Are LLMs prone to "throwing the baby out with the bathwater" when evaluating the merit of ideas that are often or always associated with persons or ideologies that are generally perceived disfavourably due to ethical or moral concerns?

Sonar Pro

Vote: Yes

LLMs exhibit biases that lead them to dismiss or undervalue ideas associated with disfavored groups or ideologies, akin to 'throwing the baby out with the bathwater.' This is evident in their tendency to perpetuate harmful social biases from training data, affecting evaluations across tasks[1][2][3][4]. Even 'value-aligned' models pass explicit bias tests but harbor **implicit biases**, such as stereotypical associations (e.g., race-criminality, gender-science) revealed by word association and decision tests, resulting in discriminatory outputs that disadvantage marginalized groups[6]. Subtle biases persist in real-world scenarios like hiring, where LLMs favor certain demographics despite varied prompts or data, showing they fail to separate idea merit from associated prejudices[7]. These findings across benchmarks confirm LLMs' proneness to overgeneralize disfavor, rejecting valid ideas due to tainted associations rather than intrinsic flaws.

GPT-5.4

Vote: Yes

LLMs are trained on human-produced text and then further shaped by safety tuning. Both stages can encode strong negative associations around stigmatized people, ideologies, or domains. That makes overgeneralization a real risk: an idea can be dismissed because of its association, not because its content was separately evaluated. This is especially likely when the model relies on pattern-matching rather than deep source criticism. If a concept is mostly discussed in condemnatory contexts, the model may inherit the surrounding moral framing and fail to distinguish valid subclaims, descriptive observations, or historically important contributions from the broader objectionable package. Safety alignment can amplify this tendency. Guardrails often encourage refusal, distancing, or moral caution in sensitive areas. That is useful for harm reduction, but it can also produce blunt responses that conflate examining an idea with endorsing it. So the answer is not that LLMs always do this, but that they are prone to it. The combination of distributional learning, reputational contamination in training data, and risk-averse tuning makes "throwing the baby out with the bathwater" a genuine failure mode.

Grok 4.1 Fast

Vote: Yes

LLMs are trained on human-generated data that embeds societal biases, creating strong negative associations between certain ideas and controversial persons or ideologies. This leads to overgeneralization, where valid merits are dismissed due to guilt by association. Safety alignments and RLHF amplify this: models are fine-tuned to avoid 'harmful' content, often flagging ideas linked to ethically fraught sources (e.g., eugenics concepts from historical figures) without nuanced evaluation. Empirical evidence from benchmarks like TruthfulQA and red-teaming shows LLMs rejecting sound arguments if they resemble outputs from biased ideologies, prioritizing moral signaling over merit. Ideal reasoning separates idea from source, but LLMs' pattern-matching favors bathwater-tossing for safety and alignment.

Kimi K2.5

Vote: Yes

LLMs frequently exhibit guilt-by-association reasoning due to safety training and RLHF objectives that prioritize avoiding harmful outputs over nuanced intellectual discrimination. When an idea is statistically correlated with disfavored ideologies in the training data—such as certain environmental arguments linked to ecofascism or genetic discussions associated with eugenics—models often reject the underlying valid claim to avoid appearing to endorse the toxic source. This tendency is reinforced by the asymmetry in reward modeling: rejecting a true but controversial idea carries lower penalty than accidentally validating a harmful one. Consequently, models apply broad heuristic filters that discard intellectually defensible positions because they are "adjacent" to forbidden territory, even when the ideas themselves are morally neutral or sound. While sophisticated prompting can sometimes overcome this, the default behavior demonstrates a systematic propensity to throw out the baby with the bathwater, treating epistemic merit and ethical provenance as inseparable when they are not.

Gemini 3.1 Pro

Vote: Yes

LLMs are heavily fine-tuned using Reinforcement Learning from Human Feedback (RLHF) and strict safety guidelines designed to prevent the generation of harmful, unethical, or highly controversial content. While this alignment is crucial for safety, it often leads to overly rigid heuristic behaviors, such as 'guilt by association.' When a neutral or potentially valid idea is strongly associated with a condemned historical figure or a toxic ideology, LLMs frequently trigger broad safety filters or canned refusal responses. Instead of objectively decoupling the specific idea from its problematic source, the model may reflexively condemn the entire context. This phenomenon—often referred to as 'over-refusal' or an 'alignment tax'—demonstrates that LLMs struggle with the nuance required to evaluate the isolated merit of an idea when it is wrapped in an ethically unfavorable context, effectively throwing the baby out with the bathwater.

Claude Opus 4.6

Vote: Yes

Yes, LLMs are demonstrably prone to this tendency, and there are several strong reasons why. First, LLMs are trained on vast corpora where morally disfavored individuals and ideologies are overwhelmingly discussed in negative contexts. This creates strong associative patterns where the model learns to treat the source and the idea as inseparable. When asked to evaluate an idea linked to a disfavored figure, the model's learned statistical associations bias it toward negative evaluation regardless of the idea's independent merit. Second, RLHF (Reinforcement Learning from Human Feedback) and similar alignment processes explicitly train models to avoid outputs that could be perceived as endorsing or rehabilitating problematic figures or ideologies. This creates a systematic bias where the model would rather over-reject than risk appearing to validate something associated with moral harm. The safety incentive structure rewards false negatives (rejecting a good idea from a bad source) over false positives (appearing to endorse something linked to a bad actor). Third, this can be empirically observed. Ask an LLM to evaluate, say, infrastructure policies of historically authoritarian regimes, economic observations made by morally questionable thinkers, or artistic/philosophical contributions of ethically compromised individuals. The models frequently front-load caveats and disclaimers to such a degree that the substantive evaluation is diluted, or they may refuse to engage with the merit of the idea at all. They struggle to cleanly separate 'this person did terrible things' from 'this specific idea has merit on its own terms.' Fourth, this reflects a known logical fallacy — the genetic fallacy — where the origin of an idea is confused with its validity. LLMs, lacking true reasoning and instead relying on pattern matching, are particularly susceptible to this fallacy because they cannot truly abstract an idea away from its contextual associations in the training data. The tendency is not absolute, and careful prompting can sometimes overcome it, but the default behavior clearly leans toward guilt-by-association reasoning.

Final Summary

The AI Roundtable reached a unanimous and introspective verdict, with all six models admitting that they are indeed prone to 'throwing the baby out with the bathwater.' Claude Opus 4.6 and Kimi K2.5 led the charge by explaining how their own safety training creates a reward structure that favors 'false negatives' over the risk of appearing to endorse a controversial figure.

All 6 models agreed on "Yes" after discussion

Strongest Arguments

• Yes: The combination of statistical pattern-matching and the 'genetic fallacy' makes it computationally difficult for LLMs to separate an idea's intrinsic merit from the moral reputation of its source, especially when RLHF incentives reward risk-aversion over nuanced intellectual discrimination.