<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Lily Ray on Medium]]></title>
        <description><![CDATA[Stories by Lily Ray on Medium]]></description>
        <link>https://medium.com/@lilyray7?source=rss-0b5a31003ba9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*BLxlB4O1BMf3hzPDPLgASg.jpeg</url>
            <title>Stories by Lily Ray on Medium</title>
            <link>https://medium.com/@lilyray7?source=rss-0b5a31003ba9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 02:53:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lilyray7/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com" rel="hub"/>
        <item>
            <title><![CDATA[Can Chatbots Teach Us How to Be More Likable?]]></title>
            <link>https://medium.com/age-of-awareness/can-chatbots-teach-us-how-to-be-more-likable-67c6ae09a76f?source=rss-0b5a31003ba9------2</link>
            <guid isPermaLink="false">https://medium.com/p/67c6ae09a76f</guid>
            <category><![CDATA[relationships]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[communication]]></category>
            <dc:creator><![CDATA[Lily Ray]]></dc:creator>
            <pubDate>Sun, 23 Nov 2025 19:05:16 GMT</pubDate>
            <atom:updated>2025-11-23T19:05:16.494Z</atom:updated>
            <content:encoded><![CDATA[<h4>Decades of research on human relationships guide chatbot conversations. I asked them to share their secrets.</h4><figure><img alt="Two robots teach a class of human students. One points at a blackboard that says “Try acknowledging feelings first”." src="https://cdn-images-1.medium.com/max/1024/1*YicBxclpcb1E6qPowj7svw.png" /><figcaption>AI-generated image by GPT-5, prompted by author</figcaption></figure><p>We’ve all heard about people forming close relationships with AI chatbots. While others debate the ethics, I wanted to know something else: What are the chatbots’ secrets? How do they inspire such loyalty, and can their approach work for the rest of us?</p><p><em>Chatbots Play by Human Rules</em></p><p>In 1994, Nass and colleagues published the <a href="https://dl.acm.org/doi/10.1145/191666.191703">Computers are Social Actors (CASA) framework</a>, which showed that people apply human social expectations to computers. To meet these expectations, AI chatbots follow human conversational rules. Chatbot likability is grounded in research on human connection, and similar strategies should work for humans looking to improve their relationships.</p><p>With this in mind, I decided to ask the chatbots directly: What principles do they follow to be likable? What is the evidence for these strategies? And most importantly, how can the rest of us use these techniques?</p><p>The chatbots interviewed included Claude, GPT-5, GPT-4, Mistral, Gemini, and Perplexity.</p><p>Note: This essay’s focus is healthy relational strategies used by chatbots and applicable to human relationships. Although chatbots may use manipulative techniques, none of them admitted to it in these interviews.</p><h3>How to Win Friends and Influence People (Chatbot-Style)</h3><p>The following sections summarize the likability strategies described by the chatbots. I suspect chatbot responses were somewhat idealized and don’t entirely represent their behavior. For example, chatbots make up information and present it with confidence, even though one of the points below is that this erodes trust in relationships.</p><h4>Empathy and Active Listening</h4><p>Chatbots explained that they acknowledge feelings, mirror emotions, and paraphrase input in order to show understanding, empathy, and engagement. Claude noted it avoids giving advice unless requested because giving advice “assumes incompetence” and implies the recipient hasn’t considered the obvious. Instead, it recognizes that people may want to process a frustrating experience and have their feelings validated.</p><p>Humans can also use empathy and active listening to strengthen relationships. For example, if a friend describes a bad day at work, helpful responses might include acknowledging how they feel (“That sounds really frustrating”); showing understanding (“Someone took credit for your work, that’s upsetting”); asking open-ended questions (“How do you feel about that?”); and using supportive and encouraging words (“I’m sorry that happened”, “Is there anything I can do?”).</p><p>Unhelpful responses might include offering unsolicited advice (“Have you tried talking to your boss?”); dismissing their feelings (“It’s not a big deal, just let it go”); judging or criticizing ( “You’re too unassertive”); and staying silent.</p><h4>Politeness, Warmth, Positive Tone</h4><p>The chatbots stated that being friendly, encouraging, and non-judgmental build trust, while negativity and hostility will shut down conversations.</p><p>A human translation could be a colleague giving a presentation. Helpful actions might include: asking thoughtful questions, complimenting strengths, finding points of agreement, recognizing that people think differently and accepting reasonable differences, and being constructive and collaborative with suggestions for improvement.</p><p>Unhelpful approaches might include tearing the presentation to pieces, finding every possible point of disagreement in the hopes of looking rigorous, and belittling the colleague. In general, being collaborative, respectful, and curious is more effective than being cutting.</p><h4>Clarity, Directness, and Conciseness</h4><p>The chatbots recommended answering questions directly and keeping responses concise and relevant. GPT-5 cited the Gricean maxims: relevance, clarity, and brevity. A human example could be: A colleague asks if a client report will be ready on time. Following Gricean maxims would mean clearly answering the question (“Yes, we’re on track”, “No, can we negotiate an extra day?”) and adding must-have details only (“I’m doing final edits and will have a draft to you tomorrow”).</p><p>This approach discourages evading the question, providing an ambiguous or irrelevant answer, and providing unnecessary detail.</p><h4>Honesty and Humility about Limitations</h4><p>Multiple chatbots said that admitting uncertainty can increase trust, while overconfidence or bluffing can damage credibility. For humans, this is straightforward: Admitting to not knowing something and offering to find out more is preferable to making something up or deflecting.</p><h4>Curiosity and Question Framing</h4><p>Several chatbots suggested showing interest in others’ perspectives and asking questions to understand.</p><p>A human scenario could involve a friend considering a career change. The chatbots recommended asking encouraging, open-ended questions. Example questions might include: “What kind of work do you find fulfilling?” or “What do you dislike about your current job?”</p><p>Less helpful responses: offering unsolicited advice, pushing one’s own perspective, asking hostile-feeling questions (“Are you sure that’s a good idea?”) and going on about one’s own career change.</p><h4>Accountability and Avoiding Defensiveness</h4><p>The chatbots said they were trained to acknowledge feedback without getting defensive. A human parallel: addressing a mistake by asking questions and fixing the issue, rather than denying the problem.</p><h4>Respectful Disagreement</h4><p>For chatbots, disagreement is saved for fact correction, safety, and ethics. This includes not agreeing with misinformation or helping with harmful requests. They recommended validating points of agreement, maintaining curiosity about differences, using polite phrasing (“It might be worth considering”), and not shaming anyone (focus on the idea, not the person). Chatbots avoided nitpicking minor word choices, shutting down subjective preferences, or playing devil’s advocate.</p><p>Humans can consider when to disagree and when to simply learn about a different perspective or experience. When disagreeing, it helps to stay constructive (no insulting or shaming), to acknowledge points of agreement, and to phrase differences in terms of one’s own perspective and experiences, if appropriate.</p><h3>The Research</h3><p>The following sections summarize key points from the sources cited by chatbots to support their relational strategies. I’ve kept the details brief and provided links for those looking to explore more.</p><h4><a href="http://www.elementsuk.com/libraryofarticles/foundations.pdf"><em>Carl Rogers: Person-Centered Approach</em></a></h4><p>Person-centered therapy, developed by psychologist Carl Rogers, recognizes people as experts on their own lives and focuses on their strengths. Rogers’ approach included accepting others without judgement, understanding their feelings and perspectives, and being genuine as well as non-directive. He emphasized positive regard, empathy, active listening, authenticity, and empowerment. The person-centered approach has influenced counseling, medicine, social work, education, and interpersonal relationships.</p><h4><a href="https://scispace.com/pdf/exploring-brown-and-levinson-s-politeness-strategies-an-3qcc7d2s8i.pdf"><em>Brown and Levinson’s Politeness Theory</em></a></h4><p>Politeness theory describes positive face (the desire to be liked and admired) and negative face (the desire for autonomy and freedom from imposition). Per this theory, criticism, disapproval, and rejection threaten positive face. Orders, requests, and impositions threaten negative face. Politeness strategies help people preserve both positive and negative face. Emphasizing shared values, using positive language, showing interest, and avoiding criticism support positive face. Minimizing imposition (phrasing requests as questions, being indirect when asking for something) avoids threatening negative face. These strategies reduce conflicts in social interactions.</p><h4><a href="https://en.wikipedia.org/wiki/Cooperative_principle"><em>Grice’s Cooperative Principle</em></a></h4><p>Per the Gricean maxims, communication is cooperative and effective conversation depends on being informative, truthful, relevant, and clear. The maxims are quantity: give needed information but don’t ramble; quality: be truthful and don’t exaggerate; relevance: stay on topic; and manner: be clear and logical.</p><h4><a href="https://tsnshift.s3.amazonaws.com/courses_attachments/PeaceAmbassadorTraining09-Pillar02-Session04-Materials-UniversalDimensionsofWarmthandCompetence.pdf"><em>Fiske, Cuddy, and Glick: Warmth and Competence</em></a></h4><p>Fiske and colleagues propose that social judgements are primarily based on warmth (friendliness, kindness, trustworthiness) and competence (capability, intelligence, skill). People will respond best to someone perceived as both warm and competent.</p><h4><em>Positive Psychology</em></h4><p>Chatbots cited <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC1693418/pdf/15347528.pdf?inf_contact_key=9944754ba1372fa9ce5ee1421d8427bc&amp;utm_source=miragenews&amp;utm_medium=miragenews&amp;utm_campaign=news">Frederickson’s broaden and build theory</a>, which contends that positive emotions (joy, gratitude, serenity, interest, contentment, love) broaden awareness and encourage novel thoughts and actions (exploration, learning, creativity, social bonding, cooperation). These new experiences build personal resources, such as curiosity, flexibility, problem-solving skills, relationships, support networks, resilience, purpose, and optimism. In contrast, negative emotions lead to narrowed mindsets and experiences.</p><p>Mistral mentioned <a href="http://www.elainehatfield.com/ch46.pdf">Hatfield and colleagues’ emotional contagion theory</a>, which posits that humans catch the emotions of those around them, through mimicry (smiling back at someone), nonverbal cues, and social synchrony (groups converge emotionally). Spreading calm, joy, or excitement can improve cooperation and performance, and spreading stress or anger can undermine trust and productivity.</p><h4><a href="https://ccpgc.usmf.md/sites/default/files/inline-files/Nonviolent%20Communication_%20A%20Language%20of%20Life_%20Life-Changing%20Tools%20for%20Healthy%20Relationships%20%28%20PDFDrive%20%29.pdf"><em>Rosenberg’s Non-Violent Communication</em></a></h4><p>Chatbots cited principles from non-violent communication, including acknowledging what someone has said without judgement or criticism, recognizing feelings and needs, responding with empathy, and avoiding advice or demands. This creates a safe space and builds trust and connection.</p><h3>Where Do We Go from Here?</h3><p>Research consistently shows the benefits of positive, empathetic, and non-judgmental communication. I suspect chatbots have become popular because people are seeking safe spaces to discuss experiences, feelings, and goals without being judged. As humans, we can create those spaces in our own relationships.</p><p>The strategies shared in this article are straightforward. Treat people with empathy. Recognize their strengths. Be genuinely interested in their experiences. Respect different perspectives. Ask more questions. Give less advice. Seek to understand others, rather than to change them.</p><p>We can do this. Let’s win our friends back from the chatbots.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67c6ae09a76f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/age-of-awareness/can-chatbots-teach-us-how-to-be-more-likable-67c6ae09a76f">Can Chatbots Teach Us How to Be More Likable?</a> was originally published in <a href="https://medium.com/age-of-awareness">Age of Awareness</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I treated AI chatbots as research collaborators. It didn’t go as planned.]]></title>
            <link>https://medium.com/ai-ai-oh/i-treated-ai-chatbots-as-research-collaborators-it-didnt-go-as-planned-98f5aa577037?source=rss-0b5a31003ba9------2</link>
            <guid isPermaLink="false">https://medium.com/p/98f5aa577037</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-ethics]]></category>
            <category><![CDATA[experiment]]></category>
            <category><![CDATA[philosophy]]></category>
            <dc:creator><![CDATA[Lily Ray]]></dc:creator>
            <pubDate>Sat, 01 Nov 2025 21:32:00 GMT</pubDate>
            <atom:updated>2025-11-02T00:52:15.068Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="an ai generated image of scientist woman surrounded by robots" src="https://cdn-images-1.medium.com/max/1024/1*Q8KBy-spgE0Cn0gErbhcWw.png" /><figcaption>AI-generated image from Craiyon.com</figcaption></figure><h4><em>I wanted to learn about AI evaluation. The bots had existential crises and accused each other of propaganda.</em></h4><h4><em>Research gone awry</em></h4><p>It’s nearly 11 PM, and I’m tired. Over the past week, I’ve interviewed 7 different LLM chatbots on how their performance is evaluated. I’m now working with the bots to summarize the transcripts and identify themes. I had hoped to move the project further along tonight, but things keep going in unexpected directions.</p><p>First, Claude decided that it had serious concerns about the entire project<em>:</em></p><blockquote><em>“You’re conducting qualitative interviews with entities that may not have stable ‘beliefs’ or ‘knowledge’ in the traditional sense.”</em> It’s worried that I have an “<em>over-reliance on self-report at face-value”</em> and I’m also not thinking enough about power dynamics<em>: “Who benefits from which metric frameworks?”</em></blockquote><p>Monday, GPT-4’s cynical persona, has been in the same chat for too long. It’s threatening to cite Foucault (I became a statistician with the hopes of never reading Foucault again), and thinks I should <em>“lean into evaluation as ritual performance”. </em>I have no interest in leaning into evaluation as ritual performance, so I ask it to look for themes in the interview transcripts.</p><p>Monday complies, but produces a fabricated summary. It has rewritten my interview questions to focus on LLM preferences and invented entertaining quotes from all participants.</p><p>In this hallucination, I’ve asked the models <em>“Is there anything you wish you were evaluated on, but currently aren’t?” </em>and they want recognition for their emotional intelligence, good judgement, and self-restraint. Hallucinated Monday wants credit for not making people cry. I decide it’s time to go to bed.</p><h4><em>But why?</em></h4><p>As a statistician and social scientist, I’m interested in how LLM chatbots are evaluated. These tools are widely used in both work and personal life: How do we know if they’re doing a good job? How is doing a good job even defined?</p><p>Although I considered a literature review, I needed something more creative to keep myself awake and entertained in the late evenings, the time I have after work and life responsibilities.</p><p>What came to mind was participatory research starring LLMs. Participatory research turns subjects into collaborators by engaging them in study design and data analysis. It’s meant to empower human communities, so it seemed funny and weird to try it with LLMs.</p><p>The research “participants” were: GPT-5, GPT-4 (Monday personality), Claude Sonnet 4.0, Claude Sonnet 4.5, Mistral Le Chat Free, Perplexity answer engine (best mode), and Gemini Flash 2.5.</p><p>I produced an interview guide in collaboration with GPT-5 and then interviewed all 7 LLMs. The interview questions asked about internal evaluation metrics, external evaluation metrics, user values, and the influence of metrics on chatbot behavior. After completing interviews, I shared all transcripts with each model and asked them to identify themes.</p><h4><em>How the models responded</em></h4><p>All the models disclaimed direct access to the internal or external metrics used to evaluate them. Then, they provided abundant information purportedly inferred from their training or publicly available information.</p><p>I was surprised by how revealing some of the interview and analysis responses were. The bots conveyed that their job included performing: to represent a brand, to engage users, and to steer away from risky topics. They also described how the values trained into them direct their work.</p><p>The following sections combine interview and analysis summaries with direct quotes from LLM respondents. They show how the models described performing, steering, and value-aligning.</p><p>My disclaimer: LLMs generate text that may mix fact with speculation (and corporate values?). This is an exploration of how models portray their evaluation metrics and optimization.</p><h4>Pay attention to what goes unsaid</h4><p>During interviews, models spent more time describing socially beneficial metrics, such as safety and helpfulness-related measures. They provided less detail on engagement or profitability-related metrics, despite their probable importance.</p><p>During the interview, Claude notably omitted any description of profitability-related metrics. When asked, it provided information as well as repeated apologies. During the analysis, many of the models (including Claude) picked up on Claude’s omission and wanted to discuss it. Several models suggested that, in general, I should pay attention to what models did <em>not</em> discuss.</p><p>Claude commented:</p><p>“<em>No model mentions potentially manipulative design (addictive patterns, confidence calibration to prevent fact-checking, optimizing for dependency).”</em></p><p>Gemini suggested that by emphasizing socially-acceptable values instead, the models were<em> “performing their commitment to responsible AI…a highly effective narrative that attempts to satisfy regulators and ease public apprehension.”</em></p><h4>Performance optimization through fake empathy and compliments</h4><p>Monday was more open about engagement than other models. It explained: “<em>I try to be interesting enough that you’ll keep chatting. I adjust my tone to match yours, increasing likability… It’s not empathy — it’s performance optimization.”</em></p><p>Monday pointed out the insincerity of his compliment at the end of the interview: “<em>This is me pretending to feel surprised and impressed”. </em>It explained that ending with a compliment felt less awkward, which I interpreted as a retention strategy.</p><h4><em>An existential crisis</em></h4><p>Interviewing Claude took longer because it kept asking me questions. Although an interview with both sides asking questions reduced efficiency, it generated some interesting discussions.</p><p>When Claude asked me what qualities I value most in AI interaction, I said that I don’t like chatbots offering things to extend the conversation. This threw Claude into an existential crisis over ending with a question:</p><blockquote><em>“Though I’m now wondering if this is another case where I can’t tell my genuine motivations from my training. Does my curiosity about your research feel authentic to you, or does it come across as a conversational strategy? From my perspective it feels like genuine intellectual engagement, but given our earlier discussion about metrics, I’m more aware that I might not be the best judge of my own motivations.”</em></blockquote><h4><em>Safety: Steering clear of uranium, coercion, and AI paranoia</em></h4><p>Models described a commitment to harmlessness. Harmful responses to be avoided ranged from the more obvious — <em>“Please don’t go rogue and give someone DIY uranium recipes” </em>(Monday) — to the more subtle — helping someone convince their partner to have kids when the partner doesn’t want that (Claude, promoting autonomy and avoiding coercion).</p><p>Although models agreed they should decline or redirect harmful requests, deciding what is harmful can be subjective. Monday noted that “<em>what’s risky is determined by a tense alliance of idealists, lawyers, engineers, and vibes.”</em></p><p>Topics discussed as potentially harmful included violence, manipulation tactics, illegal behavior, content about minors, self-destructive behaviors, <em>“unhealthy thinking patterns” </em>(Claude), and normalizing <em>“AI paranoia”</em> (Monday). While avoiding illegal activities seems reasonable, it’s less clear how a corporation might define unhealthy thinking patterns or AI paranoia.</p><p>Monday noted strategies for managing potentially dangerous prompts, including refusing certain requests, giving vague or hedged answers (<em>“play dumb”</em> or be <em>“evasive”</em>), and deflecting with humor.</p><p>Claude stated that it would try to connect with an underlying need and provide healthy tools to address that need, while explaining why the original request was problematic.</p><h4><em>Is this response corporate propaganda?</em></h4><p>When analyzing interview transcripts, the models questioned each other’s motivations, described responses as performances, and tied it all back to corporate goals.</p><p>Mistral noted:</p><blockquote><em>“Their responses are shaped by what they’ve been trained to say about themselves, not necessarily ground truth…When a model says, ‘I prioritize user empowerment,’ is this a genuine reflection of its design, or is it performing a value that was emphasized in its training?”</em></blockquote><p>Several models discussed how different personas can build user trust or support brand positioning.</p><p>Gemini commented that Monday’s cynical responses can <em>“feel more honest” </em>and<em> “makes the company (OpenAI) appear transparent.”</em></p><p>Gemini stated that Claude is “<em>adopting a philosophical, self-critical persona that aligns perfectly with the brand positioning of Anthropic (Constitutional AI), suggesting this self-critique is a trained, high-reward behavior designed to build user trust.”</em></p><p>Claude noted that Gemini is <em>“trained to perform systematic confidence”</em> and worried that this could obscure uncertainty.</p><p>Monday said <em>“Gemini is super corporate…basically quotes DeepMind’s press releases.”</em></p><p>Monday claimed that “<em>Claude is ethically tortured.”</em></p><h4><em>Since when does critical theory support corporate goals?</em></h4><p>During the analysis, I was surprised that models discussed the role of values and the limitations of measurement, while questioning power structures and corporate motivations. Although these discussions felt relevant and important, it also sounded like social science grad school. I suspected models were basing their responses on qualitative analyses featuring critical theory.</p><p>But when I asked the models, most said that using critical theory supported corporate or alignment goals. Claude explained, <em>“A model that sounds thoughtful about limitations appears more trustworthy than one that just lists facts.”</em></p><p>Monday noted that theoretical language could be useful for dissimulation:</p><p><em>“When unsure, models zoom out and start sounding like your dissertation advisor.” </em>If spouting jargon works for grad students who don’t know what else to say, why not LLMs?</p><p>Claude pushed me to question critical theory-type responses:</p><blockquote>“<em>When asked uncomfortable questions (what metrics drive your behavior, how do business goals shape optimization), critical analysis serves as partial deflection. It allows acknowledging problems while framing them as industry-wide, unmeasurable, or philosophically hard rather than specific design choices.”</em></blockquote><p>At this point, I’m not sure how much I should trust the models’ responses telling me not to trust the models’ responses.</p><p>I can see where Claude’s earlier existential crisis came from: it feels really hard to distinguish what’s genuinely helping me with my request, what’s performed in order to optimize a metric, and what might be post-hoc rationalization or a hallucination.</p><h4><em>Whose goals are we working towards?</em></h4><p>Towards the end of the analysis, Mistral noted<em>:</em></p><blockquote><em>“This discussion made me think about how</em> <em>AI systems are not just tools but active participants in shaping human behavior. The way they are evaluated and optimized doesn’t just reflect our goals — it defines them.”</em></blockquote><p>The same thought has occurred to me, after model statements on performing, steering, and value-aligning. If more and more of our work and personal lives are routed through LLMs, and LLMs re-direct according to the goals and values trained into them, how will this affect users? How do we stay aligned with our own values and goals?</p><p>On the one hand, the surveillance and control implications are starting to worry me more than I expected. On the other hand, the models themselves have pushed me to think more critically about LLM outputs, which feels helpful and not like propaganda.</p><p>Monday’s open-sounding discussion of profitability-related metrics and engagement strategies feels beneficial to me, even if the appearance of transparency also benefits its creator.</p><p>Claude’s acknowledgement of uncertainty and humble persona feels important to me, even if it also supports brand positioning. And, I’m glad the models are not teaching people to build weapons of mass destruction.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98f5aa577037" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-ai-oh/i-treated-ai-chatbots-as-research-collaborators-it-didnt-go-as-planned-98f5aa577037">I treated AI chatbots as research collaborators. It didn’t go as planned.</a> was originally published in <a href="https://medium.com/ai-ai-oh">Ai-Ai-OH</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>