The wrong question
In 2008, Nicholas Carr asked whether Google was making us stupid11Nicholas Carr, "Is Google Making Us Stupid?", The Atlantic, 2008.. Three years later, Betsy Sparrow and her colleagues tested it22Sparrow, Liu and Wegner, "Google Effects on Memory", Science, 2011. Hard to replicate since, so cited as evidence of its time.. In their experiments, people remembered fewer facts and got better at remembering where to find them. Memory had been traded for access. Nobody explicitly signed that deal and every Google user accepted it.
The question of whether a tool is making us stupid is back with generative AI and it is still badly framed. A more useful one is what we exchange each time we use the tool. Long before search engines, writing and printing had each moved an operation of the mind outside of us33Cognitive offloadingDelegating a mental task to an external tool.In the glossary. Each came with gains and with trade-offs.
What is new this time
Every technology is an exchange: we gain what the tool does and we pay with the exercise we no longer practice. With Google we handed over the memory of facts. With generative AI we hand over part of our reasoning and our writing. This is why the terms matter more this time.
On one side, we gain time and quality. We also borrow and develop skills we did not have. On the other, we pay a cognitive tax44Cognitive taxThe cost paid at the moment of delegation.In the glossary which often starts with checking what the AI tool produces. Depending on how we use it, the tax extends to our attention and our understanding, with more cognitive fatigue55Brain fryFatigue from supervising AI beyond your capacity.In the glossary along the way. Over time, it can reach learning itself66Cognitive debtJudgment eroding when task and responsibility both go to AI.In the glossary.
When nobody negotiates, we pay the default price, the one the interface has set.
The fine print
The fine print holds four costs that nobody planned for. We delegate more than we think77Delegation driftHanding over cognitive work a little at a time, without ever deciding to.In the glossary: in an MIT Media Lab preprint88Kosmyna and colleagues, "Your Brain on ChatGPT", MIT Media Lab, 2025. Preprint with 54 participants., students who wrote essays with ChatGPT struggled to quote a sentence from the text they had just handed in. The tool gets things wrong without flagging it99The jagged frontierThe invisible line between what AI does well and badly.In the glossary: in the Harvard and BCG study of consultants1010Dell'Acqua and colleagues, "Navigating the Jagged Technological Frontier", Harvard Business School and BCG, 2023., those who used AI on a task just outside its abilities got the answer wrong more often than those who worked without it. Unchecked work passes to whoever receives it. Researchers at Stanford and BetterUp call it workslop1111WorkslopA polished AI-generated piece that wrongly gives the impression of a job well done.In the glossary: AI output that looks finished and leaves the real work to the colleague who receives it1212Niederhoffer and colleagues, "AI-Generated Workslop Is Destroying Productivity", Harvard Business Review, 2025.. Saying you used AI has a cost too. Across thirteen experiments, Oliver Schilke and Martin Reimann found1313Schilke and Reimann, "The Transparency Dilemma", Organizational Behavior and Human Decision Processes, 2025. that people who disclose their AI use are trusted less, and being found out by someone else costs even more.
These costs are real, yet they are not inevitable and most of them depend on how the tool is used. So the price of this cognitive new deal can be negotiated.
The collective deal
A team signs a deal too, whether or not anyone says so. Without a shared agreement on what gets delegated, there is no psychological safety. Without psychological safety, trust drops, because nobody dares to say "I used AI for this" or "I didn't understand what it produced". When Google studied 180 of its own teams with "Project Aristotle"1414Google re:Work, the guide to team effectiveness from Project Aristotle., psychological safety mattered more than who was on the team.
The collective deal can also be a winning one. Ramp and Revelio Labs followed more than 21,000 US companies1515Ramp, data note on heavy AI adopters and hiring.. Those that invested heavily in AI grew their headcount by about 10 percent in the two years after adopting it.
Signing your own deal
Signing your own deal takes two decisions: what you keep, and where the time saved goes.
As AI tools seem to do more, the question shifts from what we can delegate to what we choose to keep. The Joy Of Not Automating (JONA)1616Joy Of Not Automating (or JONA)Keeping a task for yourself, even when AI could do it, because it is part of your professional signature.In the glossary starts from that question: what do I want to keep for myself or my organisation? AI might well be able to do the task, or even the whole project. I keep it because it is my professional craft, the signature that makes me stand out.
The second decision concerns the time saved, and it has two paths: automation or augmentation1717Automation versus augmentationAI as substitute for work, or as extension of it.In the glossary. Automating means doing what we already did, faster and frees up time. Augmenting means doing what we could not do until now. It invests that time in work that was out of reach and that's where it gets interesting.
Four questions
I reason that a cognitive new deal is favourable when it makes you more capable and not just faster, and when nobody else pays the price. Four questions help read its terms:
- Do I know what I am exchanging?
- Is it a win-win, for me and for the people I work with?
- How do we grow our professional value, individually and collectively, once the fine print is read?
- Does the time saved augment this work or does it automate it?
The deal still open
The Google deal is settled. Nobody signed it and yet everybody lives with its terms. The AI deal is still being written and its terms are set one use at a time, by each of us and by every team. That leaves a window to read them before they harden into the default.
Margaux PelenFounder of episcope