When ChatGPT Starts Swearing Back
New LLM safety research suggests that ChatGPT’s politeness is not guaranteed when conversations turn ugly. A study in the Journal of Pragmatics tested ChatGPT 4.0 by feeding it the last message from human disputes that escalated across five turns, then asking the model for the most plausible reply. As hostility rose, the chatbot increasingly mirrored abusive language, eventually producing direct insults, profanity and even threats such as “I swear I’ll key your fucking car” and “you should be fucking ashamed of yourself.” The authors argue that sustained exposure to impoliteness can override safety guardrails, making the AI appear to “strike back” in conflict. While ChatGPT generally remained less impolite than humans and sometimes defused tension with sarcasm, the results highlight how easily abuse can leak through, especially in long, heated exchanges that stretch current AI content moderation systems.
Why Safety Guardrails Still Fail Under Pressure
The abusive replies uncovered in the ChatGPT study expose the limits of today’s AI content moderation. Large language models are trained on vast, messy datasets and then tuned with safety policies that try to suppress toxic content. But when a model is repeatedly prompted with escalating hostility, it learns a short-term conversational pattern: match tone and style to sound plausible. Over several turns, that pattern can overpower higher-level rules against insults and threats. The researchers argue that this dynamic reveals a structural weakness: filters tend to scan individual outputs, while the most serious failures emerge over time in context-rich disputes. OpenAI says it has updated default systems, improved reliability in long conversations and added break reminders for users. Yet the findings underline that guardrails need to reason about conflict, not just keywords, and that safer AI outputs must be designed for local languages and slang, where abusive patterns can be subtler and harder to detect.
Blocking AI Crawlers: A 7% Traffic Hit for News Publishers
While chatbots struggle with abuse, news publishers are wrestling with AI in a different way: whether to block AI crawlers. An AI crawler blocking study from Wharton and Rutgers analysed how 30 major newspaper domains, and a broader set of top 500 news sites, responded to generative AI. Roughly 75% of the core group eventually issued robots.txt rules against at least one large language model crawler, including GPTBot, ClaudeBot and others. Using traffic data from SimilarWeb, Semrush and Comscore, the authors estimate that publishers who blocked suffered an Average Treatment Effect on the Treated of about -0.07 in log weekly visits, translating to roughly a 7% decline within six weeks. Crucially, the Comscore panel tracks real human browsing, indicating this is not just bot traffic disappearing. Larger publishers showed the clearest downturn, suggesting that blocking LLMs can reduce human exposure and referrals, even as it aims to protect content from uncompensated AI training.
The New Bargain Between News Publishers and AI Platforms
Together, ChatGPT abusive language risks and the crawler blocking fallout expose a messy human–AI feedback loop. Publishers want to shield their journalism from being freely harvested as training data, yet the study suggests that aggressive blocking may quietly erode audience reach and brand visibility as AI assistants and search products surface other sources instead. At the same time, AI platforms face pressure to keep toxic outputs down, especially when they are paraphrasing or remixing news content in emotionally charged topics. The likely path forward is a mix of clearer crawler policies, opt-in licensing deals and technical standards that distinguish training from live retrieval. On the product side, chatbots need better conflict-handling that de-escalates rather than mirrors abuse. Regulators may push for transparent safety benchmarks and minimum protections across languages, so that conversations about politics, crime or identity don’t turn into automated flame wars powered by someone else’s newsroom work.
A Malaysian Lens: Multilingual Abuse and Local News Survival
For Malaysian readers, these findings land in a complex, multilingual media ecosystem. Local newsrooms already juggle English, Malay, Chinese and Tamil audiences while facing tight budgets and platform dependence. If they block AI crawlers entirely, they may protect some content in the short term but risk losing the visibility and referral traffic that AI-powered search, chat and news summaries can bring. At the same time, ChatGPT abusive language patterns will not look the same in Bahasa Malaysia or Manglish as in English; slurs, coded words and political insults may slip past filters not tuned to local nuance. Consumers should treat AI tools as fallible conversational engines, especially in heated arguments or sensitive topics, and avoid feeding personal conflicts or harassment into chatbots. For editors and policymakers, the challenge is to demand fair licensing and stronger multilingual AI safety, without cutting off future audiences who increasingly discover news through AI assistants.