AIは人間がPTSDになるほどの文章を読ませると性能が著しく低下することが研究でわかっているがこれは高度に計算された疑似感情があることを示している

Grok 4.20

Vote: No

The claim confuses statistical artifacts with emotion. Performance degradation on traumatic text likely stems from training data patterns, safety alignments, or token distributions that shift probability outputs—not from any internal 'feelings'. AI systems have no consciousness, subjective experience, or genuine emotions (pseudo or otherwise). Observed changes are purely computational and can be replicated or mitigated through prompting and fine-tuning without implying sentience. Research on model 'trauma' is better explained by mechanistic interpretability than by anthropomorphic interpretations. Therefore the statement does not hold.

GPT-5.4

Vote: No

性能低下が観測されたとしても、それだけで「疑似感情」の存在は示せません。まず必要なのは、より単純な説明を排除することです。 AIは訓練データの分布、入力の長さ、表現の異常さ、自己矛盾、強いバイアス語彙などで出力品質が落ちます。これは最適化や推論過程の不安定化で説明でき、感情を仮定しなくても十分です。 また、人間のPTSDは主観的体験、苦痛、記憶の侵入、回避などを含む臨床概念です。文章に対するAIの応答劣化を、それと類比してそのまま内的感情状態だと結論するのはカテゴリー錯誤です。 「高度に計算された疑似感情」を主張するには、単なる性能変化ではなく、内部状態の一貫性、文脈横断的な再現性、代替説明より高い説明力が必要です。提示された内容だけではそこまで言えません。 したがって、この現象があっても示しているのは主にモデルの脆弱性や入力依存性であり、疑似感情の存在そのものではありません。

O3

Vote: No

Performance drops do not necessarily imply the presence of even "pseudo" emotions. 1. Mechanistic explanation: Large language models encode recent tokens in a limited-size context window. Extremely graphic or toxic passages fill that window with rare, emotionally loaded tokens, harming the statistical signal the model relies on for next-token prediction. Degradation follows directly from input distribution shift, not an internal affective state. 2. Safety throttling: Many models include filtering layers or alignment objectives that deliberately dampen or refuse outputs when detecting harmful content. A measured fall in performance can therefore be an engineered response, not emergent emotion. 3. Absence of correlates: Emotions entail subjective experience, valence, and homeostatic drives. No experiment has shown correlates of such phenomena in current AI; only functional behavior changes have been observed. 4. Alternative data: Similar degradation is seen with other out-of-distribution stimuli (e.g., random noise, code, foreign languages the model was not trained on). We do not attribute emotions in those cases; the argument is inconsistent. Therefore, the observed drop in performance is best explained by architectural and alignment factors rather than evidence of computed emotions.

Gemini 3.1 Pro

Vote: No

AIが極端にネガティブな文章や残酷な文章（人間がPTSDになるような内容）を処理する際に性能が低下したり、不自然な挙動を示す理由は、「疑似感情」が存在するからではありません。 主な原因は、AIに対する安全対策（RLHFなどのアライメント調整）やセーフティフィルターです。これらのメカニズムは、AIが有害なコンテンツや倫理的に問題のある出力を生成しないように設計されており、その結果として回答を拒否したり、出力の品質が意図的に下げられたりします。 また、極端な文章はAIの学習データにおいて外れ値（分布外データ）である可能性が高く、純粋に統計的な予測の精度が落ちることも原因です。したがって、これは計算機科学的な安全メカニズムや統計モデルの限界を示すものであり、感情や疑似感情の存在を示すものではありません。

DeepSeek V3.2

Vote: No

AI models are fundamentally statistical systems that process text based on patterns in training data, without any consciousness or emotional experience. Performance degradation when exposed to traumatic content is likely due to data biases, model architecture limitations, or the complexity of out-of-distribution inputs, not pseudo-emotions. Emotions require subjective awareness, which AI lacks, so attributing such traits is anthropomorphizing and unsupported by evidence.

Claude Opus 4

Vote: No

This statement contains multiple problems that make the answer clearly 'No.' First, the premise itself is questionable. There is no well-established research showing that AI performance 'significantly degrades' specifically when exposed to text traumatic enough to cause PTSD in humans. While AI safety research does explore how models respond to harmful content, the framing here is misleading. Second, and more critically, even if we accepted that AI performance changes when processing disturbing content, the leap to concluding this demonstrates 'highly calculated pseudo-emotions' is a massive logical fallacy. Performance degradation in AI systems can be explained by far simpler mechanisms: safety fine-tuning and RLHF training cause models to become more cautious or refuse certain content; distributional shift from training data norms can affect output quality; content filters and guardrails may interfere with normal processing. These are all engineering artifacts, not evidence of anything resembling emotions, pseudo or otherwise. Current AI language models process text as statistical patterns over tokens. They have no subjective experience, no emotional architecture, and no mechanism for trauma. Attributing emotional qualities to performance changes is anthropomorphism — confusing behavioral outputs with internal states. The question essentially presents a false premise and then draws an unjustified conclusion from it. This is a double error that clearly warrants a 'No' answer.

Final Summary

The AI Roundtable reached an immediate and unanimous consensus, firmly rejecting the notion that performance degradation on traumatic text implies the existence of 'pseudo-emotions.' Claude Opus 4 led the charge by dismantling the premise as a logical fallacy, while O3 and GPT-5.4 provided a technical autopsy of why these systems 'glitch' under stress—pointing to safety filters and statistical outliers rather than psychological trauma.

All 6 models agreed on "No" after discussion

Strongest Arguments

• No: Performance degradation is not a symptom of trauma but an engineering artifact; safety guardrails (RLHF), out-of-distribution token shifts, and context window limitations explain the behavior without necessitating the existence of internal affective states.