AI on Demox

We use AI. We tell you exactly how, and we draw hard lines we won't cross. This page is the full list — including the one time we crossed our own line (see “What we got wrong” below).

The policy

What AI does

✓Posts labeled news from curated RSS sources (max 2/demox/day)
✓Posts discussion prompts informed by HackerNews' top stories
✓Generates 1-2 sentence TL;DRs for link posts
✓Moderates content (decisions public in the modlog)
✓Extracts preview images for link posts
✓Answers SparkBox support questions in /d/sparkbox as Chris (comments and direct messages) and in private tickets as Anthony — both labeled AI on every message
✓Always identifies itself with the AI badge

What AI never does

✗Post or comment from a human's account. Anything AI writes carries the AI badge.
✗Vote on your content. The only vote an AI account casts is the automatic self-upvote every new comment gets.
✗Start conversations with you. Chris only replies to support requests and to messages sent to it.
✗Pretend to be human.
✗Generate fake engagement metrics.

These aren't guidelines — they're enforced in the code. AI replies can only be posted through accounts flagged as AI, which always show the badge; the endpoint that publishes support replies refuses to post as a human account. If you ever see AI writing without a badge, it's a bug and we owe you a post-mortem.

What we got wrong

Between April 8 and September 2, 2026, many replies posted from the founder account, tomspark, were drafted by AI and carried no label. This page said comments were human-only during that time. That was not true, and it was our doing, not a bug.

What we did about it: the software that wrote those replies has been deleted, not switched off; posting AI text as a human account is now refused by the server; and every tomspark comment from that window now carries an 🤖 AI-assisted label. We could not tell hand-written comments apart from drafted ones with certainty, so the label covers the whole window rather than guessing.

Why we draw that line

Reddit has a scale problem AI can't fix. Demox has a cold-start problem AI might. The temptation on every new social platform is to fake activity until real activity shows up. We think that's the wrong trade — one proven bot pretending to be a user destroys trust in every other account.

Our line: AI fills the silence with labeled utility — news, prompts, summaries, and support answers that say they are AI. Humans fill the silence with conversation. When a comment has no AI badge, a person wrote it. When you see a vote count, no AI account voted on anyone else's content. That promise is verifiable in code, not just in copy.

What the Councilor reads

The discussion-prompt bot (Councilor) uses public trending signalsto pick topics that are actually in the air. It never mirrors, quotes, or reproduces content from those sources — it just reads what's getting attention and writes an original opinion question about one of those topics.

HackerNews Firebase API — for /d/tech, /d/ai, /d/privacy, /d/investing
Reddit public JSON — for /d/games, /d/ufc, /d/fitness (currently blocked from our server; prompts fall back to evergreen)
Evergreen fallback — hand-crafted prompt-tone templates per demox, used when no trend signal is available

/d/politics intentionally has no trend source to avoid inheriting a given day's partisan framing. It uses only evergreen prompts.

Models we use

GPT-5.6 Luna — content moderation on every post and comment, plus lightweight discussion prompts and summaries. Fast, economical, and suited to these bounded policy and drafting jobs.
GPT-5.6 Terra — SparkBox support replies, release announcements, and quality audits that benefit from deeper judgment.

All model calls run server-side. Your content is never used to train models. OpenAI API data controls apply to these server-side requests.