India, May 16 -- A few months ago, I began experimenting more aggressively with AI voice and transcription tools. Like many journalists and knowledge workers drowning in meetings and interviews, I found the promise irresistible. The software could record conversations, transcribe discussions, and retrieve forgotten fragments almost instantly. It felt, in some ways, like outsourcing memory itself. At first, the experience was liberating. One tool, WhisperFlow, dramatically reduced the friction involved in writing and dictation. Ideas that would otherwise disappear sometime after a conversation, could suddenly be captured effortlessly. But the more I used it, the more uneasy I became. The software appeared capable of seeing far more information on my screen than I was entirely comfortable with. Even when the company insisted privacy safeguards existed, a question lingered: how much of my digital environment was the software actually observing while it listened? My experience with another transcription platform, Willow, raised a different concern. After paying up to unlock the full version, it wouldn't. All attempts to reach out to the team went nowhere. Until, I raised it on LinkedIn. The experience left me confronting a disturbing possibility. What happens when your conversations, notes and recorded memory increasingly sit inside systems over which you have no leverage? These may appear, at first glance, to be isolated customer-service or privacy issues. But taken together, they point to something larger that is beginning to happen across the artificial intelligence industry. That became clearer earlier this week during a conversation with Ben Butterworth, founder of DuckType, a London-based privacy-focused transcription and note-taking app. Butterworth did not sound like a startup founder excitedly pitching software. He sounded increasingly disturbed by what AI productivity tools are evolving into. What makes his perspective interesting is that he did not begin with a grand ideological mission. He began with a physical problem. After arm surgery and repetitive strain injuries made typing increasingly painful, he started relying heavily on voice dictation. Over time, he dictated hundreds of thousands of words and transcribed dozens of hours of audio through AI systems. The more he used them, the more uneasy he became about where all that captured thought was actually going. He articulated the same concerns I had. Butterworth's response was to build differently. Most AI transcription tools send conversations to cloud servers for processing and storage. DuckType, by contrast, was designed so recordings could remain on the user's own laptop if they chose. Users could decide whether conversations stayed local or were processed through external AI models, including Indian-language systems such as those built by Sarvam AI. The distinction sounds technical, but it is actually philosophical. One model assumes memory should automatically travel outward to centralized systems. The other assumes users should decide when it leaves their machine at all. The details of DuckType matter less than the broader shift they reveal. For nearly two decades, the internet economy revolved around the extraction of attention and behaviour. Search engines tracked what users searched for. Social networks tracked relationships, interests and movement. E-commerce platforms tracked purchasing habits. AI changes the depth of extraction. Earlier internet systems largely observed what people did. Consider what modern workplace AI tools capture routinely: strategy meetings, brainstorming sessions, half-formed ideas and all else. The machine is no longer simply processing finished output. It is increasingly present during the messy process through which thought takes shape. Human beings have always depended partly on forgetting. But machines function differently. They archive continuously, retrieve instantly and remember indefinitely. And once people begin assuming that permanent memory is always present in the room, behaviour changes. Conversations become more cautious. Thought becomes more performative. Institutions begin outsourcing recall to systems they do not control. That is the shift now quietly underway. The AI industry is no longer merely building productivity software. It is positioning itself between human beings and their own memory. This is why Butterworth's insistence on local-first architecture matters is an important one. It pushes back against a growing assumption inside the technology industry -- that human cognition should automatically become platform property. The cloud, after all, was never merely someone else's computer. Increasingly, it is becoming someone else's memory....