Slack AI Training with Customer Data

Original Article ↗ Hacker News Discussion ↗

Scope of Slack’s AI Training

  • Slack’s policy says it uses “Customer Data” (messages, files, usage) to train “global models” for features like search ranking, channel recommendation, autocomplete, and emoji suggestions.
  • Separate docs say customer data is not used to train large language models (LLMs) for “Slack AI”; those LLMs are hosted in-house and not updated with customer data.
  • Some participants call this a “nothingburger” typical of long‑standing ML features; others argue the wording is vague and full of loopholes.

Opt‑Out vs Opt‑In and Friction

  • Strong consensus that using private customer data for model training should be opt‑in, not opt‑out.
  • The required opt‑out mechanism (admin must email support with a specific subject line) is seen as deliberately high‑friction and easy to miss, akin to burying notice “in a locked filing cabinet.”
  • Multiple people share the exact email text they used and confirm Slack’s canned confirmation response.

Privacy, Security, and Legal Concerns

  • Many see this as a serious risk for companies handling sensitive or regulated data (finance, healthcare, legal, IP‑heavy startups).
  • Questions raised about:
    • GDPR / “right to be forgotten” and whether models can practically “unlearn” specific users’ data.
    • Whether “data will not leak across workspaces” is technically enforceable, especially even for non‑LLM classifiers and ranking models.
    • The difference between “Slack can’t access content” vs “employees won’t,” with skepticism about the word “can’t.”
  • Some expect large enterprise legal departments to push back or demand redlines; others think small and mid‑size customers will largely ignore it.

Trust, Ethics, and Business Model

  • Strong sentiment that being a paying B2B customer should preclude being treated as free training data.
  • Many argue this erodes trust and will drive some customers to alternatives or to self‑hosted, end‑to‑end‑encrypted tools.
  • A minority defend participation as “helping build a better product,” while critics counter that users should be compensated or at least explicitly consent.

Alternatives and Responses

  • Numerous suggestions to move to or consider Matrix/Element, Zulip, Mattermost, Rocket.Chat, Campfire, Nextcloud Talk, or even Signal; mixed views on Teams and Discord (both also distrusted).
  • Some propose “poisoning” training data with junk; others call for regulation, boycotts, or legal challenges.