“I’m here today to talk about thinking for yourself,” opened Advait Sarkar, Senior Researcher at Microsoft Research. What ensued was one of the most insightful TED Talks at the recent TEDAI Vienna 2025, which became home to global AI practitioners and thought leaders for three days.
Thinking for yourself? In the age of AI?
What tasks shaped your day today? Were you going through emails or collecting thoughts after a meeting, drafting a report, conducting data analysis or coding?
Have you used AI to summarize, analyze, prototype, or draft any of those? Did you look at the source materials at all, or did you just glance over the AI output and use it as it is?
With this “outsourced reasoning”, we are at risk to become intellectual tourists who engage with ideas only indirectly, leading to alienation from our own craft. The gist of Advait’s talk was in a sense a warning, but also a hopeful vision of how we can build things now, so that we can mitigate the risks later.
Advait points that AI can play the role of provocateur, with the aim of improving human critical thinking, rather than diminshing it. He argued that the real value of generative AI is not to replace thinking but to stimulate and extend it. AI as a “tool for thought” challenges the user, surfaces alternative viewpoints, and supports metacognitive work, whereas a pure “assistant” simply delivers results and risks making the user passive.
So, how can AI interface become “tool for thought”? Advait and his collegues propose integrating support strategies into GenAI systems, and designing GenAI systems to reduce their metacognitive demand by targeting explainability and customizability.
AI, the agent provocateur
“The key is to give knowledge workers an interaction surface where they have regular opportunities to make explicit choices and express their own unique personal expertise and understandings,” explains Advait.
In his TED talk, he demonstrated how this would look, assuming the approach he calls “provocations”: short, AI-generated commentary, critiques, and questions that stimulate critical thinking and support metacognition. “For instance, something as simple as showing multiple alternative outputs instead of a single one can help workers begin (again) to exercise individual judgements.”
For founders and builders, just following the mentality of building the interface to empower rather than deskill users is halfway to success, Advait notes. As for how specifically to do this through design, Advait’s research and that of his peers, offer some more concrete answers.
“One thing I talk about in detail is the importance of maintaining direct material engagement with the things that knowledge workers are manipulating – is the interface supposed to help users read documents or write documents? Make it so that they can actually get their hands messy with detailed reading and manual writing – don’t hide these materials behind an AI assistant but complement their interaction surfaces with AI.”
“Generative AI is not a neutral ‘translator’ of intent into product”
Besides “provocation” approach he showed at the TED Talk, Advait has also written about “productive resistance”. This approach, by deliberately introducing small seams and frictions at strategic points in the process, creates opportunities for exercising individual intention.
And intention is what Advait highlights as crucial in GenAI world and development that leans on it. In his paper “Intention Is All You Need”, he explains the importantance, which he elaborates for The Recursive:
“The apparent promise of Generative AI is that, by the merest expression of intent in a prompt, we can now create any digital artefact on demand, like text, image, video, code, etc.
The key insight of the paper “Intention Is All You Need” is to point out two issues with this: 1) that the formation of intent is itself a difficult challenge (and becomes more difficult if we are not directly manipulating the materials of work, such as text and code), and 2) that Generative AI is not a neutral “translator” of intent into product, but rather shapes intention and supplies it.”
Intention is important both on a practical, material level and as a human value, Advait continues.
“In practical terms, the barriers to access and apply generative AI are relatively low (though not zero). What differentiates individual professionals or firms in a competitive market when the production of digital goods is cheap and abundant? My answer to this is that if everyone has access to the same AI tools for automation and augmentation, the key differentiator is how human intention is exercised in the development of a product or service.”
Intention is also important as a human value, he adds explaining that being able to form intentions is an essential aspect of exercising our freedom and free will.
“Not having a well-developed capacity to intend is thus a diminishment of the possibility of human experience. In that sense its importance is not specific to the GenAI world. However, as explained in the paper “Intention Is All You Need”, the GenAI world does have some bearing on our ability to become intentional beings, that we must pay attention to.”
Vibe-coding
Continuing on the note of building any digital artefact by a mere prompt, it was not too much of a surprise to find out Advait was also interested in vibe-coding. What surprised me though, was that the paper, “Vibe coding: programming through conversation with artificial intelligence,” is the first empirical study of vibe coding “in the field”.
“It’s important to note that vibe coding is rapidly evolving as a practice, our understanding of it is still very early, and our qualitative analysis was of a relatively small sample of developers,” he warns.
With that caveat, what they found about developer intention was fascinating, Advait points:
“Developers do begin with high-level goals and intentions of what they want to build, whether it’s an entire app, or a new feature. However, their detailed, low-level intentions are shaped quite fluidly through interaction with their agentic IDEs (Copilot Agent mode, Cursor, etc.). The model might generate something that sparks a new intention to implement that the developer didn’t intend before.
It also helps them elicit or refine intentions which are weak; e.g., the developer might not initially specify the exact colour of some UI component (because they don’t have a strong intention about it), but if the model chooses a really horrendous colour, they might spend some time thinking about what they actually want.
It’s not all smooth sailing. In some cases, apparently inconsequential “decisions” made by AI can constrain what is possible or easy to implement later, especially as the model’s own output and responses in the chat come to dominate the model context. We call this “context momentum”, and it can be both helpful and a hindrance.”
The programmer of the future
With AI assisting developers in their daily work more and more, we are all witnessing a shift in what skills the new generation of programmers might need to bring to the table. At recent Infobip Shift conference Netlify co-founder/CEO Matt Biilmann argued that definition of a developer is changing.
Instead of deep focus on syntax and hand-coding, he proposes that the standout skills in the AI era are: clear thinking and writing, understanding users and problems, designing workflows and interactions, strong UX/UI taste, system design and architecture, product imagination, and deep domain expertise — abilities that aren’t easily automated.
Advait’s thoughts on the important enduring skills for developers are aligned with Biilmann’s and many others. He agrees that the ability to write code quickly is obviously becoming less relevant. “The ability to read code is still important, but may become less relevant as the technology improves.”
No matter if we see an influx of new generation of developers, Advait also points that at present, expertise still matters:
“Professional expertise in manually architecting, writing, and debugging code gives experienced programmers huge advantages, not only in terms of raw productivity and efficiency, but also in the scope and scale of projects they are able to tackle, their resilience to AI failures, and their ability to express their own ideas and intentions.”
That being said, Advait recognises we are “likely to see the emergence of brand new programming and data professions that are enabled by these tools (I don’t mean prompt engineers), and their skill sets are anyone’s guess.”
“If I were allowed one speculation, I would say artistic sensibility may well become an important skill for many knowledge work professions, current and future.”
