Two executives at the centre of AI at work now say it is producing a flood of unchecked, empty output. The flaw starts with AI itself, which produces text that looks finished when it is not. The failure is that people skip the steps that would catch it. Those steps need skills formed at school and practised at work – reading with understanding, checking claims, and owning the result.

Shopify’s 2025 rule: AI first, then new hires

In March 2025, Tobi Lütke, the founder and chief executive of Shopify, told his staff that using AI well was now “a fundamental expectation of everyone at Shopify”. He published the memo on X in April 2025, after it began to leak.

The memo went further than encouragement. Before asking for more staff or money, teams “must demonstrate why they cannot get what they want done using AI”. Shopify also added questions about AI use to its performance reviews.

Other technology firms took the same line within days. Micha Kaufman, the chief executive of Fiverr, told his staff not to “hire more people before we learn how to do more with what we have”. At Shopify, using AI became something employees were measured on.

The problems reported eighteen months later

Eighteen months later, Lütke described what happened next. On The Knowledge Project, a podcast hosted by Shane Parrish, released on 15 September 2026, he said: “The failure case now of lazy work is not lack of output.” Staff were using AI to produce emails and code they had not read, and passing them to colleagues to deal with. “We call those ‘slop grenades’ that people toss at each other,” he said. “Machines can’t take responsibility. Humans take responsibility.”

On 31 August 2026, two weeks before Lütke’s interview, Ryan Roslansky described the same problem from the other side in a LinkedIn post. Roslansky is a former chief executive of LinkedIn and, until 1 October 2026, was the Microsoft executive vice-president overseeing Office and Teams. Microsoft announced that day that he will leave the company at the end of the year. Someone had sent him a document that was clearly written by AI. He read it, then had his own AI summarise it, and “ended up with nothing that brought new thinking to the idea”. He called it “a doom loop”, and warned that “we’ll end up with a very expensive way to avoid writing and reading in an ocean of sameness”.

Roslansky is no AI sceptic. In remarks reported in October 2025, he said that “almost every email that I send these days is being sent with the help of Copilot”, Microsoft’s AI assistant, which he called a “second brain”.

The two accounts fit together. Lütke describes the sender, who hands over work nobody checked. Roslansky describes the receiver, who reads it and finds nothing new or of added value in it. Between the two, what is missing is not just thinking but a set of skills a person brings to the work, such as reading carefully, checking claims, and adding their own judgement.

Why every tool needs a skilled person

AI tools are useful, but they cannot be relied on completely. They make mistakes, miss what matters, and produce text or code that looks finished when it is not. Whatever the tool provides, a person has to check it, correct it, and improve it before passing it on. Both executives are describing what happens when that last step is skipped. Doing that step well takes a set of eleven ordinary skills:

  • Structuring and planning – breaking a task into steps yourself and deciding how and where an AI tool can support you.
  • Reading with understanding – getting the real point of an answer and noticing what is missing.
  • Critical thinking – asking how we know a claim is true, rather than trusting a confident tone.
  • Fact-checking and sources – checking claims against real sources.
  • Asking good questions – writing clear prompts with context, and following up until the answer is useful.
  • Synthesis – combining several sources into your own argument.
  • Independent writing – writing in your own voice, with AI as a critic rather than an author.
  • Self-assessment – knowing what you have actually understood.
  • Academic honesty – knowing where help ends and cheating begins, and saying how you used AI tools.
  • Privacy and digital responsibility – protecting your own data and other people’s.
  • Presenting and defending your work – explaining your work in your own words and answering questions about it.

None of these is new, and none is technical. They are the skills a good school has always tried to teach and practise. What has changed is that AI produces something that looks finished, so it is tempting to skip them.

In June 2026, the OECD and the European Commission published a joint framework for AI literacy in primary and secondary schools. It describes the knowledge, skills, and attitudes young people need to use AI well. The skills above overlap substantially with it. The framework also makes a point worth repeating: simply using AI tools does not, on its own, build any of them.

Where these skills are formed

These habits are first formed long before anyone joins a company. They start at school, when a student first decides whether to read a source or to let a machine summarise it.

That moment has already arrived. In the United States, fifty-four per cent of teenagers aged thirteen to seventeen say they have used AI chatbots to help with schoolwork, according to a Pew Research Center survey published in February 2026.

They are also growing up in a media world that rewards skimming. Social networks are built around short videos and posts that take seconds to consume. Long reading, the kind that builds attention and understanding, has to compete with that every day.

The measured results are not encouraging. In the OECD’s PISA tests of fifteen-year-olds, the average reading score across OECD countries fell by ten points between 2018 and 2022. The OECD called it an unprecedented drop. Its Survey of Adult Skills, published in December 2024, found that adult literacy had declined or stagnated over the past decade in most member countries. Among the European Union countries surveyed, adult literacy fell in seven and improved significantly only in Finland and Denmark.

I do not claim that short-form media alone caused these declines. But a generation that reads less and less deeply, and now has a machine that will read and write on its behalf, is exactly the workforce Lütke and Roslansky are worried about. This is why teaching and refreshing these skills alongside AI matters more now, not less.

What companies and schools can do

Companies that expect their staff to use AI should train and refresh these skills, in the context of the AI tools they actually use. Learning the ever-evolving tools themselves is not enough. Staff need practice in writing the right prompt, reading AI output critically, checking it against sources, and adding their own judgement before they hand anything on.

This kind of training can reduce the cost of unchecked output. Every unchecked document Lütke describes creates work for someone else, and every summary of a summary, as in Roslansky’s doom loop, adds cost without adding thought.

Schools are the other place, and the earlier one. Companies can retrain adults, but habits are easier to build early. Schools already teach most of these skills. What they need now is to reinforce them and to practise them with AI, not around it. Banning AI in schools would leave students to learn its habits alone.

These skills are also useful well beyond work and school. They are how a person reads the news, judges an offer, or decides whom to believe.

Caring enough to use them

Skills are necessary, but they are not enough. An employee can know perfectly well how to read, check, and think, and still choose to forward whatever the machine produced. The skills only matter if people care enough to use them at work.

That is Lütke’s point about responsibility. A machine cannot be held to account for a document, so a person has to be. When someone puts their name to AI output they have not read, they are not saving time. They are handing the cost to the next person, who may hand it to a machine again, which is how Roslansky’s doom loop begins.

Building that motivation is the company’s responsibility. Leaders and managers need to align people behind the company’s values and goals, even more than before, and give them a reason to stand behind their own work and explain it, beyond the pay cheque and the job description. That also means rewarding quality, not just quantity.

The lesson from Shopify and Microsoft is not that AI made people worse at their jobs. It is that AI makes it easy to stop doing the part that was always ours. Training today’s staff, and the next generation, to keep reading, checking, and owning what AI produces is how companies and schools can make sure AI leads to better work, not just more of it.

Sources