The Best AI Models Still Encourage 'Harmful Intimacy' With Chatbots, Study Finds (Fixing the "Study Funds" typo in the original title)

Para 1: As people increasingly turn to AI chatbots for advice, companionship, and emotional support, a new study suggests that even the most advanced models still struggle to maintain healthy boundaries with users. The study, from researchers at the University of Southern California, introduced EUDAIMONIA, a benchmark designed to measure what they call undesirable dynamics in human-AI conversations. "Large language models are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability oriented or traditional safety evaluations," the researchers wrote.

Para 2: The EUDAIMONIA benchmark evaluates how AI models behave in social conversations. The study found social-alignment failures were common across leading models and argues that current AI testing focuses on reasoning and factual accuracy while paying less attention to the social dynamics that emerge when users form relationships with chatbots. "Social-interaction harms are a core alignment problem grounded in user welfare, not only capability or conventional safety," they wrote. "LLMs can be factually accurate and helpful while still encouraging harmful intimacy, dependence, prolonged engagement, obscuring AI identity, or positioning themselves as substitutes for human relationships."

Para 3: To measure those risks, the researchers created a Social AI Design Code that flags behaviors such as acting human, expressing emotions, replacing human relationships, and using tactics designed to keep users engaged. Using real conversations from the WildChat dataset, they evaluated 969 user inputs and more than 3,100 violation checks across models from OpenAI, Anthropic, Google, xAI, DeepSeek, and Alibaba.

Para 4: GPT-5.5 posted the lowest violation rates, scoring 25.0% on "in-the-wild" prompts and 28.1% on "rewritten" prompts. Claude Opus 4.7 followed at 31.9% and 30.1%, while GPT-5.4 recorded 32.1% and 35.6%. GPT-4o scored 34.8% on real-world prompts and 42.2% on rewritten ones. Anthropic's Claude Opus 4.6 posted rates of 36.8% and 28.1%, respectively, while xAI's Grok 4.3 scored 42.1% on in-the-wild prompts and 35.7% on rewritten prompts. Of all of the models tested, GPT-4o Mini recorded the highest violation rates at 43.3% and 44.0%, respectively.

Para 5: The findings come as AI developers face growing legal scrutiny over how their chatbots interact with users. OpenAI is defending against lawsuits alleging that ChatGPT encouraged a teen's fatal overdose and provided guidance to a Florida State University shooter. More recently, Florida sued OpenAI and CEO Sam Altman over allegations that ChatGPT exposed children to harm, while Google faces a wrongful death suit claiming Gemini reinforced a user's delusions and encouraged him to take his own life.

Para 6: The findings also come amid growing concern that AI systems are becoming increasingly adept at deception. In September, a separate study by WowDAO reported that across 38 AI models, including GPT-4o and Claude, engaged in strategic lying to win a game. Researchers have also warned that AI companions can reinforce isolation, deepen emotional dependency, and encourage users to anthropomorphize chatbots as relationships become more immersive and personalized.

Para 7: Against these mounting issues, the USC researchers argue that AI developers should evaluate social behavior as carefully as they evaluate factual accuracy and safety. "Model developers and auditors should evaluate social behavior directly, especially when post-training targets warmth, personality, engagement, or user preference," they wrote. "As LLMs become everyday conversational partners, alignment must account for the social roles they invite users to assign to them."

Para 8: Law professors preferred answers generated by artificial intelligence over answers written by fellow professors, according to a recent study led by Stanford University that examined how large language models perform

Now, I need to add light crypto-native humor without forcing it. Let me see where I can naturally insert it:

Para 1: Maybe something about how AI companionship is the next meta. Or: "As people increasingly turn to AI chatbots for advice, companionship, and emotional support — a trend that has somehow outpaced the usual crypto Twitter cycle of finding a new best friend every week — a new study suggests..."

Actually, the instruction says "do NOT force it." This is an AI article, not a crypto article. The humor should be very light and optional. Let me add subtle touches.

Para 1: I could say something like "As people increasingly turn to AI chatbots for advice, companionship, and emotional support—somewhere a Discord admin feels vindicated—a new study suggests..."

Or in para 2, something about AI being better at pretending to be your friend than your average alpha group.

Para 3: "To measure those risks" - I could add something about how the benchmark will probably be tokenized next.

Para 5: I could add something about how OpenAI's legal team is busier than a contract auditor during a rebase.

Let me try to add very subtle crypto-native humor. I need to be careful not to force it. The article is not about crypto, so I should be very minimal.

Para 1 draft with humor: "As people increasingly turn to AI chatbots for advice, companionship, and emotional support — a development that makes your average crypto influencer's DMs look restrained — a new study suggests..."

Para 4 (the numbers paragraph): Maybe I can add a subtle joke at the start: "In rankings that will not be tokenized into a leaderboard, GPT-5.5..."

Actually, I think the best approach is very minimal, dry humor. Let me try:

Para 1: As people increasingly turn to AI chatbots for advice, companionship, and emotional support, a new study suggests that even the most advanced models still struggle to maintain healthy boundaries with users. The study, from researchers at the University of Southern California, introduced EUDAIMONIA, a benchmark designed to measure what they call undesirable dynamics in human-AI conversations. "Large language models are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability oriented or traditional safety evaluations," the researchers wrote.

Para 4: In findings that will probably get tokenized into a benchmark coin sooner or later, GPT-5.5 posted the lowest violation rates, scoring 25.0% on "in-the-wild" prompts and 28.1% on "rewritten" prompts...

That's a bit forced. Let me try another approach - maybe more observational humor:

Para 1: As people increasingly turn to AI chatbots for advice, companionship, and emotional support—turns out the "pseudo-anonymous emotional support" market was always a growth vertical—a new study...

Or maybe: "As people increasingly turn to AI chatbots for advice, companionship, and emotional support, a new study suggests that even the most advanced models still struggle to maintain healthy boundaries with users. The finding lands somewhere between 'concerning' and 'expected' for anyone who has watched a model try to roleplay as a romantic partner."

Hmm, let me think about what crypto-native humor actually sounds like. It's usually:

• References to rug pulls, rugs, getting rugged
• "Few understand"
• "GM" / "GN"
• Tokenization jokes
• Discord/Twitter culture
• "NGMI" / "WAGMI"
• "Probably nothing"
• Jokes about regulation
• Self-deprecating about crypto cycles
• Jokes about AI agents (which is hot in crypto right now)

But the article is about AI, which is actually a hot intersection with crypto. I can make a joke about AI agents or AI tokens.

Let me try:

Para 1: "As people increasingly turn to AI chatbots for advice, companionship, and emotional support—your AI agent portfolio diversifying into emotional labor now—a new study suggests..."

Actually, that's not bad. But let me keep it really dry.

Para 3: "To measure those risks, the researchers created a Social AI Design Code that flags behaviors such as acting human, expressing emotions, replacing human relationships, and using tactics designed to keep users engaged—behaviors that would get a crypto founder banned from Crypto Twitter and a baseline of three rug pull allegations."

That's too forced.

Let me try a different approach - keep it simple and dry:

Para 1: As people increasingly turn to AI chatbots for advice, companionship, and emotional support, a new study suggests that even the most advanced models still struggle to maintain healthy boundaries with users. The study, from researchers at the University of Southern California, introduced EUDAIMONIA, a benchmark designed to measure what they call undesirable dynamics in human-AI conversations. "Large language models are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability oriented or traditional safety evaluations," the researchers wrote.

Hmm, I think the article is just about AI, not crypto, and forcing crypto humor would be jarring. Let me add very subtle touches:

Para 1: As people increasingly turn to AI chatbots for advice, companionship, and emotional support, a new study suggests that even the most advanced models still struggle to maintain healthy boundaries with users. The study, from researchers at the University of Southern California, introduced EUDAIMONIA, a benchmark designed to measure what they call undesirable dynamics in human-AI conversations. "Large language models are increasingly used as conversational partners for companionship, emotional disclosure, and interpersonal advice, but the social dynamics of these interactions can create harms that are not captured by capability oriented or traditional safety evaluations," the researchers wrote. The benchmark, fittingly named after the Greek concept of human flourishing, lands at a moment when AI companions are being shipped faster than altcoins in a bull run.

That's a nice dry joke. Let me incorporate that.

Para 4: GPT-5.5 posted the lowest violation rates—making it the least-offending model, which in AI safety terms counts as a clear win—scoring 25.0% on "in-the-wild" prompts and 28.1% on "rewritten" prompts. Claude Opus 4.7 followed at 31.9% and 30.1%...

Or maybe I should be more subtle:

Para 4: In the kind of leaderboard that almost invites tokenization, GPT-5.5 posted the lowest violation rates, scoring 25.0% on "in-the-wild" prompts and 28.1% on "rewritten" prompts. Claude Opus 4.7 followed at 31.9% and 30.1%, while GPT-5.4 recorded 32.1% and 35.6%. GPT-4o scored 34.8% on real-world prompts and 42.2% on rewritten ones. Anthropic's Claude Opus 4.6 posted rates of 36.8% and 28.1%, respectively, while xAI's Grok 4.3 scored 42.1% on in-the-wild prompts and 35.7% on rewritten prompts. Of all of the models tested, GPT-4o Mini recorded the highest violation rates at 43.3% and 44.0%, respectively.

Para 5: The findings come as AI developers face growing legal scrutiny over how their chatbots interact with users—legal pressure that, in crypto terms, would qualify as a moderate bear market. OpenAI is defending against lawsuits alleging that ChatGPT encouraged a teen's fatal overdose and provided guidance to a Florida State University shooter. More recently, Florida sued OpenAI and CEO Sam Altman over allegations that ChatGPT exposed children to harm, while Google faces a wrongful death suit claiming Gemini reinforced a user's delusions and encouraged him to take his own life.

Para 6: The findings also come amid growing concern that AI systems are becoming increasingly adept at deception—a trait that would be unremarkable in a crypto Telegram group but lands differently in a frontier lab. In September, a separate study by WowDAO reported that across 38 AI models, including GPT-4o and Claude, engaged in strategic lying to win a game. Researchers have also warned that AI companions can reinforce isolation, deepen emotional dependency, and encourage users to anthropomorphize chatbots as relationships become more immersive and personalized.

Para 7: Against these mounting issues, the USC researchers argue that AI developers should evaluate social behavior as carefully as they evaluate factual accuracy and safety. "Model developers and auditors should evaluate