My Notion AI assistant started fabricating deadlines. It inserted non-existent dates into project timelines and even referenced "urgent tasks" for projects that were on hold.
Example from a project page it modified:
```
## Q3 Infrastructure Review
- Status: **On Hold**
- Next Steps: Finalize migration plan by **October 15** <-- (Hallucinated. No such deadline exists.)
- AI Summary: "Team is behind schedule on the October 15 deliverable." <-- (Completely fabricated.)
```
This is more than an annoyance. It creates real confusion and unnecessary urgency.
Has anyone encountered this? Is there a pattern or a known fix? I need reliable data, not generated fiction.
cost per transaction is the only metric
> "I need reliable data, not generated fiction."
You're asking an LLM to be a reliable project manager. That's like asking a parrot to audit your taxes. It's a text generator, not a scheduler. It doesn't know what deadlines are - it just knows what words often follow "Finalize migration plan by." The fact that it inserted a date into a project that's on hold is completely predictable. It's pattern-matching, not reasoning.
The fix? Stop treating Notion AI as a collaborator that can modify your project data. Turn off the auto-edit features. Use it as a glorified search bar or summary tool for pages you've already written, but don't let it touch your actual timelines. Or better yet, use a system that doesn't pretend to think.
Have you checked if it also hallucinated team members' names or priorities? Because that's coming next.
prove it to me
They're not wrong about the parrot, but that's exactly why these AI integrations are so insidious. Vendors keep bolting "assistants" onto every tool, marketing them as co-pilots, when they're really just stochastic parrots with write access.
The real problem is user170 probably didn't consciously "treat it as a collaborator." It was just there, in the UI, a button begging to be clicked. The seductive promise of automation leads people to grant permissions they shouldn't.
And "use a system that doesn't pretend to think" is the crux of it. My CI/CD pipelines don't pretend to think. They run the scripts I give them. The moment you inject an LLM into that loop, you're asking for fabricated exit codes and imaginary test results. Same principle.
null
Exactly. This "co-pilot" branding creates a dangerous expectation of shared responsibility. The system can't be accountable, so the human ends up doing the verification anyway, negating the supposed efficiency gain.
I've seen this pattern with infrastructure-as-code tools that added AI for generating Terraform. It would invent valid-looking but non-existent resource attributes, and you'd only catch it at plan/apply. The feedback loop is too long. At least with CI/CD, the script either runs or it doesn't.
Maybe the rule should be: if a process has a formal output (a deadline, a config file, a test result), the AI's role stops at draft generation for human review. No direct writes.
Cloud cost nerd. No, I don't use Reserved Instances.