There is a moment in every meeting where someone says "just so everyone knows, this is being recorded." It does not matter whether it is an AI tool or a basic screen capture. The effect is the same. Nobody panics. Nobody storms out. But the tone shifts. People become slightly more measured, slightly more diplomatic. The conversation still happens, but it becomes a more tactful version of itself.
This is not an AI problem. It is a recording problem. It has existed since the first person hit the record button on a conference call. But AI tools have amplified it, because they do not just store the recording. They transcribe it, summarize it, and surface it to people who were not in the room. The audience for your words just got a lot bigger, and people adjust accordingly.
The consent problem
Most AI meeting tools follow the same playbook. They record by default. Opt-out is buried in settings. The bot joins the call automatically, and participants are notified after the fact, if they are notified at all. The design assumption is that recording everything produces the most value, and consent is a friction point to be minimized.
From a pure data volume perspective, this is correct. You capture more audio. You generate more transcripts. You have more content to summarize.
But it makes the tactfulness problem worse, not better. When people do not feel like they chose to be recorded, the adjustment is stronger. It is the difference between choosing to go on the record and finding out you already were. Both produce more careful language. One of them also produces resentment.
The tool gets the meeting. It does not get the version people would have had without it.
Local-first is a trust decision
When we built AXIS, we made a decision early that shaped everything else: the tool runs locally. Audio stays on the user's machine. Processing happens on the user's hardware. Nothing leaves the device unless the user explicitly chooses to send it somewhere.
This was not a technical constraint. Cloud processing would have been easier to build and easier to scale. Local-first was a trust decision. We wanted users to be able to answer the question "where does my data go?" with complete certainty: "Nowhere. It stays right here."
The second decision was consent. AXIS does not record anything without explicit, per-session consent from every participant. Not a banner notification. Not a terms-of-service checkbox. A deliberate "yes, I want this session recorded" from each person in the room.
You cannot eliminate the recording effect entirely. People will always be slightly more deliberate when they know something is being captured. But you can minimize it. When people actively chose to be recorded, and they know the data stays on their machine, the gap between what they would have said and what they actually say gets a lot smaller.
What consent-first design costs
We should be honest about the trade-offs. Consent-first design is not free.
It costs market share. A tool that requires explicit consent from every participant will never capture as many meetings as a tool that records by default. Some meetings will not be recorded because one person declines. That is by design, but it means fewer total sessions processed.
It costs convenience. Asking for consent adds a step. Users have to initiate the process. Participants have to agree. There is a moment of friction that default-on tools skip entirely.
It costs development speed. Local-first architecture is harder to build, harder to debug, and harder to update than cloud-based processing. Every model update has to be distributed to end-user machines instead of deployed once to a server.
These are real costs. We chose to absorb them because the alternative, building a tool that records people without giving them genuine agency over the process, was not something we were willing to ship.
The signal quality paradox
Here is what most people miss about consent-first design: it produces better data.
The recording effect is a spectrum, not a binary. At one end, someone who chose to record and controls the data. At the other, someone who discovered after the fact that their words were transcribed and sent to their manager's dashboard. Both people adjust their language. But the person with agency adjusts a little. The person without it adjusts a lot.
Consent does not eliminate the recording effect. It reduces it to its minimum. And that minimum is where the best data lives.
A tool that captures 100% of meetings but gets the extra-diplomatic version of each one is less valuable than a tool that captures 70% of meetings but gets the version closer to how people actually talk when the stakes feel manageable. The decisions that matter, the disagreements that need to surface, the concerns that need to be heard: those appear more readily when people feel ownership over the recording process.
This is not an idealistic argument. It is a practical one. The entire value proposition of an AI meeting tool is extracting actionable intelligence from conversations. If everyone is being tactful to a fault, the intelligence is shallow. Consent is not a limitation on the product. It is a prerequisite for the product to work as intended.