Generative AI and Cognitive Work

A conversation without a listener

In 1966, Joseph Weizenbaum published a conversation between a young woman and a computer at MIT.

She typed that her boyfriend had made her come there. The machine asked for an example. She said she was depressed. It replied that it was sorry to hear this. When she mentioned her mother, it asked about her family.

The exchange looked attentive. Its mechanism was available for anyone to inspect.

Weizenbaum's program, ELIZA, searched the user's sentence for keywords. A script called DOCTOR selected decomposition and reassembly rules. A reference to “mother” could trigger a question about family. A sentence beginning “I am” could return in another grammatical form. If no useful keyword appeared, the program retrieved an earlier topic or supplied a general response. The capitalised machine lines in the published transcript were assembled without a model of the woman's life.

Years later, Weizenbaum recalled an incident that unsettled him. His secretary had watched him work on the program and knew what it was. After a few exchanges, she asked him to leave the room. In Computer Power and Human Reason, he used that request to show how readily people supplied an appearance of understanding with meaning of their own.

ELIZA was not a modern generative model. Its rules were hand-built, its domain was narrow and its operation was comparatively transparent. The scene matters because it separates three things that fluent software invites us to join:

1. the form of an output;
2. the work a person does with that output; and
3. the understanding, authority and responsibility we attribute to its source.

Generative AI has changed the first of these beyond recognition. It can produce useful text, code, images, audio, plans and classifications across domains that no ELIZA script could cover. It can also change the second by shortening, rearranging or removing parts of a person's task.

The third does not follow automatically from either change.

That is the central problem of generative AI and cognitive work. A plausible artefact can arrive before anyone has established whether it is true, relevant, original, lawful, safe or fit for a particular decision. The work may have become faster. The claim has not yet become warranted.

The short answer

Generative AI changes cognitive work by generating or transforming representations: words, numbers, images, code and other symbolic material. In a suitable task and workflow, it can accelerate drafting, retrieval, translation, classification, explanation and software production. Its effect is uneven because performance depends on the precise task, model, context, worker, evidence available for checking and design of the surrounding organisation.

A generated output is a cognitive artefact, not a completed act of human judgement. It does not carry its own source chain, understand the consequence of using it or accept responsibility for the result. People and institutions still frame the purpose, supply context, test the output, authorise action, meet obligations and learn from what happens.

The practical question is therefore more exact than “Can AI do knowledge work?” Ask:

Which part of this cognitive task changed, what outcome improved, what new work appeared, and who can verify and carry the consequence?

This page uses cognitive work as a working term for activity whose immediate objects are representations. It includes finding, classifying, interpreting, transforming, composing, comparing, explaining and deciding with words, numbers, images, code or other symbols.

That is a task description, not a social rank. A nurse interpreting observations, an electrician diagnosing a fault, a carpenter reading a plan, a case worker assessing a file and a software engineer tracing a dependency all perform cognitive work. So does a student writing an essay. The term does not divide “thinking jobs” from everyone else.

From a script to a generator

ELIZA rearranged language through rules that Weizenbaum and his collaborators could enumerate. A modern language model learns statistical relationships across a large training corpus and generates a continuation conditioned on the material supplied to it.

The technical history contains many steps. One decisive step arrived in 2017, when eight researchers introduced the transformer. Its attention mechanisms let the model weigh relationships among positions in a sequence while permitting much more parallel training than the recurrent systems then dominant in machine translation. Later systems combined that architecture with large-scale pretraining and prompting. The GPT-3 paper showed one model performing varied language tasks from instructions and examples without updating its parameters for each task.

The result is a generality of generated form. A user can supply a question, a document, a codebase excerpt, a table or an image and ask for a new representation: a summary, a translation, an explanation, a program, a classification or a set of alternatives. Applications can connect the model to retrieval systems, calculators, software tools and organisational data.

Those additions matter. “The AI” in a workplace may include a base model, a search index, a prompt template, private data, business rules, an interface, permissions, monitoring and a person making a final decision. Capability belongs to this configured system under stated conditions, not to a brand name in the abstract.

The United States National Institute of Standards and Technology defines generative AI as models that emulate characteristics of input data to produce derived synthetic content. Its generative-AI risk profile also explains the stubborn factual problem. A language model's statistical generation can yield accurate and coherent text. It can also produce false, internally inconsistent or input-divergent material, sometimes with a confident explanation.

This is commonly called hallucination. NIST uses confabulation, partly to avoid borrowing a clinical term. Whatever word is chosen, the operational fact is the same: fluency and factual warrant are different properties.

The distinction is older than the transformer. Emily Bender and colleagues argued in “On the Dangers of Stochastic Parrots” that patterns of linguistic form in training data should not be confused with meaning grounded in a speaker, purpose and situation. The point is not that generated language can never be useful. It is that the usefulness has to be established in the human situation where the output will be used.

Cognitive work has more than one stage

It is easy to see a finished document and call the task “writing”, or see a recommendation and call the task “analysis”. The label hides the allocation of work.

A cognitive task can be opened into at least seven stages:

This is Guidebeam's working map, not a universal cognitive theory. The stages can overlap and loop. A discovery during evaluation may change the frame. An action may reveal missing evidence. A conversation can generate and evaluate in the same minute.

The map prevents a common mistake. If a model drafts a competent response, it has changed generation. It may also reduce the time needed to represent the problem or evaluate alternatives. It has not thereby chosen the organisation's purpose, verified every fact, acquired authority or accepted the consequence.

The map also prevents the opposite mistake. Work that remains with a person has not necessarily remained unchanged. The person may spend less time composing and more time checking. They may move from handling a whole range of cases to supervising standard outputs and resolving exceptions. Pace, attention and discretion can change even when the job title and headcount do not.

Generative AI · Cognitive workflow

Where did the cognitive work move?

A fluent artefact is one stage in a longer workflow. Framing, evidence, evaluation, decision authority, action and learning can remain elsewhere.

Lower consequence

Check fit, tone and obvious errors before use.

Material consequence

Verify sources, requirements, tests and affected people.

Rights or safety consequence

Require authorised review, traceable evidence and a route to challenge.

Cognitive work allocation across three generative-AI cases

Stage	Customer-support reply	Software change	Public-benefit case summary
1 · Frame	Person defines the customer problem; failure: the wrong issue is answered.	Developer defines the requirement and constraints; failure: a plausible but wrong task.	Case officer defines the decision question; failure: relevant rights or facts are omitted.
2 · Represent	Tool receives selected account context; verify consent, completeness and recency.	Tool receives code and tests; verify repository scope and hidden dependencies.	Tool receives authorised records; verify provenance, privacy and the governing rule.
3 · Generate	Model drafts a reply; fluent wording can contain a false policy claim.	Model proposes code; compilable output can still implement the wrong behaviour.	Model drafts a summary; confident prose can misstate evidence or law.
4 · Evaluate	Agent checks facts, policy and tone against the record.	Developer reviews reasoning and runs tests against requirements.	Officer checks every material statement against the authorised file.
5 · Decide	Authorised agent decides what promise or remedy can be offered.	Responsible maintainer decides whether the change may merge.	Legally authorised decision-maker determines the case; the model has no authority.
6 · Act	System sends only after approval; failure: an unauthorised commitment.	Pipeline deploys only through release controls; failure: an unreviewed change.	Agency acts through its lawful process; failure: a summary is treated as the decision.
7 · Learn	Outcomes and corrections improve guidance without hiding recurring errors.	Defects and review comments feed back into tests and practice.	Appeals and overturned cases reveal where the workflow needs repair.

A seven-stage table compares a customer-support reply, software change and public-benefit case summary. Every cell identifies the human and tool contribution, verification or authority boundary and a possible failure. A consequence scale shows that higher stakes require a heavier verification burden.

The writing experiment

In 2023, Shakked Noy and Whitney Zhang recruited 444 college-educated professionals for a preregistered experiment. Participants completed short writing tasks resembling work done by marketers, grant writers, consultants, managers and other professionals. Half were randomly given access to ChatGPT for the second task.

The effect was substantial within that setting. Completion time fell by 0.8 standard deviations and independent quality scores rose by 0.4 standard deviations. Lower-performing participants benefited more, compressing the measured productivity distribution.

The work changed inside the task too. Participants with the tool reported spending less time on rough drafting and more on generating ideas and editing. The tool did not simply make the same sequence faster. It moved effort among stages.

This is good evidence for a bounded claim: access to that system improved speed and assessed quality on those short, incentivised professional-writing tasks.

It is not evidence that all professional writing improved, that fact-checking became unnecessary or that an occupation could be removed. The tasks did not include months of stakeholder negotiation, access to confidential records, legal sign-off, responsibility for a false claim or observation of whether the resulting document worked in the world.

The difference is important because cognitive artefacts are judged through different tests. A persuasive internal note may be valuable because it helps a team see options. A regulatory filing must also be complete and accurate. A public explanation needs sources that readers can inspect. A condolence message can be grammatically flawless and still fail because the sender has outsourced the relationship it was meant to express.

One measured gain cannot carry all of those meanings.

The support system that carried expert practice

A later study moved from simulated writing into a live organisation.

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered introduction of a generative-AI assistant among 5,172 customer-support agents at a software company. The tool listened to text conversations and suggested responses based partly on patterns in earlier successful interactions.

In the peer-reviewed study, access increased issues resolved per hour by 15 per cent on average. Less experienced and lower-skilled agents gained more. They improved speed and quality, while the most experienced, highest-skilled agents saw small speed gains and small declines in quality.

The pattern suggests a form of knowledge redistribution. Practices that experienced agents had developed through many conversations became available as suggestions to colleagues who had not yet accumulated the same history. International agents also improved English fluency, and some gains persisted in later unaided work.

This was not knowledge appearing from nowhere. The system was trained on records of human service. Top performers had helped create the examples from which it learned. The study's authors note that their contribution to training data became another source of value, even as their own measured gains were small.

The result complicates a familiar story about skill. A tool can reduce a performance gap on a defined task. It can also change whose past work becomes infrastructure, who receives credit, which novel solutions enter the record and how compensation recognises contribution.

The authors keep the result within its boundary: one firm, one occupation, one assistant and a relatively stable set of support problems. A hospital ward, legal practice, school, engineering project and public agency have different records, consequences and routes for escalation.

The relevant lesson is not “novices always gain most”. It is that worker, task and accumulated organisational knowledge can change the effect.

The frontier has a broken edge

Fabrizio Dell'Acqua and colleagues designed an experiment with 758 Boston Consulting Group consultants. Participants were assigned to work without AI, with GPT-4, or with GPT-4 plus a short overview of prompting.

For 18 tasks selected to sit within the model's capabilities, AI users completed 12.2 per cent more tasks and worked 25.1 per cent faster. Their responses also received higher quality scores. The tasks included creative, analytical, writing and persuasion work familiar to consultants.

Then the researchers placed a case beyond the system's frontier. Solving it required joining spreadsheet data with a subtle fact in interview notes. The model's tempting reading of the numbers pointed towards the wrong recommendation.

Participants with AI were 19 percentage points less likely to reach the correct answer. They worked faster. Their wrong answers could still be coherent and persuasive.

The researchers called this a jagged technological frontier. Tasks that look similarly difficult to a person can fall on different sides of a system's capability. Prompting skill did not turn the frontier into a smooth line.

The phrase should not become a permanent map of jobs. The experiment tested one model on constructed cases at one time. Systems change. Tasks are decomposed and tools are connected. The frontier moves with the configuration and the standard of performance.

Its durable lesson is methodological. You cannot infer the sign of the effect from the fact of use. The same assistance that raises quality on one task can lower correctness on its neighbour. Speed can rise in both cases.

Sixteen developers and the missing speed-up

Software development provides another useful collision between expectation and measurement.

An earlier controlled experiment asked developers to complete a bounded JavaScript task. Participants with GitHub Copilot finished substantially faster. That result showed a capability in a short, greenfield setting.

In 2025, METR studied a different kind of work. Sixteen experienced open-source developers selected 246 genuine issues from large repositories to which they regularly contributed. Each issue was randomly assigned to permit or prohibit AI tools. The work went through the repository's normal review process.

The developers expected AI to help. Experts asked to forecast the result expected the same.

The measured result went the other way. With early-2025 tools available, the developers took 19 per cent longer. Time was spent prompting, waiting, reading and correcting suggestions that did not always fit the repository's implicit requirements. Even afterwards, participants believed the tools had made them faster.

This was a small and unusual sample: 16 experienced contributors, repositories they knew deeply, issues averaging about two hours and tools from a specific period. It does not prove that generative AI slows software development. The earlier coding experiment does not prove the reverse for mature systems.

Together, the studies show why “coding” is too large a unit. Producing a plausible function, understanding a codebase, preserving an interface, satisfying an undocumented convention, passing review and maintaining the result are different parts of the work. A generated artefact can shorten one and lengthen another.

Perceived productivity is different again. Saving the discomfort of a blank page can feel fast. Elapsed time, accepted output, defects, maintainability and learning each need their own observation.

Human plus AI is a hypothesis

The phrase human in the loop often arrives as reassurance. It names a position in a diagram, not a demonstrated safeguard.

Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone examined 106 experiments that reported the performance of humans alone, AI alone and the combination. Their preregistered meta-analysis covered 370 effect sizes from studies published between 2020 and June 2023.

On average, combined systems performed better than humans alone. They performed worse than whichever of the human or AI had done best alone. Decision tasks showed losses relative to the better component. Creation tasks had a more favourable pattern, though the estimate of synergy was imprecise.

The results were highly heterogeneous, and the review predates much of the current generative-AI evidence. They do not supply a rule that people and systems should work separately. They remove the assumption that combining them automatically captures the best of both.

A reviewer can miss an error because the output is fluent, because checking costs more than generating, or because the system has shifted their attention towards its proposed answer. A person can also reject useful assistance after false alarms or poor earlier experience. Raja Parasuraman and Victor Riley called these patterns misuse and disuse in their 1997 human-factors review.

Human oversight therefore needs a design:

• access to the underlying evidence, not only the generated answer;
• enough time to inspect it;
• skill relevant to the failure that matters;
• information about the system's state and limits;
• authority to reject, redo or escalate;
• feedback when an accepted output later fails.

Without these conditions, “a human checked it” can describe a ritual rather than a control.

Bainbridge's irony returns to the desk

In 1983, the psychologist and human-factors researcher Lisanne Bainbridge wrote a five-page paper about highly automated industrial plants.

Designers, she observed, tended to automate the tasks they could specify and leave people to monitor the system. The person was then expected to take over during a rare abnormal condition, precisely when the situation was unfamiliar and the routine practice needed to understand it had been removed. The more reliable the automation, the less often the operator exercised the remaining skill.

“Ironies of Automation” was written about process control. Its logic now appears in cognitive work.

A model handles routine correspondence, ordinary claims or common code changes. The person receives ambiguous cases, contradictions, angry customers and outputs that are almost right. Average production can improve while the remaining human work becomes less predictable. Verification replaces composition, yet the checker may have fewer opportunities to learn how the whole artefact is made.

That outcome is possible. It is not inevitable.

Work can be designed so that people retain meaningful practice, see the evidence behind suggestions, compare alternatives, inspect uncertainty and learn from outcomes. A system can expose expert reasoning rather than merely output an answer. Teams can rotate between generation and review. Organisations can preserve cases that teach, instead of sending every simple example into an invisible pipeline.

Current research does not justify a universal verdict about deskilling. A 2025 survey of 319 weekly workplace users found self-reported changes in where they applied critical thinking, especially towards verification and integration. Because the study relied on participants' accounts of 936 use examples, it cannot show that generative AI caused cognitive decline.

The honest conclusion is narrower. Learning depends on the activity that remains, the feedback it produces and the opportunity to understand more than the final answer. Removing effort can remove drudgery, practice or both. The effect must be observed over time.

AI, Entry-Level Work and Career Formation takes up the deeper question of how beginners acquire tacit knowledge, judgement and credibility when parts of junior work change.

Verification is part of production

Generative output is cheap enough to reverse an old relationship. Producing a first draft may take seconds; establishing whether it deserves to travel can take longer.

The NIST generative-AI profile recommends comparing output with known ground truth, documenting fact-checking, testing content flows and monitoring human–AI configurations. These are institutional controls, not claims that every output can be made reliable with a checklist.

The burden changes with the consequence.

A private brainstorm can tolerate false starts because the point is to widen the option set. A summary used to navigate a long document needs traceable passages. Code needs tests, review and an account of the environment in which it runs. A benefit decision, clinical note, legal filing or safety instruction requires the governing evidence and authority of that domain.

Verification can fail in several different ways:

• factual error: the output states something untrue;
• unfaithfulness: the output does not represent the supplied record;
• omission: a decisive fact or affected person is absent;
• misclassification: the right fact is placed in the wrong category;
• poor judgement: the facts are correct but the proposed action is unsuitable;
• illegitimate use: the action is outside the decision-maker's authority or violates an obligation;
• unowned consequence: no one is prepared to monitor or repair what follows.

These failures need different evidence. A citation can help with factual provenance and still leave the decision unjustified. A unit test can establish one behaviour and miss a security boundary. A second language model can reproduce the first model's misconception.

Verification is therefore part of production, not friction applied after the “real” work. If an organisation budgets for generation and leaves checking to spare attention, it has not measured the new task.

What the evidence supports

There is now credible evidence that generative AI can improve speed or assessed quality in particular cognitive tasks. The writing experiment, customer-support deployment and consultant study are not demonstrations alone. They observed people doing specified work against comparison groups.

There is equally credible evidence that the effect is uneven. It changed by worker experience in customer support, by task frontier in consulting and by the difference between a short coding exercise and experienced repository work. A broad 2024 meta-analysis found improvement over humans alone without average superiority over the better component.

The International Labour Organization's 2026 review reaches a compatible conclusion. Task-level gains are real but heterogeneous. Evidence of large-scale displacement remains limited, and reported time savings have not yet translated clearly into aggregate output, earnings or employment.

This is not a contradiction. It is the distance between a task result and a labour-market result described in How Technology Changes Tasks, Jobs and Occupations. Adoption, workflow redesign, demand, prices, new tasks, work relationships and the distribution of gains sit between them. The five terms in Automation, Augmentation, Exposure, Displacement and Transformation keep those stages attached to their evidence.

For cognitive work, the evidence supports five conclusions.

Generated form is a genuine capability. Text, code and other representations that once required direct human composition can now be produced or transformed under broad instructions.

The artefact is not the whole task. Framing, representation, evaluation, decision, action and learning may remain, move or expand.

The effect has a local sign. A system can improve one task, degrade its neighbour and impose verification costs that reverse an expected saving.

Skill changes the arrangement. Less experienced workers may gain access to patterns accumulated by experts. Experts may gain less, correct more, or spend time reconciling output with context the system lacks.

Responsibility survives generation. A model does not become accountable because its sentence is fluent. The person or institution that deploys, accepts and acts on the output still owns the evidential and human consequence.

The work after the answer

ELIZA's responses came from a script. Modern systems generate across domains with a reach that Weizenbaum did not live to see. Treating them as the same technology would be absurd.

Treating the old scene as irrelevant would miss its lesson.

The woman at the terminal performed the act that made ELIZA's sentence meaningful. She supplied a life, read attention into a transformation and decided that the exchange deserved privacy. The program produced form. The human furnished situation and consequence.

Today's generated artefact may be much more useful. It may contain a good explanation, a working program or an option no one in the room had considered. That makes evaluation more important, because useful systems earn entry into consequential work.

The future of cognitive work will not be decided by whether a model can produce an answer. It will be decided in the allocation that follows:

Who frames the question? Who supplies the record? Who notices what is absent? Who can reject the fluent mistake? Who signs, sends, deploys or pays? Who learns when the outcome returns?

Generation has become a new participant in that chain. It has not made the chain disappear.

Notes on the evidence

ELIZA is used as a historical case about output and attributed understanding, not as a technical analogue of a transformer. The working definitions of cognitive work, cognitive artefact and the seven-stage workflow are Guidebeam synthesis. Every productivity result is bounded to its study's task, system, sample and outcome. The Noy and Zhang tasks were short simulations; the Brynjolfsson, Li and Raymond result comes from one customer-support setting; the Dell'Acqua frontier was constructed around one model and set of cases; and the METR result comes from 16 experienced developers using early-2025 tools. The critical-thinking study is self-report evidence and does not establish cognitive decline. No study cited here proves universal productivity, deskilling, job loss or economy-wide employment change.

Sources and further reading

• Weizenbaum, Joseph. “ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine”. 1966.
• Weizenbaum, Joseph. Computer Power and Human Reason. 1976.
• Vaswani, Ashish et al. “Attention Is All You Need”. 2017.
• Brown, Tom B. et al. “Language Models are Few-Shot Learners”. 2020.
• Bender, Emily M. et al. “On the Dangers of Stochastic Parrots”. 2021.
• NIST. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. 2024.
• Lin, Stephanie, Jacob Hilton and Owain Evans. “TruthfulQA”. 2022.
• Bainbridge, Lisanne. “Ironies of Automation”. 1983.
• Parasuraman, Raja and Victor Riley. “Humans and Automation: Use, Misuse, Disuse, Abuse”. 1997.
• Vaccaro, Michelle, Abdullah Almaatouq and Thomas Malone. “When Combinations of Humans and AI Are Useful”. 2024.
• Noy, Shakked and Whitney Zhang. “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence”. 2023.
• Brynjolfsson, Erik, Danielle Li and Lindsey Raymond. “Generative AI at Work”. 2025.
• Dell'Acqua, Fabrizio et al. “Navigating the Jagged Technological Frontier”. 2026.
• Becker, Joel et al. “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”. 2025.
• Peng, Sida et al. “The Impact of AI on Developer Productivity”. 2023.
• OECD. The Impact of AI on the Workplace. 2023.
• Merola, Rossana et al. The Impact of GenAI on Jobs, Productivity and Work Organization. ILO, 2026.
• Eloundou, Tyna et al. “GPTs are GPTs”. 2023.
• Lee, Hao-Ping et al. “The Impact of Generative AI on Critical Thinking”. 2025.
• Autor, David. “Applying AI to Rebuild Middle Class Jobs”. 2024.