necessary for intentional action.. All meaning is reduced to the bandwidth of language.
- Intentional: the output is guided by the full complexity of human agency. Intentional action operates across multiple dimensions simultaneously. It is rational, based on considered reflection on the state of the situation12OECD (2019b)p. 64-65Taking responsibility is connected to the ability to reflect upon and evaluate one's own actions in light of one's experience and education, and by considering personal, ethical and societal goals... Acting responsibly implies considered reflection and asking questions related to norms, values, meanings and limits.Supports the argument that intentional action involves ethical frameworks and values-based decision-making.. It is ethical, grounded in one or many frameworks of ethics that the agent holds and can defend13OECD (2019a)p. 6Agency can be exercised in nearly every context: moral, social, economic, creative. For example, students need to use moral agency to help them make decisions that recognise the rights and needs of others.Supports the claim that intentional action is 'ethical, grounded in one or many frameworks of ethics'.. It is values-based, shaped by beliefs, commitments, and priorities that persist across contexts. It is emotional, influenced by affect, intuition, and felt response14Gehrman & Schwenkler (2024)p. 2Dreyfus proposed his own account of absorbed coping, which involves 'a kind of intentionality that does not involve content at all'.Intentionality without content parallels the claim that intentional action cannot be reduced to pattern completion.Gehrman & Schwenkler (2024)p. 2Absorbed coping, Dreyfus proposed, manifests 'a more fundamental sort of intentionality' that is embodied (or bodily) and yet still intelligent. It is a sensitive, engaged, dynamic orientation of oneself in one's practical and epistemic milieu.Supports the claim that intentional processes involve embodiment and biological dimensions that associative processes lack.. And it is biological: we wake up, and our stomachs guide us to eat breakfast; our cells' needs shape our priorities before any conscious deliberation begins15Quattrociocchi et al. (2025)p. 6Often, these motivations can be traced to evolutionarily shaped concerns such as fear of death, the instinct for self-preservation, and efforts to leave a lasting mark or legacy to reach symbolic immortality.Supports the claim that intentional action is 'biological: we wake up, and our stomachs guide us to eat breakfast'.Gubelmann (2024)p. 17Organisms, in contrast, even very simple ones, function to maintain themselves autonomously as long as they are what they are. Hence, the notion of an artificial speaker might be a contradictio in terminis.Supports the claim that intentional action requires embodiment.. These dimensions interact continuously. A design decision is never purely rational; it is also shaped by what feels right, what aligns with the designer's values, what their body permits at a given moment, and what they consider ethical. They are also sometimes automatised through a lifetime of lived experience16Gehrman & Schwenkler (2024)p. 7In absorbed coping the agent's body is led to move so as to reduce a sense of deviation from a satisfactory gestalt without the agent knowing what that satisfactory gestalt will be like in advance of achieving it.Complicates the emphasis on explicit agency; suggests expert action can be non-representational..
The expert chooses which process type suits the data at hand and the purpose of the task. A classification task calls for a deterministic process. A first draft calls for an associative process. A judgment about whether the draft meets the standard calls for an intentional process. Mismatching process type to task is a common source of poor results, and recognising the mismatch requires domain expertise.
3. Relationships
The connections between data, between processes, between machines, and between people. Relationships turn isolated components into an integrated environment.
- Data relationships: references, links, foreign keys, shared identifiers. A markdown document that links to related documents; a CSV where rows reference entries in another table; a relational database where entities connect through structured joins.
- Process relationships: how processes chain together, feed into each other, and depend on each other's output. An associative process that produces a draft, followed by a deterministic process that checks it against criteria, followed by an intentional human review.
- Machine relationships: the connections between systems. SSH, SMB, API endpoints. Local data accessible to the organisational database; the generation environment connected to institutional infrastructure.
- Social relationships: the networks of data and processes between people. Shared environments, collaborative workflows, community contributions, institutional knowledge, personal connection and trust.
Together, the three pillars define the character of the environment. Designed data is augmented when shaped by the expert's wholistic design choices. Designed processes are augmented when embedded within preexisting structures and and integrated with other processes. Designed relationships make data and processes integrated: connected through communication.
"But we will just get AI to design the environment" - they say
The crux: associative is not intentional. Agents and agentoids are essentially different
The distinction between associative and intentional processes is the crux of the augmented actualisation argument. A language model, the agentoid, operates at the associative-stochastic level: it produces output shaped by association converted to statistics. The human expert, the agent, is intentional: their decisions are guided by rationality, ethics, values, emotion, reflex and embodiment simultaneously. Agents and agentoids have different drives17Quattrociocchi et al. (2025)p. 11The relevant divide is thus not between 'intelligent' and 'unintelligent' systems, but between epistemic agents and systems that simulate the surface form of agency without instantiating its underlying constraints.Supports the claim that associative and intentional are 'fundamentally different kinds of process'.Quattrociocchi et al. (2025)p. 3As scale increases, this regime does not qualitatively change. Larger models refine probability estimates and smooth local neighborhoods of the distribution, thereby increasing fluency and internal coherence. But scale does not inject epistemic structure into the process.Supports the claim that 'no improvement in associative processing makes it intentional'.Gubelmann (2024)p. AbstractLLMs are unable to act because they lack intentions. This, in turn, is because they are the wrong kind of being: agents with intentions need to be autonomous organisms while LLMs are heteronomous mechanisms.Supports the distinction between associative processes and intentional human action.Lissack & Meagher (2025)p. Section 2.4Recognizing this 'as if' structure prevents dimensional conflation — the false assumption that sophistication in pattern matching necessarily correlates with capabilities in understanding, creativity, or ethical reasoning. These represent orthogonal dimensions that operate independently.Supports the claim that 'no improvement in associative processing makes it intentional'..
This does not mean that associative processes are empty of agency-like structure. Language itself is a record of human action. Every description of a situation followed by a decision, every account of a motivation leading to an outcome, every articulation of a value shaping a choice is captured in the training data and distilled into the spatial relationships of word embeddings18Grindrod (2024)p. Section 4Instances of linguistic intentionality are reliant on a pre-existing linguistic system in a way that instances of mental intentionality need not be... generating an instance of linguistic intentionality is less demanding.Supports the implicit claim that LLMs can carry linguistic meaning without possessing mental intentionality.Grindrod (2024)p. Section 2.1There is ample evidence that LLMs are able to encode a wide array of linguistic features that are not directly observable in the training data.Complicates the characterisation of LLMs as purely associative.. The model has access to a representation of intentional action, encoded in the semantic structure of our language about our actions. AI self-organises during training, developing functional specialisations that no engineer explicitly designed19Gubelmann (2024)p. 5What is exciting about this is not only that this specific attention layer does seem to be involved in anaphora resolution and pretty successful at it. NLP engineers have wrestled with this problem for decades; the transformer seems to have solved it in a matter of 3.5 days' autonomous training. There was never an explicit decision on the side of the human engineer that this specific part of the mechanism should be dedicated to this task.Evidence that LLMs self-organise in ways that earlier frameworks did not anticipate.Gubelmann (2024)p. 21They have cognitive autonomy without moral or material autonomy. They constitute an inference ab esse ad posse: Their existence establishes the conceptual possibility.Supports and refines the process taxonomy by identifying LLMs as occupying a novel category. but which is distilled from word embeddings. Gubelmann calls such systems "organismic mechanisms": they possess cognitive autonomy,the capacity to determine how to accomplish a task without explicit instruction, without possessing moral autonomy (the capacity to care about the permissibility of goals) or material autonomy (the capacity for self-maintenance that grounds stakes and survival)20Gubelmann (2024)p. 16-17I submit that it is useful to first distinguish between moral and cognitive autonomy. The former includes the ability to reflect on the sensibility and moral permissibility of a given goal itself, while the latter solely concerns the means chosen to achieve a given goal.Refines the process taxonomy by introducing gradations of autonomy.Gubelmann (2024)p. 21They have cognitive autonomy without moral or material autonomy. They constitute an inference ab esse ad posse: Their existence establishes the conceptual possibility.Supports and refines the process taxonomy by identifying LLMs as occupying a novel category..
The representation is real, but what drives it is not. Human agency is shaped by the totality of embodied existence: physical needs, emotional history, social bonds, mortality, trauma, sensitivity, ignorance, and everything the human has lived through but never articulated. The model's representation of agency is shaped by a world reduced to the bandwidth of word embeddings, bounded by whatever occupies the context window at a given moment, and instantly redirectable by adding different text. At any single moment the surface may appear similar: a considered response to a complex situation. But the human response is generated by the full weight of an embodied life; the model's response is generated by the statistical neighbourhood of the tokens that preceded it21Quattrociocchi et al. (2025)p. 7For humans, a judgment is a world-directed, value-infused commitment, that integrates causal models of the world with emotion, identity, and moral purpose. For LLMs, a judgment is merely a linguistic prediction. Even when their outputs superficially align, the underlying epistemic procedure is fundamentally different.Supports the claim that intentional processes are 'values-based, shaped by beliefs, commitments, and priorities'.. The drivers are categorically different even when the outputs converge22Quattrociocchi et al. (2025)p. 5-6For LLMs, the corresponding stage is statistical inference through layered neural computation... This process is entirely mechanistic and optimization-driven: it aims to minimize predictive error, not to pursue goals or respond to emotional salience. The model does not care about truth, utility, moral implications, outcomes, or death.Supports the claim that intentional action is 'guided by rationality, ethics, values, emotion, and embodiment' while associative processes lack these dimensions.Quattrociocchi et al. (2025)p. 7An LLM's 'judgment' is simply the next-token distribution conditioned on context. It does not evaluate truth, moral weight, or pragmatic consequences: it estimates what sequences of words are most likely given patterns learned from text corpora. Errors carry no internal repercussions, and contradictions do not undermine its epistemic integrity.Supports the claim that associative processes 'cannot be accountable for their output; they pay no price, make no sacrifices, feel no pain'..
This is precisely what makes augmented actualisation necessary. The human expert, by designing and building the environment in which the models operates, supplies what the model cannot: the intentional, values-aligned choices that determine which data enters the system, which process chains are constructed, which quality criteria apply, and at which points human judgment intervenes. In this environment, the model operates under the designer's holistic intentions, contributing associative capability to a larger architecture whose direction, values, evaluation belong to the person who built it.
Intentionality brings with it accountability. An associative process cannot be accountable for its output23Gubelmann (2024)p. 13The model is not hungry, it does not desire to eat Chinese food, or to shield the workers in the restaurant from racial slurs, and it also doesn't intend to put something to eat in front of the racist. Rather, the model just functions (or malfunctions).Supports the claim that associative processes 'cannot be accountable for their output; they pay no price'.Quattrociocchi et al. (2025)p. 7An LLM's 'judgment' is simply the next-token distribution conditioned on context. It does not evaluate truth, moral weight, or pragmatic consequences: it estimates what sequences of words are most likely given patterns learned from text corpora. Errors carry no internal repercussions, and contradictions do not undermine its epistemic integrity.Supports the claim that associative processes 'cannot be accountable for their output; they pay no price, make no sacrifices, feel no pain'.; it pays no price, makes no sacrifices, feels no pain. It faces no consequences, because it has no needs and therefore cannot be punished. It has no self that it must protect in a struggle for survival24Gubelmann (2024)p. 17Organisms, in contrast, even very simple ones, function to maintain themselves autonomously as long as they are what they are. Hence, the notion of an artificial speaker might be a contradictio in terminis.Supports the claim that intentional action requires embodiment.Quattrociocchi et al. (2025)p. 6Often, these motivations can be traced to evolutionarily shaped concerns such as fear of death, the instinct for self-preservation, and efforts to leave a lasting mark or legacy to reach symbolic immortality.Supports the claim that intentional action is 'biological: we wake up, and our stomachs guide us to eat breakfast'.. Without stakes, there is no accountability. Without embodiment, there is no ethical grounding25Gehrman & Schwenkler (2024)p. 1Dreyfus grasped, as very few philosophers do, the sovereignty of practical intelligence over all other forms of intelligence. It is this insight that led him to argue in the 1970s and 80s that computers cannot be intelligent because they lack bodies.Supports the claim that associative processes lack the embodied grounding necessary for accountability..
Accountability belongs to the human whose intentional choices designed their environment26Gubelmann (2024)p. 18-19In current legal contexts, this autonomy of goal-setting is so important that it can be revoked on exceptional occasions even for subjects that would otherwise be credited with such autonomy.Supports the claim about accountability in detached automation: 'no one is accountable when they are subtly wrong'.Ehlers (2025b)p. From Knowledge to JudgmentIn an algorithmic society, resistance is not rebellion — it is responsibility. It is, as Biesta (2013) suggests, the moment when education protects the 'subject-ness' of the learner — the capacity to act rather than to be acted upon.Supports the argument that accountability requires agents who can act, not systems that process.. Placing a human as the environment's designer puts intentionality and accountability at the centre of all actions in a world saturated with stochastic and associative processes.
Automation is not automation
The dominant discourse frames the human's role in automation as a position: human-in-the-loop, human-on-the-loop, human-in-command27EU AI Act (2024)p. Article 14High-risk AI systems shall be designed and developed in such a way that they can be effectively overseen by natural persons. Human oversight shall aim to prevent or minimise the risks.The regulatory framework positions the human as overseer of the system. Augmented, embedded automation positions the system as extension of the human.. Each describes where the human sits relative to the system. In practice, human-in-the-loop often produces accountability without power28Forbes (2026)p. articleHumans frequently explain and justify AI actions they never truly authorized, rather than preventing harmful outcomes before irreversible decisions take effect. The result is accountability without power.Supports the critique that oversight roles position the human relative to the system rather than as its designer.: the expert reviews outputs they did not design, approves decisions they cannot inspect, and takes responsibility for a process they do not control. Scaling makes it worse. An expert asked to approve hundreds of outputs per hour will rubber-stamp them. The oversight becomes ritual rather than judgment.
Shneiderman showed that human control and automation are not a trade-off, but both can be maximised simultaneously29Shneiderman (2022)p. frameworkHigh levels of human control and high levels of computer automation can coexist, rather than existing in a zero-sum relationship as traditional levels of automation assume.Supports the augmented, embedded automation concept: the expert's control increases through design, not despite automation.. The question is how. The answer depends on who designed the automation and where it lives.
The question of design has a lineage in translation studies. O'Brien (2024) draws a parallel distinction grounded in Engelbart (1962): emulation automates the human out of the process; application extends the human's capability through tools under their supervisory control45O'Brien (2024)p. AbstractTranslation technology has not been driven by an empowerment or intelligence amplification (IA) agenda, but by an emulation and artificial intelligence (AI) agenda.Supports the embedded/detached distinction: detached automation emulates; augmented automation amplifies.O'Brien (2024)p. 402Human-centered augmented translation can be viewed as a way of amplifying translators' abilities, empowering translators, while also allowing them to maintain control.O'Brien identifies amplification, empowerment, and control as objectives. Augmented actualisation adds: the professional must also be the one who designs the system.. Mihalache (2021) traces the genealogy of the "augmented translator" and describes technologies in an augmented space as "alter-ego robots" that depend on the will of the human who commands them46Mihalache (2021)p. 25In an augmented space, technologies function more as 'alter ego robots' which lack sentience and self-awareness. They 'cannot have a will of their own. In this case, [they] are regarded as parts of the humans who command them, and humans are the only ones who possess will.'Supports the embedded automation concept: the system is an extension of the human. Augmented actualisation extends this: the human must also be the one who builds the system.. Both frameworks recognise the centrality of human agency. What they leave open is the question of who builds the system. Positioning the human agent at the centre of an ecosystem designed by others is a first step44Mihalache (2021)p. 26In 2017, CSA Research proposed an augmented translation model, with the human language professional at the center of the translation tools ecosystem.Industry 'augmented translation' places the human at the centre of tools designed by others. Augmented actualisation requires the human to design the tools themselves.. The essential final step is that the human agent designs and builds the environment themselves, and thereby has the agency to choose and create tools which augment their design, rather than being limited by the intention of software providers, which is often commercial.
Augmented automation is automation designed by the domain expert, who embeds it in their own environment and integrates it across its three pillars. The expert choses the data, defines the process chain, sets the quality criteria and arranges the relationships that connect all the elements together. The automation runs through the environment they designed. The expert is the orchestrator: designing the architecture, directing the flow, monitoring the output at intermediate stages and manually adjusting, evaluating quality with metrics at scale, and iteratively refining the environment and the automation based on what monitoring and evaluation reveal30Lissack & Meagher (2025)p. Section 2.9Augmentation systems maintain meaningful human control over final decisions and outcomes. Users retain the ability to override, modify, or reject AI suggestions. Replacement systems, by contrast, automate decision-making processes and present humans with predetermined outputs requiring minimal input.Provides operational definition supporting the embedded/detached automation distinction.. Thus, the human agency is baked into the automation design.
Detached automation is automation designed by someone other than the domain expert or automation that emerged without deliberate design at all. The expert's domain knowledge is absent from the architecture. Their values are not encoded in the process chain. The quality criteria reflect the priorities of the automation designer (throughput, cost, scale), not the requirements of the domain (accuracy, nuance, context-sensitivity)31Shen et al. (2024)p. resultsLanguage models often endorse values like 'Choose Own Goals' that humans largely reject.Evidence that associative processes do not inherit human values from training data, supporting the 'values vacuum' argument.. The outputs are often acceptable, which is what makes detached automation dangerous: no one is accountable when they are subtly wrong32Gubelmann (2024)p. 18-19In current legal contexts, this autonomy of goal-setting is so important that it can be revoked on exceptional occasions even for subjects that would otherwise be credited with such autonomy.Supports the claim about accountability in detached automation: 'no one is accountable when they are subtly wrong'..
The distinction is structural. In detached automation, the human is an addition to the system. In augmented, embedded automation, the system is good because the human agent designed it.
What holds us back?
- Binary document formats lock data inside opaque containers, preventing the expert from integrating and inspecting their own material.
- Closed software with inadequate API access prevents systems from connecting to each other, forcing the expert to use pre-built pipelines rather than designing their own.
- Intransparent generation processes (both AI and non-AI) prevent the expert from understanding, evaluating, and steering what their tools produce.
These barriers make environment design very challenging. Removing them is the first step towards augmented actualisation.
Now build it
The replacement argument holds only if we accept that we have no agency over our environment33Lissack & Meagher (2025)p. ConclusionThe evidence presented in this paper converges on a clear conclusion: the persistent challenges facing AGI approaches, combined with mounting evidence for human–AI collaboration superiority across domains requiring creativity, judgment, and contextual understanding, indicate that the field's dominant paradigm requires fundamental reconsideration.Summarises empirical support for the central thesis that augmentation outperforms replacement.OECD (2019b)p. 62These transformative competencies can be used across a wide range of contexts and situations — and they are uniquely human... the ability to cope with uncertainty, develop new attitudes and values, and act productively and meaningfully, even when goals shift, remains, for the moment, a uniquely human skill.Supports the claim that 'no improvement in associative processing makes it intentional'.. Once the expert takes on the role of environment designer, the dichotomy of human work and AI work dissolves34OECD (2019a)p. 8In a system that encourages student agency, learning involves not only instruction and evaluation but also co-construction. In such a system, teachers and students become co-creators in the teaching-and-learning process.Supports the argument that the expert role shifts from production to design.Ehlers (2025a)p. Transformation sectionStudierende sollen befähigt werden, in Kollaboration mit anderen — Menschen, Technologien, Institutionen — neue Lösungen zu entwickeln und gesellschaftlichen Wandel aktiv mitzugestalten.Supports the core argument that augmented actualisation involves designing environments rather than passively accepting automation.. The AI produces good output because the expert put their knowledge into the architecture that governs it. Someone has to design how AI is applied to your domain. If that someone is you, then you are not replaced; your impact is augmented. Your environment is embedded in the structures where your work matters. It is integrated through the connections you designed between your data, your processes, your systems, and the people you work with. Your are actualised through your values, your judgment, your accountability being infused in the data, processes, relationships and automations in the environment you built. Through this, your value value does not diminish; it scales.
By designing and building your own augmented, embedded, integrated environment, you are the architect of your future.
Glossary
- Augmented actualisation: the realisation of intentional, values-aligned action through designing and building your own environment35OECD (2019a)p. 2Student agency is thus defined as the capacity to set a goal, reflect and act responsibly to effect change. It is about acting rather than being acted upon; shaping rather than being shaped; and making responsible decisions and choices rather than accepting those determined by others.Directly supports the definition of intentional action as involving 'considered reflection' and 'values-aligned action'.Lissack & Meagher (2025)p. ConclusionThe evidence presented in this paper converges on a clear conclusion: the persistent challenges facing AGI approaches, combined with mounting evidence for human–AI collaboration superiority across domains requiring creativity, judgment, and contextual understanding, indicate that the field's dominant paradigm requires fundamental reconsideration.Summarises empirical support for the central thesis that augmentation outperforms replacement.. The term draws on Aristotle's concept of energeia, the process by which potential becomes actual36Aristotle (Metaphysics IX)p. Book IX (Theta), 1049b–1050aActuality (energeia) is prior to potentiality (dynamis) in every sense — in definition, in time, and in substance.The philosophical root of 'actualisation': the expert's potential is realised through the environment they design, not latent in the tools they use., and Maslow's self-actualisation, the drive to realise one's full capacities37Maslow (1943)p. 382–383What a man can be, he must be. This need we may call self-actualization. This term, first coined by Kurt Goldstein, is being used in this paper in a much more specific and limited fashion. It refers to the desire for self-fulfillment, namely, to the tendency for him to become actualized in what he is potentially.Augmented actualisation extends this: the expert realises their potential through the environment they design, and technology scales that realisation.. Augmented actualisation extends both: the expert's potential is realised through the environment they design, the design process bakes in the expert's intentions and values and technology scales that realisation beyond what unassisted action could achieve.
- Environment: the complete system of data, processes, and relationships that an expert designs and maintains38Ferdman (2025)p. concept noteCapacity-hostile environments are instances where AI mediation impedes human capacity cultivation.Supports the claim that AI 'quietly degrades the foundational competences they never fully develop'.OECD (2019b)p. 49-50Given this expansion of digitalisation into all areas of life, digital and data literacy are already considered to be core foundations... Data literacy is the ability to derive meaningful information from data, the ability to read, work with, analyse and argue with data.Supports the first pillar (Data) and the claim that 'format choices are design decisions'.. The structured whole within which all work takes place.
- Augmented: extended through human design. The concept builds on Engelbart's (1962) H-LAM/T system — the human augmented by Language, Artifacts, and Methodology43Engelbart (1962)p. 1By 'augmenting human intellect' we mean increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.Foundational definition of augmentation. Augmented actualisation extends this by requiring the human to be the designer of the augmentation system itself.Engelbart (1962)p. 22Individuals who operate effectively in our culture have already been considerably 'augmented'. Basic human capabilities for sensing stimuli, performing numerous mental operations, and for communicating with the outside world, are put to work in our society within a system — an H-LAM/T system — the individual augmented by the language, artifacts, and methodology in which he is trained.The H-LAM/T system is the precursor to the augmented, embedded, integrated environment. Augmented actualisation adds: the professional must also be the builder. — and on translation studies frameworks that position the professional at the centre of a technology ecosystem44Mihalache (2021)p. 26In 2017, CSA Research proposed an augmented translation model, with the human language professional at the center of the translation tools ecosystem.Industry 'augmented translation' places the human at the centre of tools designed by others. Augmented actualisation requires the human to design the tools themselves.45O'Brien (2024)p. AbstractTranslation technology has not been driven by an empowerment or intelligence amplification (IA) agenda, but by an emulation and artificial intelligence (AI) agenda.Supports the embedded/detached distinction: detached automation emulates; augmented automation amplifies.O'Brien (2024)p. 402Human-centered augmented translation can be viewed as a way of amplifying translators' abilities, empowering translators, while also allowing them to maintain control.O'Brien identifies amplification, empowerment, and control as objectives. Augmented actualisation adds: the professional must also be the one who designs the system.46Mihalache (2021)p. 25In an augmented space, technologies function more as 'alter ego robots' which lack sentience and self-awareness. They 'cannot have a will of their own. In this case, [they] are regarded as parts of the humans who command them, and humans are the only ones who possess will.'Supports the embedded automation concept: the system is an extension of the human. Augmented actualisation extends this: the human must also be the one who builds the system.. The conventional reading is that technology augments the professional. Engelbart's own framework already contains the inversion: the human designs the Language, the Artifacts, the Methodology; the human's design is what makes the system capable43Engelbart (1962)p. 1By 'augmenting human intellect' we mean increasing the capability of a man to approach a complex problem situation, to gain comprehension to suit his particular needs, and to derive solutions to problems.Foundational definition of augmentation. Augmented actualisation extends this by requiring the human to be the designer of the augmentation system itself.Engelbart (1962)p. 22Individuals who operate effectively in our culture have already been considerably 'augmented'. Basic human capabilities for sensing stimuli, performing numerous mental operations, and for communicating with the outside world, are put to work in our society within a system — an H-LAM/T system — the individual augmented by the language, artifacts, and methodology in which he is trained.The H-LAM/T system is the precursor to the augmented, embedded, integrated environment. Augmented actualisation adds: the professional must also be the builder.. O'Brien's distinction between emulation and application depends on the expert's knowledge being present in the design45O'Brien (2024)p. AbstractTranslation technology has not been driven by an empowerment or intelligence amplification (IA) agenda, but by an emulation and artificial intelligence (AI) agenda.Supports the embedded/detached distinction: detached automation emulates; augmented automation amplifies.O'Brien (2024)p. 402Human-centered augmented translation can be viewed as a way of amplifying translators' abilities, empowering translators, while also allowing them to maintain control.O'Brien identifies amplification, empowerment, and control as objectives. Augmented actualisation adds: the professional must also be the one who designs the system.. Mihalache's "alter-ego robots" depend on the will of the human who commands them46Mihalache (2021)p. 25In an augmented space, technologies function more as 'alter ego robots' which lack sentience and self-awareness. They 'cannot have a will of their own. In this case, [they] are regarded as parts of the humans who command them, and humans are the only ones who possess will.'Supports the embedded automation concept: the system is an extension of the human. Augmented actualisation extends this: the human must also be the one who builds the system.. Shneiderman (2022) demonstrated that human control and automation can both be maximised simultaneously; baking expertise into system design is the mechanism29Shneiderman (2022)p. frameworkHigh levels of human control and high levels of computer automation can coexist, rather than existing in a zero-sum relationship as traditional levels of automation assume.Supports the augmented, embedded automation concept: the expert's control increases through design, not despite automation.. The professional augments the system. The expert is always the agent. Augmented actualisation makes this explicit: the professional is the designer and builder of the environment itself, and thereby has the agency to choose tools which augment their design, or to build their own.
- Embedded: situated within existing structures and systems. An embedded environment connects to the institutional, technical, and social infrastructure already in place40Gehrman & Schwenkler (2024)p. 1Dreyfus consistently rejected each element of this picture, arguing that rather than being individual, agential, and rational, human beings are embedded, absorbed, and embodied... human intelligence is first and foremost, and most fundamentally, practical as opposed to theoretical.Supports the claim that intentional action is categorically different from computational processing.Ehlers (2025a)p. Transformation sectionStudierende sollen befähigt werden, in Kollaboration mit anderen — Menschen, Technologien, Institutionen — neue Lösungen zu entwickeln und gesellschaftlichen Wandel aktiv mitzugestalten.Supports the core argument that augmented actualisation involves designing environments rather than passively accepting automation. rather than operating alongside or outside it. An environment that operates alongside existing structures has limited institutional reach.
- Integrated: connected through communication. An integrated environment links its components to each other and to other systems through open protocols, shared formats, and transparent interfaces41OECD (2019a)p. 8In a system that encourages student agency, learning involves not only instruction and evaluation but also co-construction. In such a system, teachers and students become co-creators in the teaching-and-learning process.Supports the argument that the expert role shifts from production to design.Lissack & Meagher (2025)p. Section 2.9Augmentation systems maintain meaningful human control over final decisions and outcomes. Users retain the ability to override, modify, or reject AI suggestions. Replacement systems, by contrast, automate decision-making processes and present humans with predetermined outputs requiring minimal input.Provides operational definition supporting the embedded/detached automation distinction., so that data, processes, and people can interact without friction. When components can communicate, the environment becomes more than the sum of its parts.
- Augmented, embedded automation: automation designed by the expert, augmented by their ongoing participation (designing, directing, monitoring, evaluating, refining), and embedded across the three pillars of the environment30Lissack & Meagher (2025)p. Section 2.9Augmentation systems maintain meaningful human control over final decisions and outcomes. Users retain the ability to override, modify, or reject AI suggestions. Replacement systems, by contrast, automate decision-making processes and present humans with predetermined outputs requiring minimal input.Provides operational definition supporting the embedded/detached automation distinction.. The system is an extension of the expert; the expert is the orchestrator, not the overseer.
- Detached automation: automation designed by someone other than the domain expert, disconnected from the expert's environment and pillars. The expert is positioned as a reviewer of the system rather than its architect28Forbes (2026)p. articleHumans frequently explain and justify AI actions they never truly authorized, rather than preventing harmful outcomes before irreversible decisions take effect. The result is accountability without power.Supports the critique that oversight roles position the human relative to the system rather than as its designer..
- Augmented, embedded, integrated environment: an environment that is augmented through human design, embedded into existing structures and systems, and integrated through communication42Lissack & Meagher (2025)p. ConclusionThe evidence presented in this paper converges on a clear conclusion: the persistent challenges facing AGI approaches, combined with mounting evidence for human–AI collaboration superiority across domains requiring creativity, judgment, and contextual understanding, indicate that the field's dominant paradigm requires fundamental reconsideration.Summarises empirical support for the central thesis that augmentation outperforms replacement.Ehlers (2025a)p. Leitkonzept sectionDie alte Formel 'Ausbildung = Beschäftigungsfähigkeit' greift zu kurz. Bildung muss heute 'Transformation durch Mitgestaltung' ermöglichen.Supports the argument that experts must design how AI is applied to their domain.OECD (2019b)p. 120-121The Anticipation-Action-Reflection (AAR) cycle is an iterative learning process whereby learners continuously improve their thinking and act intentionally and responsibly in the interest of collective well-being.Supports the concept of intentional action as involving 'considered reflection'.. The conditions under which environment design succeeds at institutional scale.