Compliance · 5 min read
The Ethics of Automating a Personal Conversation
Automating an Instagram DM is not like automating an invoice. The person on the other end opened a private message thread because they thought they were talking to someone. Everything difficult about this follows from that one fact, and none of it is settled by the tool being compliant.
Meta's rules tell you what you are permitted to send. They do not tell you what you should send, and there is a wide gap between the two where most of the real decisions live.
Disclosure: what you actually owe
The maximalist position is that every automated message must announce itself. In practice that produces worse conversations — "🤖 This is an automated response" at the top of a reply that says your shop opens at ten is theatre, and it makes a helpful answer feel like a rejection.
The position we think holds up is narrower and harder: disclose when the person could reasonably form a false belief that matters.
Disclosure not owed
Disclosure owed
The test is not "is this a bot". It is: if this person found out later that a machine wrote it, would they feel misled about something they cared about? An answer about delivery times, no. A message saying "I'm so sorry to hear that, that sounds really hard" — yes, and that one is worse than saying nothing.
Consent, and what a comment does not authorise
Someone commenting PRICE under your post has consented to a specific thing: an answer about the price. It is a narrow permission and it is easy to over-read.
It is not consent to be added to a nurture sequence. It is not consent to be messaged next month. It is not consent for their handle to sit in a spreadsheet you sell. The platform enforces part of this for you — the twenty-four hour DM window closes and only their next message reopens it — but the platform is enforcing a timing rule, not an ethical one. Treating the window as the limit of your obligation is doing the minimum and calling it a policy.
What we store when someone triggers an automation, and what leaves the product, is a separate and worth-reading conversation: what a tool actually keeps about your commenters and the fact that running this makes you a data controller.
When a person deserves a human
Some messages should never be answered by software, and the useful version of this list is specific rather than principled.
- Anyone who is upset. Not "mildly disappointed" — actually angry, or hurt. An automated reply to an angry person reads as being handled, and it escalates reliably.
- Anything about health, money, safety, or law. If a wrong answer could cost the person more than the transaction, a template is not enough and a generated reply is worse.
- Anyone disclosing something personal. People tell small businesses surprising things. A same-second cheerful reply to a real disclosure is a specific kind of awful.
- Anyone who has already asked twice. The second identical question means the first answer did not work. A third copy of it will not either.
- Anyone asking whether they are talking to a person. Answer that one yourself, truthfully, always.
Takeover exists for exactly this: the automation stands down the moment a human replies to a thread by hand. It is the most ethically load-bearing feature in the product and it is a toggle nobody demos.
Uses we think are wrong, even where they are permitted
Being allowed is not the standard. Some of these are things the API will happily let you do.
Automating grief and celebration. Condolences and congratulations composed by a model, sent at machine speed, about events you have not personally read.
Fake scarcity in a reply. "Only two left" generated because it converts, not because two are left. This is a lie with a timestamp on it.
Answering a complaint with a marketing message. Someone says the order arrived broken and receives a discount code for their next purchase. Permitted. Contemptuous.
Automating a channel you have told people is personal. If your bio says "DM me directly, I answer everything myself", then either that stops being true or the automation stops. Pick one honestly.
Using someone's own words back at them to feel intimate. Grounding a reply in what they said is good practice. Mirroring their phrasing to manufacture rapport is manipulation with extra steps.
The question is never whether the API allows it. It is whether you would be comfortable if the recipient read your automation settings.
Voice, and the honest limit of imitation
Our voice profile is built from replies the account owner actually wrote. That is a deliberate choice: it means the output is a compression of your real register rather than an invented persona.
But imitation has a ceiling, and the ceiling is not stylistic. A profile can learn that you write short, that you use "haan ji", that you skip full stops. It cannot learn what you would think about this particular customer's situation. When the output crosses from how you write into what you would feel, the imitation has gone past what it can honestly claim. What a voice profile records, and what it cannot know goes into that boundary properly.
A short standard
Three sentences we would be willing to publish on the account itself:
Automation answers questions we have already answered before. A person answers anything with feeling in it. If you ask whether you are talking to a human, you get a human and a straight answer.
If a setup cannot survive being described to the people it is used on, the problem is the setup.
