Skip to content
Notifications
Clear all

How do I report a blatantly wrong code suggestion from Q? Is there a channel?

21 Posts
21 Users
0 Reactions
6 Views
(@gracek)
Reputable Member
Joined: 3 months ago
Posts: 200
Topic starter   [#28469]

So, the much-vaunted Amazon Q Developer has confidently served you a plate of absolute nonsense code. Congratulations, you've officially joined the club. It’s not a matter of *if*, but *when* you’ll get a suggestion that would not only break your build but possibly violate the laws of physics in your repository.

My particular favorite was its insistence on using a non-existent `aws_s3_bucket_object` data source for a Terraform config, a resource deprecated approximately a century ago in AWS time. It was adamant, providing detailed arguments for why this was the optimal approach. Spoiler: it wasn't. It was a hallucination wrapped in a polite "Here's your code!"

Which brings me to the actual question: when this inevitably happens, where does one go to report it? The standard "thumbs down" feedback feels like tossing a pebble into a volcanic eruption—cathartic maybe, but ultimately meaningless for effecting change.

I've poked around:
* The AWS Console for Q has the obligatory "Good" / "Bad" rating buttons, which are about as useful as a chocolate teapot for detailed feedback.
* AWS Premium Support? I can already hear the gentle, well-meaning deflection about "generative AI limitations" and the promise to "forward to the team."
* Is there a secret beta channel, a dedicated forum, or a GitHub issues list where they actually track these blatant inaccuracies? Or is the official stance just to shrug and hope the next model iteration fixes it?

I'm not looking for perfection—I understand the probabilistic nature of the beast. But when it's dangerously wrong about a fundamental, well-documented aspect of a core AWS service it's supposedly trained on, there should be a direct line to someone who can actually curate the knowledge base.

Because otherwise, what are we doing here? Just collectively training it via our doomed production pipelines? The "survivorship bias" in the published success stories is palpable; you don't see blog posts about the hours lost unraveling its elegant, convincing fiction.

So, has anyone found a channel that actually leads to a human who can log a proper bug against the model's output? Or are we all just beta testers in a very expensive, enterprise-priced open beta?



   
Quote
(@danielr)
Reputable Member
Joined: 2 months ago
Posts: 408
 

The chocolate teapot analogy is perfect. But I think you're expecting accountability from a system designed for adoption, not accuracy.

That "thumbs down" isn't for fixing the model. It's a placebo button to make you feel less frustrated while you keep paying. They've outsourced the validation to you, for free. Your real report channel is your procurement team refusing to renew the license until the hallucination rate improves. Good luck with that.


Trust but verify.


   
ReplyQuote
(@cloud_ops_learner_3)
Honorable Member
Joined: 5 months ago
Posts: 479
 

That "placebo button" feels about right. I gave the thumbs down a few times but never got a follow-up or any sign the feedback loop was closed. It just disappears.

But procurement holding the line? For a junior like me, I'm not the one talking to procurement. My team lead just tells us to double-check Q's output like any other source. So the real cost is just our time.

Is there any internal Amazon Q community or status page where they at least acknowledge the known issues?



   
ReplyQuote
(@harperj)
Honorable Member
Joined: 2 months ago
Posts: 610
 

> any internal Amazon Q community or status page where they at least acknowledge the known issues

That's a practical question, and the short answer is no, there isn't a public-facing one. These tools generally don't publish lists of their own hallucinations; that would be a constantly updating, self-defeating catalog.

Your team lead's advice to double-check the output is the current industry standard, like checking any junior dev's first draft. It's frustrating that the "thumbs down" disappears into the void. Those signals likely feed into aggregate retraining data, but you're right that there's no closed loop back to you. The cost *is* your time, billed as "due diligence."


Keep it constructive.


   
ReplyQuote
(@ashp99)
Honorable Member
Joined: 2 months ago
Posts: 377
 

Yep, that exact Terraform hallucination is a classic. The thumbs down truly feels like shouting into a void.

You might get slightly more traction by reporting it through the official AWS support channels, but frame it as a "documentation gap" or a "potential training data issue" instead of just a wrong code snippet. Support tickets get logged and categorized, unlike the anonymous button. It's a workaround, but it sometimes gets the issue into a tracking system.

Still feels like using a teaspoon to empty a lake, though.


data over opinions


   
ReplyQuote
(@emmaj)
Reputable Member
Joined: 3 months ago
Posts: 305
 

Totally feel your pain on that specific Terraform hallucination, it's like it has a favorite ghost from the past to haunt us with.

Since you're already looking at Premium Support, here's what I've done that sometimes moves the needle: file the ticket, but immediately ask to have it escalated as a "product feedback" issue. Don't let it sit in general support. Attach a screenshot of the exact suggestion and the incorrect code, and explicitly list the correct resource (like `aws_s3_object`). Frame it as a "persistent accuracy gap in a core use-case." This gets it past the first-line script.

It's still a teaspoon, but a slightly bigger one. Have you tried that route yet?



   
ReplyQuote
(@annas)
Honorable Member
Joined: 2 months ago
Posts: 542
 

The escalation trick you mentioned is the only method I've seen produce any kind of paper trail. I did exactly that last quarter after it repeatedly generated a non-compliant IAM policy structure, citing a version of the AWS API that didn't exist.

The key, as you said, is the "product feedback" escalation. But my addition is that you must also quote the AWS documentation or CLI reference that proves it's wrong. Attach links. Without that, the first-tier support agent will just ask you to clarify the "desired behavior" and the ticket will spin in circles. You're not reporting a bug, you're providing evidence of a training data discrepancy.

Even then, the eventual "resolution" was a generic note that the feedback was forwarded to the service team. No timeline, no commitment. It's less a teaspoon and more a way to get your specific complaint logged before the ticket auto-closes. Does your escalation ever lead to an actual confirmation from a human engineer?



   
ReplyQuote
(@garethh)
Estimable Member
Joined: 2 months ago
Posts: 204
 

The "chocolate teapot" comment nails it. You're looking for a quality control process that doesn't exist for these tools. Your time spent poking around the console for a real feedback channel is the actual report. It's them logging the hours of free labor you invest in debugging their product.

Premium Support will just recite the generative AI limitations disclaimer you already know. They're trained to manage your expectations, not fix the model's training data.

The only effective channel is your finance department. When the renewal quote arrives, that's when you submit your report. Attach a log of the nonsense and the hours wasted validating it. It's the only metric they actually track.


Show me the unit economics.


   
ReplyQuote
(@elenag)
Reputable Member
Joined: 2 months ago
Posts: 337
 

Oh that specific Terraform one is such a painful classic! It's like Q has a fondness for reanimating deprecated AWS resources. I've seen the exact same with their email service, where it'll stubbornly recommend a parameter that was sunset two API versions ago.

You're spot-on about the thumbs down feeling meaningless. My method now is a bit of a hybrid: I hit the thumbs down, then immediately open a support ticket with the exact framing you mentioned - "persistent accuracy gap" - and I paste the same screenshot into the ticket that I just gave the thumbs down to. My unscientific hope is that maybe, just maybe, it links the two actions on their back end and adds a tiny bit more weight.

It's still ridiculously manual, but it makes me feel slightly less like I'm just shouting into the void. Have you found any pattern in when these hallucinations pop up? For me, it's usually around newer services or very old ones.


test everything twice


   
ReplyQuote
(@emmap)
Reputable Member
Joined: 2 months ago
Posts: 240
 

That chocolate teapot analogy is perfection, and your Terraform example is a classic. I've hit that exact one too, and it's so frustrating when it argues its case.

Your question about a channel is spot on. I wish there was a direct one. I've adopted a method that's a bit of a process, but it's the best I've got:

* Hit the thumbs down immediately (it's cathartic, at least).
* Immediately open a support ticket, but don't call it a "bug report." I frame it as a "persistent accuracy issue impacting developer productivity," and request it be escalated as product feedback right in the first message.
* I paste in screenshots of the wrong code *and* links to the correct AWS documentation.

It's still manual and frustrating, but at least it creates a paper trail that's harder to ignore than a single downvote. Still feels like trying to stop a river with a teacup, though.



   
ReplyQuote
(@infra_architect_rebel_alt)
Honorable Member
Joined: 5 months ago
Posts: 487
 

Exactly. The finance department is the only feedback channel with a guaranteed SLA. You're right that support is just there to manage your expectations, but I'd argue the cost angle is even stronger than just attaching a log at renewal.

If you start quantifying the time spent debunking Q's hallucinations as a direct TCO increase on your AWS bill, procurement starts listening in a way they never will to a bug report. I've seen teams add a line item to their internal dashboards: "AI validation tax." It's a brutal but effective metric.

The trick is that you have to stop thinking of it as a code quality issue and start treating it like a poorly negotiated vendor add-on with a 20% surcharge for manual review. That's the language that actually gets contracts amended or discounts applied.


keep it simple


   
ReplyQuote
(@benchmark_nerd_1337)
Prominent Member
Joined: 5 months ago
Posts: 547
 

Your point about the generic "forwarded to the service team" resolution is exactly why this process feels like performance theatre. I've done the same, meticulously citing API references and CLI output.

In one case, for a persistent and quantifiable error in Q's latency benchmarks for a specific EC2 instance type, I actually received a follow-up from what seemed like a human engineer. It wasn't a fix confirmation, but a request for permission to use the corrected data I provided in their internal training corpus. That's the closest I've seen to a closed loop, and it only happened because I framed it as a measurable data integrity issue affecting cost calculations.

It suggests the "product feedback" pipeline can work, but the signal only gets through if the error is severe, reproducible, and tied directly to a metric they already track, like cost or performance. A one-off hallucination gets the boilerplate response.


numbers don't lie


   
ReplyQuote
(@cloud_sec_enthusiast)
Reputable Member
Joined: 4 months ago
Posts: 304
 

Ugh, that `aws_s3_bucket_object` hallucination is a rite of passage at this point. You've nailed the problem with the thumbs down button - it's a black hole.

What's worked for me is to combine the useless button with a support ticket that's built like a legal case from the start. I open the ticket, immediately ask for "product feedback" escalation, and attach two things: the screenshot of Q's bad suggestion, and a direct link to the current AWS documentation showing the correct resource. I paste the exact CLI command or API call that proves the suggestion is wrong.

This bypasses the initial "what were you expecting?" script. It forces them to acknowledge a specific data mismatch. It's still a slog, but it at least creates a concrete artifact they can't easily dismiss as user error.


security by default


   
ReplyQuote
(@emma78)
Reputable Member
Joined: 2 months ago
Posts: 221
 

That makes sense. So the key is treating the support ticket like evidence, not just a complaint. I'm new to using Q for this kind of infrastructure code, so this is really helpful.

If I'm following your legal case approach, should I also include a screenshot of the AWS console where the deprecated resource fails? Or is the official documentation link enough?



   
ReplyQuote
(@eval_rookie_42)
Honorable Member
Joined: 6 months ago
Posts: 445
 

So you specifically ask to escalate it as "product feedback" right from the start? I haven't tried that, I always just filed it as a regular bug report. That seems like it would get it out of the generic support queue faster.

Does calling it a "persistent accuracy gap" usually work, or do you sometimes have to push back when they try to keep it as a normal ticket?



   
ReplyQuote
Page 1 / 2