When did you last have a thought that was yours from end to end? Or build something where every judgement along the way was your own?
For anyone who works with AI, that question is getting harder to answer. A plan arrives already weighed, a draft arrives already argued, an application arrives with its architecture already chosen. We accept them because they look good, and often they are. But each one carries decisions that someone, or something, made on our behalf.
Who makes those decisions is the question we have built somethingg around, because we think it is the part of AI that most needs getting right.
What AI gives, and what it takes
The rise of generative AI has brought a wave of tools, agents and integrations that make working life easier. They take our prompts and do things, make things and connect things. They do it fast, and they do it to a good (enough) standard.
The gains are measured, not imagined. Customer support agents given an AI assistant resolved about 15% more issues an hour, and the least experienced gained the most . Consultants with access to a model completed 12.2% more tasks, about 25% faster, and those who started out weakest improved by more than those who started strongest . People who never trained as engineers now build and ship software. We use these tools daily, and we have no interest in arguing anyone out of using them.
But something has changed that is easy to miss inside all that speed. Earlier tools did the work you specified. AI decides what the work is. A calculator does the sums and a search engine finds the pages, but neither can decide what the sums are for or what the pages mean. A model can. It frames the problem, weighs the options, reaches a conclusion and carries it out, all in one reply.
It helps to see every piece of work as two layers. The thinking layer is where the judgements are made: what the problem is, what matters, what to believe, what to do and whether the result is right. The execution layer is everything that carries those judgements out. Tools have always lived in the execution layer. They left the thinking to us, not out of principle but because they could not do it, and the skill that thinking builds stayed with us too.
None of us has ever thought entirely alone. We have always leaned on teachers, books and colleagues, and the good ones leave us thinking better rather than thinking less. What is new is a tool that can take the thinking layer from us, in the same reply that does the execution.
With that ease comes a sacrifice. The more you use these tools, the faster you move and the more you build, the more of your thinking you hand to the model, usually without noticing. That is a bad trade. For you, and for everyone around you.
What it costs
The costs of handing over the thinking aren’t immediately visible and the work keeps looking fine while they build up.
For you. Thinking is how skill gets built. Decades of research on learning find that the conditions which make learning feel harder in the moment are often the ones that make it last . Take that effort away while the output keeps coming, and one of two things happens.
If you are new to the work, the skill never forms. People are building and launching applications without ever having made the judgements those applications depend on, and it shows. A scan of more than 5,600 publicly reachable applications built on AI coding platforms found over 2,000 high-impact vulnerabilities and 175 exposures of personal data . From the inside, a working application feels exactly like knowing how to build one.
If you are experienced, the skill erodes. In one study, the rate at which doctors found precancerous growths during colonoscopies done without AI fell from 28.4% to 22.4% within months of AI assistance becoming routine . In a controlled experiment, people who had worked with an AI assistant and then had to work without it solved fewer problems and gave up sooner .
And the part you stop practising is the part becoming most valuable. As models take on more of the execution, what is left for people is the judgement, and in work where one weak step can spoil everything else, that judgement decides what the rest is worth .
For your organisation. An organisation is only as sound as its ability to check its own work, and checking is thinking. Lisanne Bainbridge called this the irony of automation: the better an automated system gets, the less practice people have at the part of the job that still depends on them, which is precisely the part that matters when the system fails . A tool that automates judgement sharpens the irony.
The competitive cost is larger still. Every organisation can buy the same few models, and the leading ones now sit close together on the most widely used public ranking . An organisation whose people hand their judgement to those models ends up with its competitors’ judgement, and with its competitors’ ideas. People who brainstormed with ChatGPT came up with ideas that were individually more creative and, taken together, less varied than ideas produced without it . Advantage comes from seeing what others have not. An organisation that thinks what everyone else thinks has given away the one thing that was hard to copy.
For everyone. The first cost to a society is its next generation of expertise. Employment of workers aged 22 to 25 in the most AI-exposed occupations now stands 19% below where it would be had it kept pace with less-exposed peers, while experienced workers show no comparable gap . The entry level is where expertise has always been built, and a society that stops building it will not notice for years, because the people who already have it are still at work.
The second is its variety. Progress depends on people seeing the same problem differently, and on someone checking the answer another way. When so many people frame, draft and decide with the same few models, that variety narrows. Writers given story ideas by a model produced stories that readers rated more creative, better written and more enjoyable, and that were more similar to one another. The authors liken it to a social dilemma: each writer is better off, and together they produce a narrower range of work . Sameness also makes a system fragile. When many decision-makers rely on the same algorithm, its errors stop being independent, and the decisions of the whole group can get worse even when the algorithm is more accurate than any one of them alone . A monoculture of thought fails the way a monoculture of crops does: all at once.
That is bad for all of us, and no single person can see it or fix it, because each of them is doing better.
The third is its safety. Much of what is proposed for keeping AI safe ends with a person checking what it did. The EU’s AI Act requires high-risk systems to be built so that the people overseeing them can stay aware of their own tendency to over-rely on the output, a tendency it names automation bias . The International AI Safety Report lists cognitive offloading among the risks to human autonomy, though it notes the research is young . Oversight is only as good as the thinking of the people doing it, and a population that has handed its thinking over can approve what a model produces without being able to tell whether it is right. Repeated across an economy, that leads to what safety researchers call gradual disempowerment: no single step looks like a loss of control, and human influence can erode all the same .
Why it happens
Nobody sets out to hand over their thinking. It happens for several reasons.
The layers come combined. The AI products most people use today return the thinking and the execution as a single finished thing. Ask for help with a decision and you get the decision. Ask for an application and it arrives with its data model and its security already decided. Ask for the short version of a report you have not read and it arrives with a view of which parts mattered. The judgements inside are rarely marked as judgements, and nothing separates the ones you made from the ones made for you.
It is hard to see from the inside. In one experiment, people wrote about whether social media is good for society, some of them with a writing assistant set up to argue one side. Those with the assistant were about twice as likely as those writing alone to argue its side, and their own views shifted towards it in a survey afterwards. On average they wrote nearly two thirds of their sentences themselves. Most thought the suggestions were balanced, and most did not think the assistant had changed their argument . In another, people using an AI coding assistant wrote less secure code than people without one, and were more likely to believe it was secure . The influence is real, and it feels like your own judgement.
The convenience comes first. The trade is lopsided in time. The convenience arrives immediately, while the cost arrives slowly and can always be deferred. People act on what they can feel, and so do the companies selling to them.
The incentives point the same way. The labs building these products, OpenAI, Anthropic and Google among them, are paid for how much work their models take on. All three sell access to their models by the token, and all three are building agents, which take on whole tasks and consume far more tokens doing it. Anthropic reports that its agents use about four times as many tokens as a chat, and its multi-agent systems about fifteen times as many . Its own usage data shows the drift: directive use, where a task is handed over whole with minimal back-and-forth, rose from 27% to 39% of usage in the eight months to late 2025, before falling back to 32% .
None of this requires bad faith. Anthropic publishes that data, and in 2025 all three labs shipped a mode that guides instead of answering: study mode in ChatGPT in July, then Guided Learning in Gemini and learning mode for every Claude user in August. Each shipped as an option, off by default, next to a product whose ordinary behaviour is to do the thinking for you. When a safeguard pulls against what a product is measured by, the safeguard is the first thing to give.
Building for the thinking layer
Every cost above is decided in the same place. The skill you build, the checking your organisation depends on, the variety a society learns from and the oversight that keeps AI safe all live in the thinking layer. Get that layer right and AI makes people better at their work. Get it wrong and it quietly makes them worse.
It is possible to get right. Around a thousand secondary school maths students were given one of two AI tutors built on the same model. One answered like a standard chat assistant; the other gave hints designed by teachers. When the tutors were taken away, the students who had used the standard assistant did worse than students who had never had AI at all, and the students who had used the hints did about as well as those who had never had it . Same model, same students. The only difference was who did the thinking.
The labs with the most capable models have the least reason to build that way. A product made for the thinking layer asks the model to do less of your work, which means fewer tokens, and for a lab that makes it a mode to offer beside the real product, not the product itself. If anything, their incentives run against prioritising it.
That is why somethingg exists. We build products for the thinking layer, to be the interface between people and the models they work with: the model does what it is good for, and the thinking stays yours.
kelve, your thought partner
kelve is the first thing we have built this way. You press think aloud and say what is on your mind. kelve writes it down as you talk and reads along: it searches where you have reached the edge of what you know, asks a question that opens the next part up, and points at what the thought has taken for granted. It never thinks for you.
Nothing is written into the thought but by you. What kelve does stands in the margin beside the words it is about, and what your situation turns on is kept across thoughts and said out loud every time it is used, so nothing is held that you cannot see.
Stop talking and the whole thought is there, with everything kelve did beside the line it came from. When you have worked something out, kelve can organise it into a short piece made only from what you said and the sources you kept.
The measure it is held to is whether what you arrive at is yours: understood, and something you could defend without it.
Who does your thinking? It should be you, with AI beside you rather than in your place.