# Contenu

> Contenu is the practice of Robert J. van Raamsdonk (Almere, The Netherlands): AI, automation and transformation consulting for global enterprises, and focused software products built with AI-assisted development.

A linguist by training (Japanese and French, Leiden University and Duke University) who began in software translation and localisation, Robert works at the interface of business and technology: clear multilingual, cross-cultural communication combined with strong hands-on technology skills. He has 30+ years in enterprise technology and has worked through every major wave: the early web, multilingual content management, Documentum, SharePoint, the Power Platform and generative/agentic AI. Since 2025 he is Solution Architect & Product Owner, AI & Automation in EY Client Technology (EY's four global service lines); 2024-2025 Product Owner, AI & Automation for EY GCR Finance Managed Services. Founder of Contenu Global ECM Consulting BV (2009-2024), co-founder of Informed Consulting (2007-2009); earlier LogicaCMG and Lionbridge (Waltham, MA). 15+ years transforming EY (Tax, Law, Finance, Client Technology) in Amsterdam, Rotterdam, London, Belfast and Washington, DC, including a global document collaboration rollout to 100,000+ users in 90+ countries. Clients include EY, IBM, Nestlé, Pfizer, Heineken, Philips, Dassault Systèmes, Biogen, ING, Robeco, Springer and Legrand. Hands-on AI builder with Claude Code, Cursor, Codex, Replit, GitHub Copilot, Factory.ai Droid, LangGraph, LangChain and the Microsoft cloud. On-site work in 41 cities across North and South America, Europe, Africa, the Middle East, Asia and Australia. Languages: Dutch, English, French, German.

Contact: robert@contenu.nl · https://www.linkedin.com/in/rvanraamsdonk

## Pages

- [Home](https://www.contenu.nl/): overview, approach (frame, build, test, shape, land, embed: through rollout, communication and adoption) and about
- [Consulting](https://www.contenu.nl/consulting/): services (AI & automation direction, rapid prototyping, global product & transformation delivery), selected work, career timeline
- [Software](https://www.contenu.nl/software/): LyricsRadar and bespoke prototypes
- [Writing](https://www.contenu.nl/writing/): essays
- [FAQ](https://www.contenu.nl/faq/): short answers about the consulting work, LyricsRadar, the essays and how they are written

## Writing

On writing with AI: Robert writes openly with AI. The thinking, experience and judgement are his; Claude helps test arguments, check sources and edit. Authorship and accountability stay with him.

- [The Quiet Retreat](https://www.contenu.nl/writing/the-quiet-retreat/): Enterprise AI is stalling on people: rationed access, quiet resistance and the stigma on anything a machine has touched. Argues that rationing (compliance-driven capability stripping, flat cost caps), employee resistance and the stigma on AI-assisted work, made detectable by EU AI Act watermarking, suppress the broad everyday use AI value depends on; proposes explicit accountability for AI processes, disclosure norms that separate assistance from authorship, and governed rather than withdrawn access. Cites ECB, IW Köln, Bitkom, PNAS, Sveriges Radio and the EU AI Act.
- [Even If It Stopped Today](https://www.contenu.nl/writing/even-if-it-stopped-today/): Even if AI stopped improving today, it would still transform knowledge work. Applies a "freeze test" to current models, argues that hype and doom both trap organisations in the future tense, and describes the "capability overhang" between what models can do and what organisations let people do. Cites Eurostat, Humlum and Vestergaard (Denmark), Dell'Acqua et al. (BCG/Harvard) and Paul David's dynamo analogy.
- [Embrace the Mess](https://www.contenu.nl/writing/embrace-the-mess/): Large organisations stall on AI because they insist on order first. Argues that the cleanup fallacy, platform-first thinking and blanket human-in-the-loop review are one control reflex that keeps AI pilots from scaling; cites MIT NANDA, Gartner, BCG, Bainbridge and the EU AI Act, and proposes tactical wins first, proportionate oversight and letting structure emerge from use.
- [The Last Interface Is Language](https://www.contenu.nl/writing/the-last-interface-is-language/): AI is weakening our ability to read, write and think at length, just as those skills become the key to using it. Argues that reading and writing are now the core skills for directing and supervising AI, cites NAEP, PISA 2025, Anthropic, MIT, BCG and Bainbridge, and proposes "protect the struggle first, then add the tool".

## More

- [Full text of all essays](https://www.contenu.nl/llms-full.txt)
- [RSS feed of essays](https://www.contenu.nl/writing/feed.xml)

## Products

- [LyricsRadar](https://www.contenu.nl/software/#lyricsradar): a VST3 lyric and songwriting plugin for Cubase (Windows, in development). Keeps lyrics in the Cubase project and shows them as a teleprompter that follows the playhead: karaoke word fill, full-screen second-monitor prompter, and a phone prompter over Wi-Fi. Timing is marked by listening, tapping or typing.

# Full text of essays

---

# The Quiet Retreat

By Robert J. van Raamsdonk · Published 2026-11-03 · https://www.contenu.nl/writing/the-quiet-retreat/

> Why enterprise AI adoption is stalling on people: rationed access, quiet resistance, and the stigma attached to anything a machine has touched.

Organisations are rationing AI for the people who do the work while betting their headcount on it.

On Monday 26 October, staff at Sweden’s Government Offices opened their browsers and found that ChatGPT and Claude no longer loaded. From now on, they were told, it would be Microsoft Copilot: the one tool the ministry had vetted. Nobody had done anything wrong. The decision was careful, defensible and entirely typical of what is now happening in large organisations across Europe and beyond.

Two years ago every large organisation wanted its people using AI. Now many of them are quietly taking it away. The tools have got better, and also more expensive and more visible. AI inside the enterprise is meeting three forces that have nothing to do with model quality: rationing from above, resistance from below, and a stigma on work that a machine has touched. Each is understandable, but together they suppress what AI’s value depends on: broad, open, everyday use.

## The rationing

The Swedish block is the visible version of a quieter process. In large companies, compliance and governance rarely ban AI outright. They strip it: features switched off, web access disabled, the stronger models removed, the choice narrowed to whichever vendor procurement already trusts. Step by step, that often leaves a tool that no longer does what people needed.

I have seen this from the inside, and I think it is among the most damaging things an organisation can do to its own AI programme. People try the stripped-down tool, conclude that “AI isn’t that useful”, and stop. The organisation then reads the low usage as proof that the technology was overhyped.

Cost is the second lever. In 2026, AI vendors moved from flat-rate licences towards paying per use, and budgets that had been experiments became line items finance noticed. In June, Walmart, which had given staff unlimited use of its in-house AI agent, put every employee on a fixed allowance. Controlling spend is reasonable. But a flat cap measures only cost. It hits the employee who has rebuilt a whole workflow around AI exactly as hard as the one generating slides nobody reads, and in practice it lands first on the heaviest users, who are often the ones getting the most out of it.

Rationing also drives use out of sight. Germany’s Institut der deutschen Wirtschaft found that 29% of employees who use AI at work use tools their employer didn’t provide. The figure rose to 45% where employers actively encouraged AI skills. Organisations are teaching people to use AI and then not handing them the tools, so people bring their own. The organisation saves on licences and gains a shadow AI problem it can’t see, govern or learn from.

## The resistance

Use is still rising. A European Central Bank survey of some 20,000 people found that the share of euro area workers using AI on the job doubled in two years, to 52% in 2026. Enthusiasm is not keeping pace. The share who view AI positively slipped from 43% to 41%, and 41% say they are simply not interested.

It doesn’t help that AI has become the stated reason for cuts. In the same week Sweden blocked its chatbots, the Dutch government announced it would shed an estimated 25,000 central government jobs over the coming years, leaning on AI to do the work. When employees hear “AI strategy” and think “redundancy plan”, resistance is the rational response. In an April 2026 survey of 2,400 knowledge workers in Europe, the UK and the US by Writer and Workplace Intelligence, 29% admitted to actively undermining their company’s AI strategy, rising to 44% among Gen Z.

## The stigma

The subtlest force is social: using AI makes you look worse. In a 2025 study published in PNAS, researchers at Duke University asked some 4,400 participants to evaluate colleagues. Those who used AI were rated lazier and less competent than those who got the same help from a person, and participants anticipated this and hid their own AI use. Everyone has by now received AI “workslop”, output that looks like work but leaves the recipient to do the thinking, so the prejudice has some basis. But it falls on careful users as much as careless ones.

Regulation is about to make that stigma detectable. Since 2 August 2026, the EU AI Act has required AI providers to mark generated text in a machine-readable way. Anthropic now watermarks Claude’s output worldwide; OpenAI began doing the same for ChatGPT in the EU in October. Detection is not yet public, but it is coming, and the watermarks work at the level of the model, which can’t tell a full draft from a corrected comma. Anthropic says a detected mark means a text “may have been processed by Claude”, and that translating or summarising your own writing can carry it. Lawmakers explicitly exempted assistive editing. The watermark doesn’t. Worse, heavy rewriting erodes the mark: in reported tests of OpenAI’s scheme, swapping a quarter of the words cut detection to 17%. The watermark will catch the honest professional who asked AI to check their grammar, and miss the one who had it write everything and paraphrased the result.

Authorship has never been binary. Editors, ghostwriters, translators and speechwriters have shaped “personal” texts for centuries, and we judged the result by its quality and by who put their name to it. A watermark answers a narrower question: did a machine touch this? In a culture that already penalises AI use, that is enough to make a careful professional think twice before letting AI review their work, which is exactly the low-risk, high-value use organisations should want most.

## Who answers for it

The same bias runs through how we judge AI decisions. In compliance-critical work, human judgement is known to be fallible, which is why we built four-eyes principles and layered reviews. Even those miss things, and we accept that as the cost of doing business. An AI system that errs once is declared unfit. Holding AI to a standard no human process meets is stigma in a risk officer’s language. The fair test is to measure both against the same standard: accuracy, consistency and whether the reasoning can be audited afterwards.

Where AI and human review really differ is accountability. A layered human review works because several people look and each of them answers for what they signed. When a reviewer misses something, there is a name, a conversation and, if needed, a consequence. An AI can’t be held to account, and employees sense the gap: in a Bitkom survey this spring, 62% of German workers said it was unclear who is responsible when AI gets something wrong. If nobody owns the machine’s decision, the safest move is to distrust it. The fix is to make ownership explicit: a named person accountable for how the AI review is designed, tested and monitored, just as a head of compliance answers for the review process rather than for every individual file. Accountability can move up a level, but someone still has to hold it.

## What it costs

Put these forces together and you get a predictable pattern. Access is rationed, so experience stays with a small group. Resistance and stigma push the rest underground, so the organisation doesn’t learn from it. What remains visible is the central programme: a handful of sanctioned use cases, a dashboard of licence counts, and a growing suspicion that AI “isn’t delivering”.

The ECB’s numbers show why that matters. The median AI user reports saving about three hours a week, nearly 8% of working time. But because only half of all workers both use AI and report any saving, the gain across the whole workforce shrinks to under 4%. The value of AI in an organisation depends less on how good the tool is than on how many people use it well. Every measure that narrows that base, whether a block, a cap or a reason to hide, cuts directly into the return.

## Two objections

**“We can’t afford unlimited access.”** True, and nobody should offer it. But a cap is a design choice. Tie budgets to demonstrated outcomes rather than flat allowances, and keep basic access universal, because the long tail of occasional users is where most of the learning happens.

**“Compliance comes first.”** It should. But meeting compliance shouldn’t mean shipping a tool that no longer works. Govern the data and the use; don’t amputate the capability.

## Make it normal

The way forward is to remove the reasons people hide. Leaders should use AI visibly themselves and say so. They should judge work on its quality and on who stands behind it. They should write disclosure norms that distinguish assistance from authorship, and give every AI process a named owner. And when they need to control cost or risk, they should govern access rather than withdraw it, because every employee pushed onto a personal account is a lesson the organisation will never learn.

The tools work. Most enterprises have yet to let their people be seen using them.

## Sources

- Sveriges Radio Ekot (23 October 2026); The Next Web, “Sweden’s Government Offices block ChatGPT and Claude on work computers” (October 2026)
- Bloomberg, “Walmart caps usage of an AI tool for employees after high demand” (8 June 2026)
- Engels and Hammermann, “Schatten-KI” (IW-Kurzbericht 63/2026, Institut der deutschen Wirtschaft, 19 September 2026)
- Dias da Silva, Lebastard and Sondermann, “AI adoption and the productivity promise: what workers report” (ECB Blog, 26 August 2026)
- Rijksoverheid, announcement on reducing central government jobs (26 October 2026), via The Next Web (28 October 2026)
- Writer and Workplace Intelligence, enterprise AI survey (April 2026), via Fortune (30 July 2026)
- Reif, Larrick and Soll, “Evidence of a social evaluation penalty for using AI” (PNAS, May 2025)
- Niederhoffer et al., “AI-Generated ‘Workslop’ Is Destroying Productivity” (Harvard Business Review, September 2025)
- Regulation (EU) 2024/1689 (AI Act), Article 50
- The Next Web, “Claude watermark: EU AI Act exemptions and detection API” (13 August 2026)
- Cloud Security Alliance, research note on OpenAI textGrain and EU AI Act text watermarking (9 October 2026)
- Bitkom, “Ein Drittel nutzt KI mindestens einmal pro Woche” (press release, 28 April 2026)

---

# Even If It Stopped Today

By Robert J. van Raamsdonk · Published 2026-10-23 · https://www.contenu.nl/writing/even-if-it-stopped-today/

> Even if AI stopped improving today, it would still transform knowledge work. Why leaders should judge AI on what it delivers now.

If AI never got any better than it is this afternoon, we would still spend the next ten years learning how to use it.

Every conversation about AI I sit in ends up in the future tense.

The optimists talk about what the next model will do: agents that run whole departments, scientific breakthroughs, general intelligence by the end of the decade. The sceptics talk about what it won’t: the bubble that will burst, the plateau that is surely coming, the returns that never arrive. They disagree about almost everything, except the tense. Both are arguing about tomorrow.

That shared habit has a cost. It lets everyone in the room off the hook for today. If the real thing is still coming, there is no hurry to use the current thing well. Wait for the next model. Wait for the hype to settle. Wait for the market to sort out the winners. Meanwhile, the tools on our desks are already doing things that would have sounded like science fiction five years ago, and most organisations are judging them on promises instead of performance.

Try a different test. Imagine that every AI lab stopped today. No new models, no new releases, no further progress. Just what we have now, for the next ten years. Would that be a disappointment, or a revolution?

I think it would still be a revolution, which means most of the debate is about the wrong question.

## The future-tense trap

Hype and doom are mirror images. Hype says: what you see now is only the beginning, so invest in the vision. Doom says: what you see now is overrated, so wait for the correction. Both push the decision into the future, and in large organisations, a decision pushed into the future rarely gets taken.

The futurist Roy Amara put it neatly: we tend to overestimate the effect of a technology in the short run and underestimate it in the long run. The AI debate manages both at once. We overestimate what next year’s model will do, and underestimate what this year’s model already can.

There is also a quieter version of the trap. When AI is always judged against what it is supposed to become, everything it does today looks provisional. A prototype is “just a demo”. A good analysis is “promising”. A working tool is “a pilot”. Nothing counts yet, because the real thing is always one release away. That is how capability that already exists becomes invisible.

## The freeze test

Let’s take the test seriously. What can these tools do, today, if nothing ever improves?

I can only answer from my own work. I use AI every day, for three things in particular.

**Prototyping.** An idea that used to need a team, a budget and a few weeks now becomes something you can click on in an afternoon. The code doesn’t have to be perfect for the conversation to change: people stop debating a slide and start reacting to a thing.

**Making hard ideas graspable.** This is the use I find most underrated. A model can take a dense concept, such as a regulatory regime, a data architecture or a statistical argument, and turn it into an analogy, a diagram or an interactive explainer that someone gets in two minutes. For a consultant, that is half the job. For an organisation, it is the difference between a strategy that a few people understand and one that everybody can act on.

**Writing.** As I say on this site, I write with AI, openly. It tests my arguments, checks my sources and cuts what doesn’t earn its place. The thinking is mine. The speed and the rigour are better than I could manage alone.

None of this depends on the next model. And the evidence was already there with the last one. In 2023, researchers from Harvard, Wharton, MIT and Warwick gave 758 Boston Consulting Group consultants access to GPT-4, a model that now looks old. On tasks within its reach, they completed 12% more work, 25% faster, at 40% higher quality. The models on our desks today are far more capable than the one in that study. Freeze them here, and that is the floor.

Open-weight models make the thought experiment even more concrete. Mistral, from Paris, publishes models that anyone can download and run on their own servers. Even if every lab closed its doors tomorrow, those models would keep working, indefinitely, on hardware you own.

## The capability overhang

If the capability is already here, why doesn’t it show?

According to Eurostat, only one in five EU businesses with ten or more employees used any form of AI in 2025. Denmark leads at 42%; most countries are far below. And among those who do use it, the measured effect is modest. A large Danish study by Anders Humlum and Emilie Vestergaard followed 25,000 workers across 7,000 workplaces and found that AI chatbots saved them, on average, about 3% of their working time.

At first glance that supports the sceptics. Read closely, the figures describe the gap between what the models can already do and what organisations let them do. OpenAI has started calling this the “capability overhang”: the distance between what the technology can do and how it is really used.

We have been here before. Electric motors entered factories in the 1880s, yet productivity barely moved for decades. The economic historian Paul David showed why: factories first simply swapped the steam engine for an electric one and kept everything else the same. The gains came only in the 1920s, when a new generation of managers redesigned the factory floor around the new power source.

AI is in its steam-engine-swap phase. Most organisations have dropped a chatbot into their existing processes, rationed it, stripped it of the features that made it useful, and then measured the result. In “Embrace the Mess” I wrote about how the instinct to establish order first keeps pilots from scaling. It is the same story: the bottleneck sits in the organisation around the model.

## Judge it on its merits

This doesn’t mean we should stop thinking about the future. Progress may well continue at its current pace; it may slow down; it may surprise us. Serious people should prepare for all three. But a strategy built on the next model is a strategy that never starts. The question for any leader is simpler, and more useful: what can this technology do for us now, and what is stopping us from getting it?

As a linguist, I notice how much of the debate is carried by grammar. “AI will transform”, “AI could replace”, “AI might never”. Modal verbs, all of them. Try describing it in the present tense instead, and the conversation changes. AI drafts. AI explains. AI prototypes. AI checks. Those are claims you can test on Monday morning.

## But what about…

**“It still makes mistakes.”** It does. So do people, which is why we built reviews, four-eyes principles and audits. The fair comparison sets AI plus a good process against the same process without it.

**“Why invest now if next year’s model will be better?”** Because the skills, processes and data your people build with today’s tools carry over to tomorrow’s. The organisations that will get the most out of the next model are the ones that learned on this one. Waiting only postpones the learning curve.

**“The productivity numbers don’t show it.”** Not yet, and the dynamo explains why. Productivity statistics measure organisations. Flat numbers usually mean the work hasn’t been redesigned yet.

**“It’s a bubble.”** Perhaps the market is. Valuations can crash while the technology stays. The internet survived the dot-com bust; the companies that had learned to use it came out ahead.

## The present tense

Here is what I’d ask of any leadership team. For one quarter, ban the future tense from your AI discussions. No roadmaps for models that don’t exist, no fear of a correction that hasn’t come. Ask instead what today’s tools can do for your work, and measure that, honestly, against how you work now. Then remove the things that stop your people from using them.

If AI never got any better than it is this afternoon, we would still spend the next ten years learning how to use it.

## Sources

- Amara’s law, as attributed to Roy Amara (Institute for the Future), widely cited
- Dell’Acqua et al., “Navigating the Jagged Technological Frontier” (Harvard Business School Working Paper 24-013, September 2023)
- Eurostat, “Use of artificial intelligence in enterprises” (data extracted December 2025; news release 11 December 2025)
- Humlum and Vestergaard, “Large Language Models, Small Labor Market Effects” (NBER Working Paper 33777 / BFI Working Paper 2025-56, April 2025)
- OpenAI on the “capability overhang” (January 2026), via Axios (19 January 2026)
- David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox” (American Economic Review, 1990)
- Mistral AI, open-weight model releases (Apache 2.0 licence)

---

# Embrace the Mess

By Robert J. van Raamsdonk · Published 2026-10-14 · https://www.contenu.nl/writing/embrace-the-mess/

> Large organisations stall on AI because they insist on order first, when order is what using AI produces.

Every large organisation I work with has a pilot graveyard.

Picture the steering committee. Twelve proofs of concept, all of them demoed, all of them applauded. A contract summariser, a service-desk assistant, an agent that drafts tender responses. Eighteen months later, none of them is in production. Nobody killed them. They were simply asked to wait: for the data clean-up, for the platform decision, for the governance framework, for the review board. By the time the wait was over, the business sponsor had moved on, and the model they were built on was two generations old.

The models are good enough, and have been for a while. What holds these pilots back is a reflex: the large organisation’s instinct to establish order before it allows value. That reflex served us well in finance and in safety. With AI, it is the single biggest reason pilots never scale.

## The pilot graveyard

The numbers are now familiar. MIT’s NANDA initiative reported in 2025 that around 95% of organisations saw no measurable return from their generative AI investments. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The usual explanations are cost, risk and immature technology.

In my experience, most of the pilots that die worked in the lab. They die when they are asked to meet the organisation’s conditions for scale.

## The cleanup fallacy

The first condition is always the data. “Before we do AI, we need to get our content in order.”

I have heard this sentence for thirty years. In the Documentum era, it justified multi-year taxonomy programmes. With SharePoint, it justified migrations that would finally give us one source of truth. Each wave promised that this time the information would be clean, classified and findable. None of them delivered it, because large organisations produce working information: drafts, versions, exceptions, emails that contradict the policy, and the policy that everyone knows is out of date.

A clean-up programme in a large organisation has no end date. Making AI wait for one means it waits forever.

## The bane of platform thinking

The second condition is the platform. Nobody wants a tactical solution. Everyone wants the unified, enterprise-grade AI platform that every future use case will run on.

Two arguments carry this. The first is “let’s not reinvent the wheel”: why build ten point solutions when one platform could serve them all? The second is IP protectiveness: our data is too valuable to touch someone else’s tools, so we must build our own, inside our own walls.

Both sound responsible. In practice, they guarantee that nothing useful reaches the business for a year or more. The platform needs requirements, so it needs use cases. The use cases need a platform, so they wait. The internal chatbot built to protect the IP arrives late, a generation behind what employees use at home, and they notice.

Platforms are worth having. But starting with the platform is a near-certain way to lose the short- and medium-term interest of the business stakeholders who fund the whole thing. The better order is the reverse: ship tactical wins, watch which patterns repeat, and let the platform grow out of what proved useful.

## The human-in-the-loop trap

The third condition sounds the most reasonable: a human must review everything the AI does.

For a single pilot, that works. At enterprise scale, it collapses. Every agent produces proposals, drafts, approvals and escalations, and every one of them lands in someone’s queue. The small group of experts trusted to review them becomes the bottleneck. The oversight meant to make AI safe ends up moving the work to another desk.

It is also bad for the reviewers. In “The Last Interface Is Language” I wrote about what close supervision of AI does to the people doing it. A 2026 Boston Consulting Group study found that intensive AI oversight was the most draining way to use it, and that exhausted overseers made more major errors. The psychologist Lisanne Bainbridge predicted this in 1983: the more you automate, the harder the human’s remaining job becomes. Tired reviewers skim and then rubber-stamp, so a blanket human-in-the-loop rule mostly produces the appearance of control.

The impulse is sound, but it needs limits. Oversight should be proportionate to risk: human judgement where the stakes call for it, automated checks everywhere else. The technology is moving that way. Models such as Jev, launched by TypeSafe AI in September 2026, don’t generate text at all. They return structured decisions, such as a choice, a score or a yes or no, with a confidence value attached. That is what routine triage needs: let the machine decide the obvious cases, and send people only the ones it is unsure about. It is early, and the claims are the vendor’s own. But it points to a future where human attention is spent where it counts.

## Built for mess

For seventy years, computers demanded order. Structured fields, controlled vocabularies, exact syntax. If information was messy, the machine couldn’t use it, so we spent fortunes cleaning it up.

Language models are the first technology built the other way around. I am a linguist by training, and what strikes me most is that these models handle ambiguity the way people do. They read the outdated policy and the email that contradicts it, and they can tell you that the two disagree. They don’t need a taxonomy to do it.

Insisting on clean data before using AI throws away the one advantage that makes it different from everything that came before.

## What embracing the mess looks like

In practice, it comes down to a handful of habits:

**Start where the work is.** Deploy into the real environment, with the real documents.

**Go tactical first.** Solve one painful problem for one team, fast. Fund the platform from what repeats.

**Measure outcomes.** Ask whether the work got better.

**Make oversight proportionate.** Decide per use case where people must decide, and automate the rest of the checking.

**Let structure emerge.** Usage shows you which data matters, which processes need redesign and which controls are needed. That is where clean-up should start.

## But what about…

**“We can’t risk it with our data.”** Some data, no. Much of it, yes, and with the right contracts, the leading providers now offer stronger controls than many internal environments. Protecting IP by blocking every tool mostly protects it from being used.

**“Without a platform, we’ll end up with a hundred point solutions.”** Probably fewer than you fear, and they will teach you what the platform actually needs. A platform designed before anyone has used AI in earnest is designed on assumptions.

**“Regulators expect a human in the loop.”** For some decisions, rightly so. But “where the law requires it” is a much smaller set than “everything”. The EU AI Act asks for human oversight in proportion to risk.

## Where the mess stops

Embracing the mess still has limits. Security, privacy and accountability are non-negotiable. Regulated decisions, such as those about credit, health, employment or legal rights, need real human judgement and an audit trail. Somebody must own every AI system in production. Keep the controls, and stop making every one of them a precondition for starting.

## Order is the outcome

Large organisations will not become tidy before they adopt AI. They never have, and they never will. The ones that get value are the ones that start in the mess, learn from it, and let order emerge from use.

Order is the return on the investment.

## Sources

- MIT NANDA, The GenAI Divide: State of AI in Business 2025 (August 2025)
- Gartner, press release on agentic AI project cancellations (June 2025)
- Bedard et al., “When Using AI Leads to ‘Brain Fry’” (Boston Consulting Group, Harvard Business Review, March 2026)
- Bainbridge, “Ironies of Automation” (Automatica, 1983)
- TypeSafe AI, Jev documentation; nexos.ai, “What is Jev?” (2026)
- Regulation (EU) 2024/1689 (AI Act), Article 14 on human oversight

---

# The Last Interface Is Language

By Robert J. van Raamsdonk · Published 2026-10-01 · https://www.contenu.nl/writing/the-last-interface-is-language/

> AI is weakening our ability to read, write and think at length, just as those skills become the key to using it.

We are outsourcing reading and writing to machines at the exact moment reading and writing became the way we operate machines.

Picture a student with a chapter to read by tomorrow. She pastes it into a chatbot and asks for five bullet points. Down the hall, a junior engineer watches four AI programs write code he can’t quite follow, and approves it anyway. Both are skipping the part that used to make them better at what they do, and neither would call it cheating.

For seventy years, using a computer meant learning its language: menus, syntax, formulas. That era is over. We now build software by describing what we want, research by asking questions and brainstorm by talking. The skills that matter most are understanding what we read and hear, and saying precisely what we mean. Reading and writing are where those skills are built, and they are exactly what AI now does for us.

## We stopped doing the work

AI didn’t start this decline, but it is removing the last places young people still had to do the work. Smartphones had already shortened attention, and American high-school seniors posted the lowest reading scores since testing began in 1992. What remained was the essay you had to write and the report you had to read. By late 2025, more than half of US teens had used chatbots for schoolwork. And the first international reading test taken after ChatGPT arrived, PISA 2025, showed scores falling sharply across the developed world.

The early research suggests the cost is real. In a 2026 trial run by Anthropic, one of the AI companies, junior developers who learned a new coding tool with an AI assistant scored 17% lower on a test of what they had learned than those who worked it out by hand, and they weren’t significantly faster. At MIT, most people who wrote an essay with ChatGPT couldn’t correctly quote a line of it minutes later. Reading and writing are how understanding gets built. Skip the process and you skip the understanding. Attention works the same way: holding a long document in your head is trained like endurance, and a generation that receives every text as five bullets never trains it.

## Beyond the United States

Europe’s literacy advantage lives in its older generations, and it is not being passed on. Northern European adults still read far better than Americans. But among 15-year-olds, only Ireland and the UK now match the US, and the Netherlands has fallen far behind. Meanwhile, 94% of British university students already use AI for their graded coursework. In India, where much of the world’s software is now built, the problem is foundations never laid: only 27% of rural eight-year-olds can read a simple text, while at the elite Indian Institute of Technology in Delhi, four in five students use AI, all of them for writing.

## The keyboard is now a sentence

In 2023 the AI researcher Andrej Karpathy joked that “the hottest new programming language is English.” Today it’s simply how work gets done. Developers describe what software should do instead of writing it line by line. Research is a conversation whose answers are only as good as its questions. Strategy starts as a dialogue with a model.

All of this runs on two skills. The first is saying exactly what you mean, typed or spoken. A good prompt is a good brief, and the machine can only mirror the clarity you bring. The second is understanding what comes back. AI sounds equally confident whether it is right or wrong, usually answers in code and documents far too long to listen to, and catching its mistakes takes slow, sceptical reading. People who think clearly in language get enormous leverage from AI. Everyone else becomes a passenger: able to produce text, unable to judge it.

## Managing the machines is hard work

The industry’s answer is that humans will move up a level and “manage the agents”: the AI programs that now carry out whole tasks on their own. But supervising is harder than doing. It means reading at a higher intensity: switching between parallel threads, absorbing piles of output, rebuilding each context and passing judgement every few minutes.

I know this firsthand. After a day of intensive AI-assisted development, I am wrung out; it takes a real toll on mental endurance. I’m not alone. A 2026 Boston Consulting Group study of nearly 1,500 workers found that closely overseeing AI was the most mentally draining way of using it, and those who hit what the researchers call “AI brain fry” made major errors 39% more often. The psychologist Lisanne Bainbridge predicted this in 1983: the more a system is automated, the harder the human’s remaining job becomes. Tired supervisors skim, then rubber-stamp, then burn out. Managing AI raises the demand for deep reading and attention, just as those skills are fading.

## But what about…

**“We said the same about calculators.”** Calculators took over a procedure and left the thinking to us. Reading and writing are the thinking.

**“We’ll just talk to it.”** Voice changes the channel. A spoken request still has to be as clear as a written one. And when a model turns your rambling into a tidy plan, it has done the thinking for you. Writing is still where clear expression is learned.

**“The evidence is mixed.”** It is, and the newest studies show why. In a 2026 experiment, students who used AI to explain ideas learned more than students without it, while those who let it write for them lost their gains as soon as it was gone. Anthropic’s coding trial found the same split. AI is an aid when it explains and a substitute when it answers, and answers are mostly what it is asked for.

**“AI will soon do all of this for us.”** Perhaps. For now, every task handed to AI is one more thing someone has to direct and check. And even a system that needs no supervision has to be told what we want; the more capable it is, the more faithfully it carries out a careless wish. The bet is also lopsided. If AI never needs us, a literate generation has lost little. If it does, and for years to come it clearly will, a generation that can’t read deeply has lost everything.

## Protect the struggle

Banning AI would not help. Reading and writing need a change of status: from old-fashioned basics to the core curriculum of the AI age. The principle is the same everywhere: protect the struggle first, then add the tool. Schools should guard unassisted reading and writing the way they guard exercise, with whole books read, essays written in class and work defended out loud, and only then teach AI as an extension of those skills. Employers should assess young staff on the quality of their judgement, because a company that automates all the hard work will have no senior people in ten years. And each of us can keep a simple rule: read the source before the summary, and write the first draft before asking for help.

## The last interface

Every earlier computing revolution rewarded people who learned the machine’s language. This one rewards people who master their own. A generation that can’t read deeply or write precisely will hold the most powerful tools ever built and be unable to steer them.

Reading slowly, writing clearly and thinking for long stretches were never only about operating machines. They are how we understand the world and hold power to account. The technology that finally speaks our language may leave us unable to speak it ourselves.

## Sources

- NAEP 2024 Reading, Grade 12 (NCES, September 2025)
- How Teens Use and View AI (Pew Research Center, February 2026)
- PISA 2025 U.S. Results (NCES, September 2026) and Dutch results (University of Twente, September 2026)
- Shen and Tamkin, “How AI assistance impacts the formation of coding skills” (Anthropic, January 2026)
- Kosmyna et al., “Your Brain on ChatGPT” (MIT Media Lab, preprint, June 2025)
- Survey of Adult Skills 2023 (OECD, December 2024)
- Student Generative AI Survey 2026 (HEPI, March 2026)
- ASER 2024 (ASER Centre, January 2025)
- IIT Delhi Generative AI guidelines and student survey (IIT Delhi, 2025)
- Karpathy, “The hottest new programming language is English” (X, January 2023)
- Bedard et al., “When Using AI Leads to ‘Brain Fry’” (Boston Consulting Group, Harvard Business Review, March 2026)
- Bainbridge, “Ironies of Automation” (Automatica, 1983)
- Contractor and Reyes, “Experimental Evidence on the Learning Impact of Generative AI” (arXiv preprint, July 2026)