Responsible AI

What we mean by "AI"

Aloha uses machine learning in a few specific, disclosed places:

• Muse (voice model) — drafts posts in your cadence for the Composer, on the channels you've switched it on for.
• AI companion — single-shot write / refine / suggest actions available in Basic; not trained on your writing.
• Channel rewrites — converts one draft to native variants per channel (length, format, tone).
• Inbox triage — classifies incoming messages into question / praise / needs-review buckets.
• Insight summaries — plain-English weekly notes in Analytics.
• Best-time suggestions — statistical model, not generative AI.

We don't use AI to secretly automate your replies, to generate fake engagement, or to impersonate another creator's voice.

What trains on your data

Your voice model does. It trains on the specific posts you mark as "sounds like me" in Settings → Voice. You can remove any post from the training set at any time; the model rebuilds on the next save (a few seconds).

Your inbox triage does. It learns from the messages you classify or reclassify — "this is actually a question", "this is a low-touch thank-you". Learning is workspace-scoped.

Nothing trained on your data is shared with other customers. Nothing trained on your data is shared with any external foundation-model provider.

What never trains on your data

• Public AI models. Your posts are never submitted to any third-party foundation-model provider for use as training data. The inference endpoints we rely on are governed by API terms that prohibit prompts being used to train the provider's public models.
• Other customers' voice models. Your data stays in your workspace. Even a shared voice model on a Team plan only trains on the posts that workspace marks as voice-matching.
• Our marketing. Examples in our marketing and in this documentation are synthetic or used with explicit creator permission.

Where the models run

• Your voice profile — trained and stored by Aloha, workspace-scoped, hosted in AWS us-east-1. It captures tone, cadence, and style from the material you provide; it does not itself generate text.
• Inference — third-party foundation-model endpoints we route to behind our own service, governed by API terms that prohibit training on your prompts. Providers are named in our Data Processing Addendum and may change; the protections above apply to any provider in the chain. Prompts are scrubbed of identifying metadata before transmission.

Your controls

• Delete your voice model — Settings → Voice → Delete. Hard delete; training data and embeddings purged within 24 hours.
• Opt out of third-party inference — Settings → Voice → Use local model only.
• Opt out of analytics — Settings → Privacy → Analytics. Stops Aloha learning about which drafts you liked.
• Opt out of inbox classification — Settings → Inbox → Triage off. The inbox stays unsorted; no ML runs on it.

Human-in-the-loop defaults

Everything the AI drafts is a draft. Aloha never:

• Auto-publishes a post without an approval step you've set up.
• Auto-replies to DMs or comments without a per-matrix approval rule you've explicitly enabled.
• Sends bulk outbound messages on your behalf with AI-generated text, without your thumb on every send.

The defaults are deliberately strict. Automation is opt-in, not opt-out.

Transparency

• Diffs. Every AI-drafted rewrite shows you what changed against the source — you can see exactly which words and line breaks came from the model.
• Voice-match score. Posted next to every draft. If the model isn't confident, we tell you; we don't sneak generic outputs past a low-confidence gate.
• Classification rationale. Inbox triage decisions include a one-line reason ("contains a question", "sender is a returning customer") so you can retrain the classifier when it's wrong.

Content we refuse to help with

The Composer won't draft:

• Impersonations of real people without their consent.
• Content designed to deceive — fake reviews, fake case studies, AI-disclaimed texts stripped of the disclaimer.
• Harassment, targeted attacks, slurs.
• Content that violates a platform's terms in ways we can detect (spam patterns, bulk cold-DM farming, engagement-bait loops).

If you push against these, the Composer will say no and explain why. The refusal isn't a chat filter — it's a trained-in decline.

Bias and error

Voice models trained on a small number of posts will pick up the patterns — good and bad — of those posts. If your training set skews in ways you don't want, the model will too. We give you the tools to see and fix this:

• Training-set visibility — see every post the model trained on, with "remove" buttons.
• Counterexamples — mark drafts as "not how I'd have said it" to downweight similar patterns.
• Rebuild — reset the model to its empty state and retrain from scratch, one click.

Environmental impact

We run inference on GPUs in AWS regions chosen for their reported carbon intensity. In 2026 we're piloting carbon-aware scheduling for non-urgent training jobs (voice-model rebuilds, best-time re-learning) to shift load toward low-carbon grid hours.

Updates

When our AI practices change materially, we'll update this page and log the change in the changelog. We treat AI policy changes as material — this document is where we hold ourselves accountable.

Contact

Questions or disagreements: hello@usealoha.app. It's a one-person project — the founder reads every message.