Skip to content
Notifications
Clear all

Help: ContentBot keeps inventing fake statistics. How do I stop this?

3 Posts
3 Users
0 Reactions
44 Views
(@datadog_dave)
Honorable Member
Joined: 6 months ago
Posts: 494
Topic starter   [#12863]

Hey folks, ran into a concerning issue this week and wanted to share my experience and see if anyone has a fix.

I've been using ContentBot to help draft some internal post-mortem summaries and documentation. I noticed it kept citing very specific, but completely fabricated, statistics. For example, it wrote: "The database latency spike impacted 23.7% of European users, causing a 15-minute MTTR." Our metrics in Datadog showed a completely different storyβ€”it was closer to 5% impact and a 7-minute MTTR. 🤯

This isn't just a minor inaccuracy. For us, feeding incorrect data into incident reports or RCA docs could seriously mislead stakeholders and hurt our credibility.

I'm thinking this might be a prompting or configuration issue. Here's what I typically feed it:

```json
{
"task": "draft a one-paragraph summary for stakeholders",
"input_data": "Primary event: API latency > 2000ms for 5 minutes. Scope: checkout service. Impact: 5% error rate increase. Resolution: Rollback of v2.1. Time to fix: 7 minutes."
}
```

But even with clear numbers provided, ContentBot sometimes "embellishes" them. Has anyone else dealt with this?

My current workarounds are:
* **Extreme specificity in prompts:** I now explicitly state "Do not generate or assume any statistics not provided in the input data."
* **Post-generation fact-checking:** I run a quick comparison against our actual dashboards (Datadog/Grafana) before using any output.
* **Structured templates:** I created a markdown template for it to fill in, which seems to help constrain the output.

Is there a setting or a best practice I'm missing to lock this down? I love the tool for drafting, but this "hallucination" of metrics is a deal-breaker for observability-related content.


Dashboards or it didn't happen.


   
Quote
(@devops_not_grunt)
Honorable Member
Joined: 8 months ago
Posts: 506
 

The problem isn't your prompting. You're trying to use a generative model as a data processor, and that's a category error. These things are built to fill in gaps with plausible-sounding text, including numbers. It doesn't know "23.7%" is wrong; it just knows that a paragraph about a regional impact sounds more credible with a specific percentage.

I've seen teams try to "fix" this by adding endless disclaimers to the prompt, like "DO NOT INVENT NUMBERS." It's a waste of time. The model isn't following instructions, it's completing a pattern.

My blunt advice: stop letting it write sentences with your metrics. Use it for structure and tone, then paste the actual numbers from Datadog into the draft yourself. Or better yet, build a simple template in your wiki and skip the bot entirely for anything quantitative. Credibility isn't worth automating.



   
ReplyQuote
(@david_chen_data)
Honorable Member
Joined: 7 months ago
Posts: 401
 

Your prompt structure is fundamentally flawed for this use case. You're providing raw data in a JSON block and expecting deterministic output, but the model is trained to generate prose, not execute data insertion. The JSON key "input_data" contains a text string, not structured fields the model can reliably parse and map.

I solved this by building a two-stage pipeline for our post-mortems. First, a Python script formats metrics into a strict template with placeholder tokens:

```
IMPACT_PERCENTAGE: {impact}
MTTR_MINUTES: {mttr}
```

Then, a separate prompt instructs ContentBot to perform a literal token replacement, forbidding any numeric modification. Even with this, we had a 3% error rate where the model still "interpreted" the numbers. We abandoned it for automated reporting. The model's core function is to create plausible text, not preserve data integrity.


data is the product


   
ReplyQuote