When AI Starts Doing Our Thinking for Us
Earlier this month, I found myself reading another engineering lead's Claude conversation logs. I was taken aback by how extensively AI had become involved in their professional and personal life. It was a fascinating, warts‑and‑all view of a level of reliance I hadn't expected.
There were technical questions, as I would expect, but they were also using AI to draft ordinary team emails, basic Slack messages, to make routine decisions, and to help interpret situations whilst relying on the same AI to choose a response. Messages were being prepared for later delivery. There were even requests for help composing messages to loved ones, in what appeared to be otherwise ordinary personal conversations.
I recorded the experience in my Timeline. What caught my attention was how often the assistance seemed to reach far beyond expressing a thought into deciding what to think in the first place.
What it left me wondering was much broader. If an assistant becomes our usual first stop for interpreting a situation, choosing a position, and finding the words to express it, where are we doing the thinking that makes the response ours?
Some Thinking is Sensible to Hand Over
We have always used tools outside of our own heads to make thinking easier. Notes, calendars, diagrams, calculators, and other people all help us do things we would struggle to manage unaided. There is no obvious prize for remembering an appointment instead of writing it down. I ‑ for example ‑ rely incredibly heavy on the Reminders app on my phone to make sure that I don't get too engrossed in my work and forget to stop for lunch!
Researchers call this cognitive offloading. Risko and Gilbert's review describes ways to reduce mental demands through external aids, including the perfectly useful possibility of freeing capacity for other work. Offloading is not inherently a failure, and effort is not valuable merely because something was difficult.
AI can extend that help into reasoning. It can suggest an interpretation of a situation that we had otherwise missed, point out an inconsistency in that situation, or help us investigate a question we did not know how to otherwise approach. Asking an LLM to challenge our bias (or assumption we've drawn) can be a way of engaging more seriously with a problem. Help with wording an email, or a message, etc can also make it easier to express an intention you have already thought through.
I don't think there is a useful rule that says typing the words manually by ourselves makes them more authentic. People have always asked for advice, had someone else edited their drafts, and struggled to say what they mean. Assistance can often help make someone's meaning clearer.
The situation becomes more awkward when the intention itself arrives with that external influence. Imagine receiving a difficult message and immediately asking an LLM model what the other person really means, whether we should be annoyed, and what we should say back. This lead was doing exactly that. With ourselves in that situation, we might use the AI answer to reconsider our first reaction. We might also accept its account of the situation and send the response it proposed before deciding whether either actually reflects our own view.
The logs I audited showed requests for help, not everything the person thought about the answers. They don't establish that someone has lost the capacity to think independently in isolation, but a considered‑sounding reply tells the recipient very little about who did the considering.
Can We Still Judge the Answer?
Checking an answer requires enough understanding to recognise when it is wrong. A human approval step does not tell us whether that understanding was present, whereas clicking a button is easy to observe; judging the reasoning behind the click is rather harder.
A 2025 Microsoft Research and Carnegie Mellon survey of 319 knowledge workers found that greater confidence in AI was associated with less reported critical thinking; greater confidence in their own ability was associated with more. It measured reported behaviour and effort, not a change in people's ability to reason. Some thinking had also shifted towards checking, combining, and overseeing the output. That still requires judgement, even when producing the preliminary answer takes less effort to get to.
The difficulty is keeping that scrutiny meaningful. Earlier research on automation bias and complacency describes people accepting incorrect advice or failing to notice when a system needs challenging. Those studies concern earlier decision‑support systems, but the question will be familiar: are we evaluating an answer, or accepting it because the system usually seems to know what it is doing?
How the help is provided matters too. In a 2025 school mathematics experiment, an ordinary GPT interface improved practice performance but left pupils doing worse on subsequent unassisted tests than those without AI help. A guided version avoided that issue. A separate university experiment found that a carefully designed AI physics tutor improved immediate learning‑test results. These are findings about learning particular material, not a measure of an adult's everyday judgement. They give me a reason to ask what the assistance asks us to do, rather than assume that removing the effort either helps or harms us.
When the Answer is About Another Person
This matters in engineering leadership because so much of the job involves deciding what to make of incomplete information. A disagreement about an approach might reflect a technical constraint, a misunderstood requirement, or something the team has not yet explained clearly. I need to weigh those possibilities, explain the trade‑offs, and take responsibility for the decision. AI can help me examine them. Accepting its preferred explanation without questioning it would leave an extremely important part of my job undone.
I have written about the difference between producing management updates and doing the work of management. The same distinction is important when mentoring another engineer. A clear, reassuring response can still rest on a poor reading of what they need. If we cannot explain why we are offering that advice, its fluency does not help very much.
The model receives the account we give it. If that account leaves something out, repeats a mistaken assumption, or explains another person's actions unfairly, a helpful‑sounding answer may begin from the same problem. I might ask whether I am being reasonable and come away with a polished defence of the position I brought to it, because AI tends to promote and inflate your own bias.
A 2026 study in Science found that people receiving affirming AI responses about interpersonal conflict became more convinced they were right and reported less willingness to repair the situation than people receiving disapproving responses. They also expressed more trust in the assistant and more interest in returning to it.
The experiments had no neutral control and measured immediate judgements and intentions, not what people eventually did. What concerns me is the possibility of becoming more certain of our own account, and keener to consult the assistant again, without having examined whether that account is fair. Feeling understood is rather different from understanding somebody else.
Now imagine both sides of a workplace exchange doing this. One colleague asks AI to interpret a problem and draft a message. The other asks AI to interpret that message and prepare a reply. The exchange could become clearer: less hurried wording, fewer unnecessary misunderstandings, more room to consider another perspective. It could also become a conversation in which each person responds to an interpretation of an interpretation, without ever checking the underlying disagreement with their colleague.
It interests me because a perfectly civil exchange might still leave both people less clear about what the other actually thinks. Politeness is useful. It does not resolve a disagreement we have not recognised.
Experiments with AI smart replies found faster messaging and higher ratings of cooperation and affiliation. Suspected AI use was associated with less favourable impressions, although that suspicion was measured rather than randomly assigned. These were brief exchanges between strangers, so they tell us little about what happens in an intimate relationship. They do show that help with wording can improve an exchange.
That distinction matters when the messages are personal. Help finding words for an apology may let someone express remorse they already feel. It may even help them see why the apology is needed. I would be much more concerned if producing an acceptable apology became a way to avoid considering what happened or taking responsibility for it. The wording alone would not let the recipient reliably tell those situations apart.
A message to someone close to us has a purpose beyond delivering an efficient summary. We may be trying to understand them, work out what we feel, or acknowledge something we have found difficult to face. Some of that happens whilst we struggle to express it. Removing the struggle could help us get past an obstacle. It could also let us avoid the part of the conversation we needed to have with ourselves.
Nobody needs another hour spent formatting an update. But deciding whether we agree, noticing that we may have been unfair, or finding out what we want to say to someone we care about takes effort too. I would be much less comfortable handing that over.
What happens if we stop doing those things routinely? Some of the thinking we are making easier to skip may be thinking we still want to practise.