← sources

I tried to build an AI that can't be talked out of a correct answer — here's what actually breaks that in practice

source: reddit (reddit:artificial)published 2026-09-18 17:55ingested 2026-09-19 02:37

retained text

Every "brutally honest" AI feedback tool I've tried eventually folds. You push back, express a little frustration, or just repeat yourself, and it quietly walks the criticism back — not because you gave it new information, but because it's optimized to keep you comfortable. That's not a prompting failure, it's a structural one: a model tuned to be liked and a model tuned to be accurate will diverge over a long enough conversation.

I spent the last few weeks trying to build around that specifically, using myself as the test case (placement prep — repos, resumes, project pitches). The interesting part wasn't the "give harsh feedback" prompt, which is trivial. The interesting part was making the harshness durable under pressure. A few things I learned building it:

1. Every critique needs a citation stored as its own row, not just text in a response, or you have no way to check later whether pushback is pointing at something real.

2. "Did the user provide new evidence" turned out to be a genuinely hard classification problem — early versions treated confident-sounding pushback as evidence even when it was fabricated. Fixed by having the system re-fetch the live source and verify the claim before ever updating a verdict, rather than trusting the user's description of it.

3. The failure mode I didn't expect: retrieved context (for a memory/recall feature) started leaking its formatting and tone into new responses through pure in-context pattern-matching, even with the correct system prompt in place and never bypassed. The prompt being "right" wasn't sufficient — the examples in context mattered just as much as the instructions.

Live Link - [https://raw-take-amber.vercel.app/](https://raw-take-amber.vercel.app/)

I've deployed a working version (5 modules: repo/resume/pitch analysis, a free-form "callout" mode with cross-conversation memory, and a mode that fact-checks another AI's response against the same evidence). Not asking anyone to sign up for anything — genuinely more interested in whether people here can find a way to talk it out of a verdict it shouldn't change, or whether the "won't fold" claim actually holds against real adversarial pushback. Happy for any kind of feedback

document fetched 2026-09-19 00:19 · canonical https://reddit.com/r/artificial/comments/1wjxc27/i_tried_to_build_an_ai_that_cant_be_talked_out_of