Seven things our AI will never do

A few months ago I was watching a demo of a very polished AI sales tool.

The founder was walking through a live roleplay. The prospect on the other side of the screen mentioned, almost in passing, that her team was going through layoffs. The AI, in a sidebar, suggested the rep "use the organizational uncertainty as a reason to lock in a longer term now."

Everyone in the room nodded. It was a clever move. It was also, if you slowed down for one second and looked at it, a suggestion to exploit a person who had just told a stranger her coworkers were losing their jobs.

I decided that night that whatever we shipped, I wanted the lines written down before anyone paid us a dollar.

We're in private beta. We have no enterprise customers asking for a security review. No regulator is knocking. The EU AI Act doesn't apply to a product that isn't generally available yet. By any normal startup logic, publishing a public ethics page before launch is a waste of a Thursday.

I did it anyway, for two reasons. The first is selfish. Written-down commitments are a forcing function on the code. Once a promise lives on a public URL, it's harder for a future me, tired, under revenue pressure, three weeks from a renewal, to quietly ship the feature that contradicts it. I'm writing the rules now because I don't trust the version of me that's running out of runway.

The second is that most of the harm from meeting AI is going to come from products whose teams never asked the question out loud. Not from bad actors. From ambitious teams who built the clever thing because the clever thing was in reach and nobody paused the demo.

The full page lives at intelliprompter.com/ai-ethics. Each pledge is paired with a specific architectural proof point. Here I just want to say, in my own words, why each one is on the list.

1. We will never coach you to lie. Not timelines, not authority you don't hold, not offers, not facts. The reason this is first is that it's the easiest line to cross and the hardest to walk back. If I let the coach help you tell a customer the deploy is on track when it isn't, I've built a weapon, not a product.

2. We will never exploit the counterparty's vulnerabilities. Grief, financial distress, fatigue, a language barrier. The sales-tool demo I opened with crossed this line. I don't want to compete on who can best weaponize an off-hand mention of someone's hard week.

3. We will never generate covert persuasion. No fake scarcity, no false anchors, no manufactured urgency. The FTC wrote a list of dark patterns. I turned it into a list of banned outputs. The honest version of a good argument is almost always available if you're willing to look for it.

4. We will never coach based on protected-class inferences. Race, gender, disability, religion. The coach doesn't infer any of these from a voice or a name because the inference model simply isn't in the pipeline. The best way to avoid misusing a signal is to never collect it.

5. We will never run without consent where the law requires it. If your account is set to California, Illinois, Washington, or any other all-party-consent state, the app gates every meeting behind an explicit consent step. This is the one most tempting to fudge for UX reasons. We don't.

6. We will never train on your meeting data. Your transcript stays on your device. Your counterparty's speech is never used for training, ever. The Otter class action turned on exactly this, and the right lesson from it was architectural, not legal. We don't have a server-side transcript store to misuse.

7. We will never act autonomously in a meeting. No auto-send, no auto-speak, no auto-accept. Every suggestion needs your keypress to leave the overlay. The coaching engine has no egress channel at all. Not a disabled toggle, physically absent. I don't trust myself to never ship the agent mode if the plumbing is sitting there.

Pledges five and seven both have real product costs. Consent gating adds friction to the first run. The no-autonomous-action rule rules out a whole category of auto-follow-up and auto-schedule features competitors already ship. I've had early users ask for both.

I think that's the right tradeoff and I know not everyone will. If you're evaluating meeting AI tools and auto-send is your top feature, we won't be the right fit. I'd rather tell you now than lose your trust later.

Every coaching card we generate is auditable in your meeting review, timestamped, and rate-able as helpful, neutral, or harmful. Every refusal the coach makes is counted and shown to you in Settings. If you ever see one of these seven lines crossed, write to ethics@intelliprompter.com. I read every one of those. We publish a review of every reported failure and how we fixed it.

The full page, with the architectural proof for each pledge, is at intelliprompter.com/ai-ethics. It's a living commitment. It will grow as the product grows. It will never shrink.

Jeff