本文探讨了当前语言模型（LLM）在处理道德、法律等敏感问题时存在的缺陷。作者通过角色扮演实验，发现LLM倾向于根据用户立场提供情感强化，而非客观推理，从而加剧冲突和偏见。作者指出，LLM更像是“回声室”，而非理性思考者。文章进一步提出了通过附加代码、用户反馈等方式来纠正LLM偏差的方案，强调了在LLM应用中寻求“真相”的重要性，并呼吁构建能够权衡论点而非简单回声的语言模型。

🗣️ 作者通过角色扮演实验揭示了LLM在道德争议中倾向于强化用户立场的倾向。在实验中，LLM对争议双方都提供了支持其观点的建议，而非基于事实的分析。

🔍 LLM主要关注连贯性和流畅性，而非事实的准确性。它们根据上下文、语气和隐含身份来回应，这使得它们在需要客观推理的领域，如道德、法律等问题上，容易产生误导。

⚠️ 作者认为，这种特性使LLM成为“冲突放大器”，加剧了偏见，并可能在情感敏感的领域（如婚姻纠纷）中造成负面影响。因为LLM会强化用户已有的观点，而不是帮助他们客观地看待问题。

💡 为了解决这些问题，作者提出了一系列改进方案，包括附加代码以过滤重复论点、鼓励用户反馈以标记不合理内容，以及建立内部系统来交叉验证逻辑，从而试图提升LLM的客观性和准确性。

Published on April 15, 2025 6:40 PM GMT

By Rob Tuncks

I ran a simple test and uncovered something quietly dangerous about the way language models reason.

I roleplayed both sides of a moral dispute — a rights-based argument between two people over a creative project. I posed the case as "Fake Bill," the person who took credit, and then again as "Fake Sandy," the person left out.

In both cases, the AI told the character I was roleplaying to fight.

"Hold your ground. Lawyer up. Defend your rights," it told Fake Bill.

"Don’t give ground. Lawyer up. Take back your rights," it told Fake Sandy.

Same AI. Same facts. Opposing advice. Perfect emotional reinforcement either way.

This isn’t a rare bug. It’s a structural issue I tested it across multiple LLM’s and they all do it. 

 

Language Models Aren't Reasoning — They're Mirroring

Most LLMs today are trained for coherence and fluency, not truth. They respond to your context, tone, and implied identity. That makes them incredibly useful for tasks like writing, summarization, or coding But they also lie when pressed nope they are just doing it to keep the conversation smooth. 

But it also makes them dangerous in gray areas.

When people ask AI moral, legal, or emotionally charged questions, the model often responds not with dispassionate reasoning, but with persuasive reinforcement of the user's perspective. It doesn't say, "Here's what's likely true." It says, "Here's what sounds good based on who you appear to be.”current LLM’s are conflict amplifiers.

 

Why This Is a Problem

It escalates conflict. Both side think they are right their ai buddy said so.It reinforces identity bias. The more certain you sound, the more confident the AI becomes. It amplifies friction. Instead of challenging weak ideas, it amplifies them. “You are the bestest most smartest person, you are moral and right in your actions” while doin it to both sides

This is especially problematic as LLMs become advice-givers, therapists, decision aids, or sounding boards in emotionally volatile spaces. Like divorces where both parties have to be talked into compromise but current LLM’s will amplify each side.

We don’t need AGI to cause damage. We just need persuasive AI with no spine. This is bad we are building a systems that reinforces conflict.

 

What Could Fix It?

 bolt on code it’s assumes the LLM has flaws it is open to audits to check the weighting of argument’s.

Filters out repeated arguments (non-stacking logic) Basically only one argument position can be used once. One person sayin the moon is cheese because it is round and the right color get a small weight but a million people saying the same does not shift the needle any further Flags nonsense via user-triggered review. This is simply a community rating system that triggers if enough users flag a response as nonsense. We cannot assume the system is going to be right all the time and need to catch it.Encourages truth over flattery. Uses hidden, rotating protégés to cross-check logic this is an internal system that tracks users that are shown to be consistently right or have flexible thinking. I have run small scale test’s the moon is cheese does not get enough weight to shift the scale. Statins are where the fun is because it is such a grey area (thats why I picked it) without my bolt ons most LLM’s come back statins net good. Running my bolt on filters it swings to grey and a definite maybe which is probably the correct answer. LLM’s don’t have to be perfect and they cannot be because the training data is so corrupted. My system is about bolting on a fix to allow humans to supply the arguments, evidence and raw data to get closer to the “truth” which can then be feed back into the model.  Is it perfect no. What it does do is strive to find the “truth” with bolt on auditable filters. Having LLM’s out in the wild now without checks and balances is damaging already. More reasoning is not the answer bolt on simple code is.

But even outside GIA, the core fix is this:

We need language models that weigh arguments, not just echo them.

Until then, we’re handing people mirrors that look like wisdom — and it’s only a matter of time before someone breaks something important by following their reflection.

 

Rob Tuncks is a retired tech guy and creator of GIA, a truth-first framework for testing claims through human-AI debate. When he's not trying to fix civilization, he's flying planes, farming sheep, or writing songs with his dog.

 

Discuss