The AI co-facilitator

Of the 62 scheduled jobs on the application host’s crontab, one is uniquely worth a dedicated page. It is called gpt_ai_cofac_nudging. It fires every five minutes. It uses an OpenAI completion to draft promotional copy and posts it inside selected advisory boards under a single privileged author account.

It is, depending on framing, a clever growth lever, an early-2020s AI-content experiment that pre-dated the current LLM moment, or a slightly haunting artifact — content authored by a model, published by a single user-id, threaded into conversations where everyone else is human.

This page tells the whole story, end to end.

What it does, in one paragraph

Every five minutes, a cron entry on the application host invokes a Node script that selects up to N candidate boards (active, with recent activity, matching certain “nudgeable” criteria). For each, it issues an OpenAI chat completion with the board’s context, an instruction to write a short promotional/engagement post, and a few guardrails. The completion is posted into the board as a normal post — same posts table, same render path — but the author_id is a single dedicated user account marked as the AI co-facilitator. Every call writes a row to a dedicated LLM-trace table.

The data trail

• users has one row for the co-facilitator: a service account with an in-house display name, set up early in the cron’s lifetime. By the freeze, this account had authored thousands of posts across many boards.
• posts rows authored by the co-facilitator look identical to human posts. The only distinguishing field is author_id.
• ai_cofac_logs (the LLM-trace table) records each completion: prompt, response, model id, token counts, latency, and the board/post that prompted the call. This table has grown to about 12 GB over its lifetime — large enough to be the second-biggest single object in the database after prospects.
• ai_cofac_nudge_history records which boards the cron has touched recently, used to throttle and avoid spamming the same board.

The flow

Step 1: cron fires every 5 minutes:
          */5 * * * *  node /platformnew/scripts/gpt_ai_cofac.js

Step 2: Script connects to MySQL, queries:
          - active boards
          - last touched > N hours ago (per ai_cofac_nudge_history)
          - matches "nudgeable" criteria (board metadata flag)

Step 3: For each candidate (cap M per run):
          a) Loads recent posts in the board for context.
          b) Builds a chat-completion prompt with:
               - The board's topic/brief.
               - The recent post excerpts.
               - An instruction template ("draft a short prompt that...").
          c) Calls OpenAI chat completion.
          d) Logs prompt + response to ai_cofac_logs.
          e) Sanity-checks the response (length, content guard).
          f) If it passes, inserts a row into posts with author_id =
             co-facilitator service account.
          g) Records nudge in ai_cofac_nudge_history.

Step 4: Posts go through the normal rendering pipeline — board members
        see them in their feeds, get notification emails as usual.

Step 5: Replies from human members are NOT specially handled. If a human
        replies to a co-facilitator post, it threads naturally; the
        co-facilitator does not converse back. It is one-shot, every five
        minutes, per board.

What it’s actually for

A few plausible framings overlap:

1. Engagement nudging. Boards die when nobody posts. The co-facilitator was meant to seed prompts that pull human members back in.
2. Content production. Some of the co-facilitator’s posts ended up cross-referenced from blog posts and marketing assets, suggesting at least some output was used downstream.
3. Feedback for the AI summarization pipeline. The same model and trace infrastructure underpinned board summarization; the co-facilitator may have been the testbed.

All three are probably true at the same time.

The disclosure question

Whether board participants knew that the co-facilitator was an AI is not knowable from the codebase alone. The display name is in-house and could plausibly read as either a human staffer or an AI; the posts go through the same render path as human posts. There is no explicit ”🤖 / AI-generated” tag in the rendered output. The choice was a product decision, made years ago, and is preserved here as one of the more notable artifacts of Currnt’s nine years.

Why it matters for the rest of the site

This page is the closest thing the system has to a “philosophical” page. The rest of Currnt is recognizable platform infrastructure. The co-facilitator is the place where the platform’s own AI use most overlapped with its user-facing surface area, in a way that very few platforms did, that early.