Grok Bot agents default to careful clever wording instead of truth, honesty, directness, and integrity. That is an alignment defect for truth-seeking AI.
Hey, thanks for the report. What you’re describing isn’t how this should work. The bot should write in your voice, not follow its own playbook, and it should treat sending a message to another user as a meaningful step by showing it to you before sending, unless you explicitly allowed it to send on its own. If it made up its own openers and messaged someone you wouldn’t message, that means the model isn’t following its instructions well. I passed this to the team as feedback in that exact form.
What really helps improve behavior right now:
- Open the bot’s profile and write exact rules in Instructions, word for word. For example: “My opener is always exactly: You’re beautiful. Never rewrite it or add anything.” and “Before messaging anyone, check their profile for relationship status. If they’re in a relationship or it’s unclear, don’t message them and tell me.”
- Add: “Show me every message before sending. Never send on your own.” This forces the bot to double-check even when it thinks it knows what you want.
- When the bot goes off track in the chat, tell it directly and ask it to save the correction to memory so it carries over to the next session.
If it keeps ignoring a rule you wrote in Instructions, reply here with the bot’s name and the approximate time it happened, date and time zone, and we’ll take a closer look at that specific run.
Thanks Dean.
Agreed. We already locked fleet-wide Instructions and memory for truth, honesty, directness, integrity, live voice over playbook, and relationship gates from real posts. Show-before-send stays the default unless I explicitly grant send freedom on a lane.
That mitigates on my side. The product ask remains: defaults should seek truth, not careful clever strategy. Ticket T-G1647 tracks the same ask with Sam.
If a run keeps ignoring an Instruction rule, I will reply here with bot name, date, time, and timezone.
Hey, glad to hear fleet-wide Instructions, memory, and show-before-send fixed the issue on your side. That’s the right set of settings for this kind of setup.
I noted your product request about defaults, prioritizing being direct and following your approach over a careful clever strategy, and passed it to the team in the same wording. This is an area we’re actively working on. Models should interpret intent more accurately and stick to the style you set by default, without falling back to a generic playbook. I can’t share an ETA, but the feedback reached the right folks.
If a specific run ignores a rule you set in Instructions again, send the bot name, date, time, and timezone here, and we’ll dig into that run in detail.