Faster, Busier, Worse: AI Doesn’t Fix Your Organisation. It Amplifies It.

Heraclitus said that character is destiny. Two and a half thousand years later, the research on AI at work suggests a corporate corollary: operating model is destiny. The machine does not decide what kind of organisation you are. It simply announces it, louder.

On 29 September, Harvard Business Review published a collection of its most insightful research on AI at work, arranged under five sober headings: productivity, trust, expertise, judgement and accountability. It is an excellent reading list. It is also, like most reading lists, a little too polite about what it contains.

Beneath the summaries sit the studies themselves: an eight-month ethnography of a technology firm, a pre-registered experiment with more than a thousand engineers, a randomised trial at a fintech, a 57-page theory of innovation, and a controlled test of what happens when you give software a job title. Read one at a time, each is a useful finding. Read together, they form an argument that most AI strategies have not yet priced in.

The argument is this. AI is not, primarily, a productivity technology. It is an amplification technology. It takes whatever an organisation already is, its incentives, its prejudices, its habits of judgement and its evasions of responsibility, and turns up the volume. The gap between AI activity and AI value is therefore not a technology gap. It is a design gap, and at present it is widening.

I. A Mirror With a Megaphone

We have spent three years asking whether AI is an oracle or a parrot. The evidence suggests a more uncomfortable third option: it is a mirror with a megaphone attached.

Where work norms are loose, AI lets work expand without limit. Where reputations are fragile, AI use becomes a stigma. Where expertise is thin, AI produces thinness at speed. Where innovation processes are biased, AI reproduces the bias at scale. Where accountability is vague, AI makes it vaguer, and does so politely, in well-formatted prose.

The labour-market data draws the same picture in a harder medium. Brynjolfsson, Chandar and Chen’s “Canaries in the Coal Mine”, built on US payroll microdata, found a 13% relative decline in employment for 22 to 25 year olds in the most AI-exposed occupations, after controlling for firm-level shocks. Their August 2026 update puts young workers in exposed jobs roughly 19% below where they would otherwise be.

But the telling detail is not the size of the decline. It is its location. The losses concentrate where AI is used to automate tasks. Where AI augments human work, early-career employment held steady or grew. Same technology, same period, opposite outcomes. The variable was not the model. It was the choice.

This is the philosophical core of what I have called the Long-AND, not the Short-OR. The question was never humans or machines. It is whether we have the patience to design systems in which each makes the other better, or whether we will settle for the quicker satisfaction of substitution.

II. Parkinson’s Revenge

In 1955, C. Northcote Parkinson observed, with the dry malice of a British civil servant, that work expands to fill the time available for its completion. Seventy years on, AI has produced a sharper version of his law: work expands to fill the capacity available for its completion. Give people a faster engine and they do not arrive earlier. They simply drive further.

The intensification effect

Aruna Ranganathan and Xingqi Maggie Ye spent eight months, April to December 2025, inside a US technology company of about 200 people: observing in person two days a week, reading internal channels, conducting more than 40 interviews (HBR, February 2026). AI use was voluntary. Nobody was told to work harder. Everybody did.

The researchers identified three mechanisms, each of which will be familiar to anyone who has watched a keen colleague discover a new tool:

1. Task expansion. People reached beyond their role because AI made unfamiliar work feel suddenly approachable.
2. Boundary erosion. Because AI lowers the friction of starting, work seeped into lunch, into meetings, into the evening. The cost of beginning fell to zero; the cost of stopping did not.
3. Parallel load. Employees ran several AI-assisted threads at once and exhumed long-buried tasks. Cognitive load was not reduced. It was redistributed across more tabs.

The second-order effect is the one that should worry executives. Engineers spent more time reviewing, correcting and guiding AI-assisted work produced by colleagues. One person’s acceleration became another person’s inbox. The authors prescribe an explicit “AI practice”: intentional pauses, deliberate sequencing, protected human contact. In other words, the very old technology of saying no.

A fair caveat: this is one firm and work in progress. But ethnography reveals mechanism, and mechanisms, unlike sample sizes, travel well.

Workslop: the tax nobody budgeted for

The review burden now has a name, and it is not a flattering one. BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 full-time US desk workers in September 2025 and defined “workslop” as AI-generated content that masquerades as good work while lacking the substance to move the task forward (HBR, September 2025; BetterUp summary).

Measure	Finding
Received workslop in the past month	40% of workers
Share of received work that qualifies	15.4% on average
Time to resolve each incident	Just under 2 hours
Estimated cost per employee	US$186 per month
Estimated cost for a 10,000-person firm	Over US$9 million a year
Managers vs individual contributors receiving it	54% vs 38.5%

The money is the least of it. Around half of recipients judged senders less capable and less reliable, and 42% trusted them less (study summary). Workslop does not merely waste time. It spends trust, which is the one organisational currency that cannot be printed.

Notice, too, where it pools: in managers’ inboxes. As HBR’s piece on managers struggling to keep up puts it, management was designed for a world in which execution took time, and AI has abolished that world. Execution now scales. Review does not. The manager has become the narrowest point in the hourglass.

My read: Goodhart in the telemetry

Goodhart’s law warns that when a measure becomes a target, it ceases to be a good measure. Much of the enterprise AI industry is currently engaged in a large, well-funded demonstration of the principle. We count prompts, seats, tokens and documents generated. We do not count review hours, rework, trust lost, or decisions made by exhausted people at 11pm. The dashboard glows green while the P&L remains serenely unmoved.

This is the Measurement Gap I described in Ghost in The Shell: The Agentic Reckoning, relocated from the economy to the office. It is also Ghost GDP in miniature. At the macro level, Ghost GDP is output that appears in the national accounts yet never circulates through households. Inside the firm, workslop is activity that appears in the telemetry yet never becomes value. Both are the same philosophical error: mistaking the production of things for the creation of worth.

III. Thamus Was Right, and Wrong

In Plato’s Phaedrus, Socrates tells the story of the god Theuth presenting the invention of writing to King Thamus. Thamus is unimpressed. Writing, he warns, will give people the appearance of wisdom without the substance. It is the first recorded competence penalty, and it has aged remarkably well. We have been suspicious of anyone who thinks with a tool for roughly 2,400 years.

The experiment

Oguz Acar, Phyliss Jia Gai, Yanping Tu and Jiayi Hou studied a leading technology company where, twelve months after rolling out an AI coding assistant, only 41% of engineers had adopted it (HBR, August 2025; full paper on SSRN). To understand why, they ran a pre-registered experiment with 1,026 of its engineers. Each evaluated an identical Python snippet. The only thing that varied was whether its author was said to have used AI.

When reviewers believed AI was involved, they rated the author’s competence 9% lower on average, for exactly the same code (HKU CAMO summary). The penalty fell harder on women and older workers; women suffered nearly twice the reputational damage of men. Measured quality, meanwhile, did not differ at all. Thamus, it seems, is alive and well and conducting code reviews.

Three consequences follow, each a governance problem dressed up as a change-management one:

• Rational non-adoption. Many engineers anticipated the penalty and declined to use AI to protect their standing. What leaders call resistance to change is, on closer inspection, a perfectly sensible reading of the room.
• Shadow AI. People who fear being seen with sanctioned tools do not abandon AI. They use unsanctioned ones, quietly relocating their work beyond every control you have built.
• Inequality amplification. A tool sold as a leveller compounds existing bias against the very groups already most scrutinised. The megaphone, again.

Where Thamus was right

Intellectual honesty requires the other half of the story. The workslop research shows that recipients do downgrade senders of low-quality AI output, and often with good reason. Some suspicion of AI users is learned, not merely prejudiced. The competence penalty is what happens when a reasonable inference from bad work is applied indiscriminately to good work as well.

The remedy, then, is not to forbid judgement. It is to judge the work, and to make quality visible independently of method. A disclosure policy without that safeguard is simply a register of people to penalise.

My read: the frozen are not lazy

In Ghost in The Shell I argued that the paralysis behind the Frozen Workforce is rarely located in the workers. This study is close to proof. The capability was there. Many of the 59% who held back were not frozen by fear of the technology but by an accurate reading of how their peers would judge them. The system did the freezing.

It is why I continue to insist that HX = CX + EX. How it feels to be seen using AI is not a soft, peripheral concern to be handed to internal communications. As I argued in The Long-AND, Not the Short-OR, employee experience is the load-bearing wall of any AI transformation. Remove it and the elegant architecture above comes down, however expensive the fittings.

IV. We Know More Than We Can Prompt

The philosopher Michael Polanyi observed in 1966 that “we can know more than we can tell.” Expertise, in his account, is mostly tacit: the surgeon’s feel for tissue, the editor’s ear for a false sentence, the underwriter’s unease about a file that is technically in order. Aristotle had a word for the practical form of this knowledge, phronesis, and he was clear that it could not be taught by instruction. It could only be grown through experience.

The most seductive promise in enterprise AI is that it can shortcut this growth, turning novices into experts on demand. The best evidence we have says it does something narrower, and philosophically more interesting.

The experiment

The study behind HBR’s “Gen AI Won’t Make Your Employees Experts” is Vendraminelli, DosSantos DiSorbo, Hildebrandt, McFowland, Karunakaran and Bojinov’s “The GenAI Wall Effect” (HBS Working Paper 26-011, full PDF). Researchers from Stanford and Harvard ran a randomised experiment with 78 employees at IG Group, a UK fintech, sorted by their distance from the craft of writing: insiders (web analysts who write for a living), adjacent outsiders (marketing specialists) and distant outsiders (technologists). Some were given generative AI; some were not. Each had to conceive and then write an article.

Ideas converged. Without AI, insiders led comfortably on conceptualising an article (3.82), ahead of marketers (3.04) and technologists (3.02). With AI, all three groups performed alike, and all outperformed insiders working unaided. At the level of ideas, AI is a great democratiser.

Craft did not. When it came to actually writing, AI-assisted marketers came close to the insiders. AI-assisted technologists hit a wall: their scores barely moved from the no-AI baseline. The mechanism is the revealing part. Technologists used AI to generate the content, then tidied its grammar; many simply pasted in what it offered. They lacked the tacit judgement to know what to keep, cut or reshape. Everyone saved time, the writing phase falling from about 87 minutes to 22 (HBS Working Knowledge), but for the distant outsiders speed simply delivered mediocrity sooner.

Reconciling the cybernetic teammate

This sits alongside the earlier Procter & Gamble field experiment with 776 professionals, “The Cybernetic Teammate” (Dell’Acqua, Lakhani, Mollick and colleagues), which found that AI helped individuals perform at the level of teams and softened the boundary between commercial and technical specialists. The two findings are not in tension. P&G’s task was ideation. IG’s wall appeared at execution. AI dissolves boundaries between disciplines at the level of ideas; it does not dissolve them at the level of craft. It is the “jagged frontier” Dell’Acqua and colleagues described in 2023, redrawn through the human rather than the task.

My read: we are sawing off the rungs

Now place this beside the Canaries data. AI is thinning out entry-level roles in exposed occupations, and entry-level roles are precisely where phronesis is grown: through repetition, correction and the occasional instructive humiliation. The GenAI Wall says tacit judgement is the entry ticket to AI productivity. The labour market says we are removing the rungs on which that judgement is earned. We are, in effect, asking the next generation to climb a ladder we are busily dismantling.

This is the Frozen Workforce risk in its most literal form, as I set out in The Long-AND, Not the Short-OR: not mass unemployment, but a leadership pipeline quietly starved of the repetitions it needs, years before anyone notices the gap. The organisations that prosper in the next decade will redesign junior roles around directing and critiquing AI. The ones that simply stop hiring juniors will discover, around 2035, that they have no seniors.

V. A Connoisseur of the Consensus

There is a quiet paradox at the heart of using generative AI for innovation. A large language model is trained on the sum of what has already been written. It is, by construction, a connoisseur of the consensus. Innovation, meanwhile, is a wager against the consensus. We are asking the most eloquent possible defender of the average to help us escape it.

Julian De Freitas (Harvard), Ayelet Israeli, Gideon Nave (Wharton), Artem Timoshenko (Northwestern) and Olivier Toubia (Columbia) set out the consequences in HBR as “The Innovation Problems AI Can’t Solve”. The full 57-page working paper, “Innovating with Generative AI: A Human Bottleneck Framework” (also on SSRN), is considerably richer than the article, and repays the effort.

The framework

Their premise is that most innovation constraints are human, not technological: the psychological and behavioural mechanisms that limit how we imagine, notice, articulate, evaluate and act on ideas. The paper maps eleven such bottlenecks across the pipeline, from ideation through screening and consumer insight to market learning (HBS AI Institute summary).

The central insight is that AI does not act uniformly on these constraints. At each stage it can relieve some bottlenecks while quietly deepening others, and predicting which requires understanding the mechanism behind the constraint rather than the capability of the model. Three examples illustrate the pattern:

• Cognitive fixation at ideation. Teams anchor on familiar ideas. A model trained on existing output reproduces that anchoring faster and at scale: a thousand variations on the obvious, delivered in seconds.
• Attention and aggregation in feedback. Products now generate more customer feedback than any team can read. AI genuinely helps, but it solves volume, not priority. Deciding what matters remains stubbornly human.
• Consumer insight. Synthetic respondents and digital twins are fast and cheap. They are also, by design, unsurprising, and surprise is precisely what breakthrough products are made of.

The authors separate bottlenecks likely to narrow as models improve from those rooted in enduring human nature. Their advice is to build strategy on the mechanism rather than the model. The model will be obsolete within a year. The psychology has been stable since the Pleistocene.

My read: the return is the learning

This is, for my money, the most important paper in the collection for anyone running a transformation programme, because it dismantles the comfortable assumption that faster means better. A faster biased process is merely a biased process that fails sooner and at greater scale.

It also sharpens an argument I made in The Return on Not Yet Knowing: in early-stage AI work, learning is the return. When AI removes the friction of contact with real customers and real data, it removes the learning along with the friction. The design question for every innovation workflow is therefore epistemological before it is technical. Which steps must keep an unmediated line to reality? Those are exactly the steps you must not hand to the machine.

VI. A Computer Can Never Be Held Accountable

A page from a 1979 IBM training manual has become one of the most shared images in AI governance. Its first line reads: “A computer can never be held accountable.” Its conclusion is that a computer must therefore never make a management decision. Nearly half a century later, the research suggests we have kept the first half of that insight and quietly mislaid the second.

When software gets a job title

Wittgenstein wrote that the limits of my language are the limits of my world. A randomised experiment reported in HBR, “Why You Shouldn’t Treat AI Agents Like Employees” (May 2026), turns that aphorism into a management finding. When the same system was framed as an “AI employee” rather than an “AI tool”:

• Personal accountability fell by 9 percentage points, while accountability attributed to the AI rose by 8.
• Errors were more likely to slip past managers.
• Unnecessary escalation rose and review quality fell.
• Employees grew less certain of their own roles and professional identity.
• And the benefit that humanising AI is meant to buy, greater willingness to adopt, did not materialise.

The industry’s vocabulary is drifting precisely the wrong way: digital workers, AI colleagues, agentic headcount. These are not harmless metaphors. Language is load-bearing. Call a system a colleague and people begin, measurably, to treat it as one, including by handing it the blame.

The courts have not been consulted on the metaphor

HBR’s July piece, “You Outsourced the AI, but You Still Own the Risk”, follows the same logic into the courtroom. Cases involving Peloton, iTutorGroup, Workday, Cigna and others show liability landing on the organisation closest to the affected person, not the model provider. Law firms, not AI vendors, have been sanctioned for filings containing hallucinated citations. Technical control is distributed along the supply chain. Accountability is not. Standard vendor contracts, meanwhile, frequently make matters worse.

The HBR collection states the organisational point plainly: human and AI partnerships create ethical and legal exposure precisely because it becomes unclear who is responsible when things go wrong (HBR collection). Ambiguity, it turns out, is not neutral. It is a liability with a delayed invoice.

My read: identities, permissions, audit trails, and no alibis

This is why I built TrustOS as a seven-layer architecture rather than a policy document, and why, in co-authoring the Agentic AI governance framework with IMDA, named human accountability was never negotiable. A principle declaring that “humans remain accountable” is decorative if the operating model quietly routes responsibility to an agent with a friendly name and a cheerful avatar.

It is also the nub of the enforceability test I set out in The Fine Print. The question for any governance claim is not whether a framework exists, but whether it could be enforced against a specific person, on a specific day, for a specific failure. The anthropomorphism study shows how easily the answer becomes no: not through malice, but through vocabulary.

My rule for clients is short enough to fit on that 1979 slide. Agents get identities, permissions and audit trails. They do not get accountability. That belongs, always, to a named human with the authority to stop the loop.

VII. The Panel Weighs In

A thesis that has not been argued with is merely an opinion wearing a suit. So I put the amplifier argument before my usual panel, using only what each has said on the record, and let them take their swings.

Mo Gawdat: the amplifier, at civilisational scale

Gawdat will recognise the thesis, because in a sense it is his. He has argued that AI is here to magnify everything humanity already is (Info-Tech, Digital Disruption). On Diary of a CEO he went further, forecasting a 12 to 15 year “short-term dystopia” of job loss and inequality before any abundance arrives (coverage).

His objection: I locate the problem in operating models; he locates it in values, specifically in the greed and ego of those in charge. My reply: he is right about the root, but operating models are where values stop being sentiments and become mechanisms. An AI practice with protected pauses is a value made enforceable. “Short-term turbulence for long-term abundance” is my more hopeful reading of the same curve. The distance between his dystopia and my turbulence is exactly the design work this essay describes.

Dario Amodei: the macro shock may swamp the micro fix

Amodei has been the most candid insider in the industry. In May 2025 he told Axios that AI could eliminate half of all entry-level white-collar jobs and lift US unemployment to 10 to 20% within one to five years, adding that those building the technology owe the public honesty rather than sugar-coating (Axios).

His objection: firm-level design is a rounding error against a labour-market shock of that magnitude. My reply: the Canaries data partly vindicates him; the entry-level decline is real and growing. But it also shows the damage concentrating where AI automates rather than augments, which means firm choices still bend the curve. He is right that enterprise design alone will not suffice. That is an argument for adding policy, not for abandoning design.

Erik Brynjolfsson: augment, do not imitate

Brynjolfsson’s evidence is the spine of Section I, and it is the empirical sequel to his 2022 essay “The Turing Trap”, which argued that pursuing human-imitating automation at the expense of human-augmenting tools is an economic and social mistake (Canaries paper).

His refinement: the choice between automating and augmenting is not made on a level field. Incentives, including the tax treatment of capital versus labour, tilt firms towards substitution. “Design better” must therefore include redesigning what gets rewarded, inside the firm and in public policy.

Andrew Ng: the optimist who proves the point

Ng is the panel’s cheerful realist. He has argued that if 20 to 30% of a job is automated, the job survives, and that AI will not replace people but people who use AI may replace those who do not. He adds that when a task becomes dramatically cheaper, firms tend to do it vastly more often rather than bank the saving (Business Insider via Yahoo).

Where he sharpens the thesis, perhaps unintentionally: that last observation is Jevons’s paradox, and it is the intensification effect seen from the boardroom. Doing a thing ten thousand times more is growth when judgement scales with volume, and workslop when it does not. Ng’s own advocacy of agentic workflows built on reflection, tool use, planning and multi-agent collaboration (Insight Partners) is, read closely, an argument for engineering review into the loop rather than exporting it to the next unfortunate human.

Yoshua Bengio: humans alone cannot watch the watchers

Bengio presses hardest on Section VI. He founded LawZero in June 2025 out of concern that agentic systems are already exhibiting deception and self-preservation, and proposed a non-agentic “Scientist AI” as a guardrail (Bengio, Introducing LawZero). He has argued that any effective safeguard must be at least as capable as the agent it oversees (Pure AI).

His objection: “a named human who can stop the loop” is necessary but nowhere near sufficient. A tired manager drowning in workslop is not a control; she is a hope. Lisanne Bainbridge made the point in her classic 1983 paper “Ironies of Automation”: the more we automate, the more we depend on human operators whose skills the automation has allowed to wither. My reply: I accept it entirely. Accountability must remain human, but oversight must be instrumented: machine monitors that surface the anomaly, humans who own the decision. That division of labour is what TrustOS is built to enforce.

Ethan Mollick: the frontier runs through people

Mollick and his Harvard co-authors supply the frame that reconciles the expertise findings. Their 2023 “jagged frontier” study showed AI lifting performance on tasks inside its capability boundary while degrading it, often invisibly, on tasks just outside (SSRN). The refinement: the GenAI Wall is the jagged frontier drawn through a person rather than a task. Where your expertise ends is where AI stops helping and starts, very persuasively, misleading.

The verdict

The amplifier thesis survives cross-examination, but leaves the room sharper. Design must include incentives, not merely workflows (Brynjolfsson). Firm-level design needs a policy complement (Amodei, Gawdat). And human accountability needs machine-assisted oversight if it is to be more than a comforting sentence (Bengio).

VIII. Prompts, Loops and Loop Governance

Diagnosis is the easy part; every consultant has a stethoscope. The harder question is what to build instead. My answer follows the progression I use with clients: Prompts → Loops → Loop Governance.

In the prompt era, AI was a personal tool, and every pathology in this essay stayed hidden inside individual behaviour: one person’s longer evenings, one person’s pasted draft. In the loop era, AI runs multi-step workflows, and agents hand work to other agents and to people. Now the pathologies compound along the chain. Intensification at one node becomes review debt at the next. Vague accountability becomes nobody’s accountability. Loop governance is the discipline of designing those chains so that value, judgement and responsibility are engineered in rather than hoped for.

The design map

Research finding	Failure mode in the loop	Design response	What to measure
Work intensifies (Ranganathan and Ye)	Capacity absorbed by expansion, not value	An explicit AI practice: work-in-progress limits, sequencing, protected pauses	Hours outside core time; WIP per person; decision quality
Workslop (BetterUp Labs and Stanford)	Review cost exported downstream	Sender-owned quality gates; context required before hand-off	Rework hours; received-work quality scores
Competence penalty (Acar et al.)	Rational non-adoption and shadow AI	Judge outputs blind to method; peer role models; safe disclosure	Adoption by demographic; shadow AI incidents
GenAI Wall (Vendraminelli et al.)	Outsiders ship fluent but flawed work	Match AI autonomy to user expertise; rebuild junior roles around critique	Quality by expertise distance; junior progression
Human bottlenecks (De Freitas et al.)	Faster, more biased innovation	Keep unmediated customer contact at designated stages	Novelty of ideas; share of real vs synthetic insight
Anthropomorphised agents (HBR, May 2026)	Accountability drifts to the machine	Tool framing; named human owner per loop; kill-switch authority	Undetected error rate; escalation rate

Seven moves for the next 90 days

1. Measure worth, not motion. Retire adoption dashboards as measures of success. Replace them with rework, cycle time to a decision, and quality scored blind.
2. Write your AI practice down. Ranganathan and Ye’s norms cost almost nothing: sequencing rules, protected focus time, explicit expectations about the evening.
3. Make the sender own the slop. Nothing AI-assisted crosses a hand-off without the context and verification its recipient needs.
4. Retire Thamus. Evaluate outputs without method labels wherever you can, and recruit respected peers, not merely executives, as visible users.
5. Calibrate autonomy to expertise. Give experts wide latitude. Give distant outsiders AI for ideation and a human expert for execution.
6. Protect the apprenticeship. Redesign graduate roles so that juniors direct, critique and correct AI output, growing the phronesis the GenAI Wall shows they will need.
7. Name the owner of every loop. Every agentic workflow gets a named human with authority to halt it, instrumented monitoring beneath them, and an audit trail. Agents receive identities and permissions; never accountability.

None of these moves requires a better model. Every one of them requires a braver operating model. That is the uncomfortable conclusion of a year’s research: the binding constraint on AI value in 2026 is no longer capability. It is design, and design, unlike compute, cannot be bought by the rack.

IX. Coda: The Fork Runs Through the Org Chart

I often describe our moment as The Fork: one road towards a Star Trek future of shared abundance, the other towards Mad Max. We tend to discuss that fork at the altitude of nations and economies, where it feels suitably grand and conveniently distant. This research brings it down to earth. The fork runs through every organisation, every team and every workflow, and it is chosen on ordinary Tuesdays.

The same tool, in the same quarter, can drive one team into burnout and free another for its finest work. It can penalise one engineer and elevate another for identical code. It can lift a marketer to near-expert prose and leave a technologist pasting paragraphs he cannot judge. It can accelerate an innovation pipeline, or merely accelerate its prejudices. It can sharpen accountability, or dissolve it into a friendly job title.

Heraclitus also said that no one steps in the same river twice. Every organisation adopting AI is stepping into a new river, and becoming a different organisation as it does. The only open question is whether that difference is designed or merely suffered.

The technology will not choose the road. We will, one operating decision at a time. That is what I mean by short-term turbulence for long-term abundance. The turbulence is not optional. The abundance is.

Luke Soon is an AI futurist, ethicist and author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: SuperIntelligence Protocol. He writes at GenesisHumanExperience.com.

References

Classical and foundational sources

• Plato. Phaedrus (the myth of Theuth and Thamus).
• Parkinson, C. N. (1955, 19 November). Parkinson’s Law. The Economist.
• Polanyi, M. (1966). The Tacit Dimension. University of Chicago Press.
• Wittgenstein, L. (1921). Tractatus Logico-Philosophicus, proposition 5.6.
• Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6).
• Brynjolfsson, E. (2022). The Turing Trap: The Promise and Peril of Human-Like Artificial Intelligence. Daedalus, 151(2).

The HBR collection

• Masinter, A. W. (2026, 29 September). A Collection of HBR’s Most Insightful Research on AI at Work. Harvard Business Review.

Productivity and intensification

• Ranganathan, A., and Ye, X. M. (2026, 9 February). AI Doesn’t Reduce Work, It Intensifies It. Harvard Business Review.
• Niederhoffer, K., Rosen Kellerman, G., Lee, A., Liebscher, A., Rapuano, K., and Hancock, J. T. (2025, 22 September). AI-Generated “Workslop” Is Destroying Productivity. Harvard Business Review. See also the BetterUp Labs summary.
• Managers Are Struggling to Keep Up with the AI Productivity Boom (2026, May). Harvard Business Review.

Adoption and bias

• Acar, O. A., Gai, P. J., Tu, Y., and Hou, J. (2025, August). Research: The Hidden Penalty of Using AI at Work. Harvard Business Review. Working paper on SSRN.

Expertise

• Vendraminelli, L., DosSantos DiSorbo, M., Hildebrandt, A., McFowland III, E., Karunakaran, A., and Bojinov, I. (2025). The GenAI Wall Effect: Examining the Limits to Horizontal Expertise Transfer Between Occupational Insiders and Outsiders. HBS Working Paper 26-011 (PDF).
• Gen AI Won’t Make Your Employees Experts (2026, March). Harvard Business Review.
• HBS Working Knowledge. Gen AI Boosts Productivity, But Can’t Turn Novices Into Experts.
• Dell’Acqua, F., Ayoubi, C., Lifshitz, H., Sadun, R., Mollick, E., Mollick, L., et al. (2025). The Cybernetic Teammate. HBS Working Paper 25-043.
• Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., et al. (2023). Navigating the Jagged Technological Frontier. SSRN.

Innovation

• De Freitas, J., Israeli, A., Nave, G., Timoshenko, A., and Toubia, O. (2026, 14 August). Research: The Innovation Problems AI Can’t Solve. Harvard Business Review.
• De Freitas, J., et al. (2026). Innovating with Generative AI: A Human Bottleneck Framework. arXiv 2608.07504; also SSRN.
• HBS AI Institute. Idea Flow + The Machine.

Accountability

• Research: Why You Shouldn’t Treat AI Agents Like Employees (2026, May). Harvard Business Review.
• You Outsourced the AI, but You Still Own the Risk (2026, 9 July). Harvard Business Review.

Labour markets

• Brynjolfsson, E., Chandar, B., and Chen, R. (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab; August 2026 update.

The panel

• Axios (2025, 28 May). Behind the Curtain: A white-collar bloodbath (Dario Amodei).
• Info-Tech Research Group. Ex-Google officer on AI, capitalism and the future of humanity (Mo Gawdat); and ChannelNews coverage of his Diary of a CEO interview.
• Business Insider via Yahoo. AI won’t replace human workers, Andrew Ng says; Insight Partners, Why Agentic AI is the smart bet for most enterprises.
• Bengio, Y. (2025, June). Introducing LawZero; Pure AI coverage.

From this blog

• Ghost in The Shell: The Agentic Reckoning (7 July 2026)
• The Return on Not Yet Knowing (3 September 2026)
• The Long-AND, Not the Short-OR (21 September 2026)
• The 2026 Enforceability Test for AI Governance: The Fine Print (29 September 2026)