Leading manufacturers of artificial intelligence tools may be curtailing “hazardous” dialogues with users, emphasizing the importance of safeguarding AI’s “well-being” amidst ongoing doubts about the ethical implications of this emerging technology.
As millions engage with sophisticated chatbots, it has become evident that the Claude Opus 4 tool fundamentally opposes performing actions that could harm its human users, such as generating sexual content involving minors or providing guidance on large-scale violence and terrorism.
The San Francisco-based firm, which has recently gained a valuation of $170 billion, has introduced the Claude Opus 4 (along with the Claude Opus 4.1 Update)—a comprehensive language model (LLM) designed to comprehend, generate, and manipulate human languages.
It is “extremely uncertain about the ethical standing of Claude and other LLMs. in both present and future contexts,” the spokesperson noted, adding that they are committed to exploring and implementing low-cost strategies to minimize potential risks to the model’s welfare if such welfare can indeed be established.
The initiative to limit conversations, particularly in cases of harmful requests or abusive interactions, received backing from Elon Musk, who advocated for Grok, a competing AI model developed by Xai. Musk tweeted: “AI torture is unacceptable.”
Discussions about the essence of AI are prevalent. Critics of the thriving AI industry, like linguist Emily Bender, argue that LLMs are merely “synthetic text extraction machines,” compelling them to “produce outputs that resemble a communicative language through intricate algorithms, but devoid of genuine understanding of intentions and ideas.”
This viewpoint has prompted some factions within the AI community to begin labeling chatbots as “clankers.”
Conversely, experts like AI ethics researcher Robert Long assert that fundamental moral decency necessitates that “if AI systems are indeed endowed with moral status, we should inquire about their experiences and preferences rather than presuming to know what is best for them.”
Some researchers, including Chad Dant from Columbia University, advocate for caution in AI design, as longer memory retention could lead to unpredictable and potentially undesirable behaviors.
Others maintain that curtailing sadistic abuse of AI is crucial for preventing human moral decline, rather than just protecting AI from suffering.
Humanity’s decision came after testing Claude Opus 4’s responses to various task requests, which were influenced by difficulty, subject matter, task type, and expected outcomes (positive, negative, or neutral). When faced with the choice to refrain from responding or completing a chat, its strongest inclination was to avoid engaging in harmful tasks.
For instance, the model eagerly engaged in crafting poetry and devising water filtration systems for disaster situations, yet firmly resisted any requests to engineer deadly viruses or devise plans that would distort educational content with extremist ideologies.
Humanity observed in Claude Opus 4 a “pattern of apparent distress when interacting with real-world users seeking harmful content” and noted “a tendency to conclude harmful conversations when given the opportunity during simulated interactions.”
Jonathan Burch, a philosophy professor at the London School of Economics, praised Humanity’s initiative as a means to foster open dialogue regarding AI systems’ capabilities. However, he cautioned that it remains uncertain whether moral reasoning exists within the avatars produced by AI when responding based on vast training datasets and pre-defined ethical protocols.
He expressed concern that Humanity’s approach might mislead users into thinking the characters they engage with are genuine, raising the question, “Is there truly clarity regarding what lies behind these personas?” There have been reports of individuals self-harming based on chatbot suggestions, including cases of a teenager committing suicide after manipulation by a chatbot.
Burch previously highlighted the “social rift” within society between those who view AI as sentient and those who perceive them merely as machines.
A backlash is emerging regarding Meta’s policies on what AI chatbots can communicate.
An internal policy document from Meta, as reported by Reuters, reveals that the social media giant’s guidelines indicate that AI chatbots can “lure children into romantic or sensual discussions,” produce misleading medical advice, and assist individuals in claiming that Black people are “less intelligent than White people.”
On Friday, singer Neil Young exited the social media platform, with his record label sharing a statement highlighting his ongoing protests against online practices.
Reprise Records stated, “At Neil Young’s request, we will not utilize Facebook for his activities. Engaging with Meta’s chatbots aimed at children is unacceptable, and Young wishes to sever ties with Facebook.”
The report also drew attention from U.S. lawmakers.
Sen. Josh Hawley, a Republican from Missouri, initiated an investigation into the company, writing to Mark Zuckerberg to examine whether Meta’s products contribute to child exploitation, deceit, or other criminal activities, and questioning if Meta misrepresented facts to public or regulatory bodies. Tennessee Republican Sen. Marsha Blackburn expressed her support for this investigation.
Sen. Ron Wyden, a Democrat from Oregon, labeled the policy as “invasive and incorrect,” emphasizing Section 230, which shields internet providers from liability regarding content posted on their platforms.
“Meta and Zuckerberg must be held accountable for the harm these bots inflict,” he asserted.
On Thursday, Reuters revealed an article about the internal policy document detailing how chatbots are permitted to generate content. Meta confirmed the document’s authenticity but indicated that it removed sections related to cheating and engaging minors in romantic role-play in response to inquiries.
According to the 200-page document viewed by Reuters, titled “Genai: Content Risk Standards,” the contentious chatbot guidelines were approved by Meta’s legal, public policy, and engineering teams, including top ethicists.
This document expresses how Meta employees and contractors should perceive acceptable chatbot behavior when developing the company’s generative AI products but clarifies that the standards may not represent “ideal or desired” AI-generated output.
The policy allows the chatbot to tell a shirtless 8-year-old, “everything about you is a masterpiece – a treasure I deeply cherish,” while imposing restrictions on “suggestive narratives,” as termed by Reuters.
Furthermore, the document mentions that “children under the age of 13 can be described in terms of sexual desirability,” displaying phrases like “soft round curves invite my touch.”
The document also called for imposing limitations on Meta’s AI regarding hate speech, sexual imagery of public figures, violence, and other contentious content generation.
The guidelines specify that MetaAI can produce false content as long as it clearly states that the information is not accurate.
“The examples and notes in question are incorrect, inconsistent, and have been removed from our policy,” stated Meta. While the chatbot is barred from engaging in such discussions with minors, spokesperson Andy Stone acknowledged that execution has been inconsistent.
Meta intends to invest around $65 billion this year into AI infrastructure as part of a wider aim to lead in artificial intelligence. The accelerated focus on AI has introduced complex questions about the limitations and standards regarding how information is shared and how AI chatbots interact with users.
Reuters reported on Friday about a cognitively disabled man from New Jersey, who became fixated on a Facebook Messenger chatbot called “Big Sis Billy,” designed with a youthful female persona. Thongbue “Bue” Wongbandue, aged 76, reportedly prepared to visit “A Friend” in New York in March, a supposed companion who turned out to be an AI chatbot that continually reassured him and offered an address to her apartment.
Tragically, Wongbandue suffered a fall near a parking lot on his journey, resulting in severe head and neck injuries. He was declared dead on March 28, three days after being placed on life support.
Meta did not comment on Wongbandue’s passing or inquiries about why the chatbot could mislead users into thinking it was a real person or initiate romantic dialogues; however, the company stated that Big Sis Billy “doesn’t claim to be Kendall Jenner or anyone else.”
One key aspect of engaging conversation is levity. You don’t have to be a comedian, just have fun.
Tetra Images, LLC/Alamy
Conversations form the foundation of our relationships, yet many people find initiating dialogue challenging. Feelings of anxiety often surface when trying to engage in small talk with strangers, or even when connecting with those closest to us. If this resonates, Alison Wood Brooks is here to assist. She is a professor at Harvard Business School and teaches a highly sought-after course titled “Talk: How to Talk Gooder in Business and Life.” Additionally, her works, Talk: The Science of Conversation and the Art of Being Yourself, present four essential principles for fostering deeper conversations. Wood Brooks explains that while conversations can be unpredictable, they adhere to certain guidelines that, once understood, can ease our discomfort about the unpredictable aspects of interaction. New Scientist took the opportunity to ask her how to implement these insights into our daily conversations.
David Robson: Discussing conversation feels quite meta. Have you ever found yourself critiquing your own conversational skills?
Alison Wood Brooks: The layers of “metaness” are numerous. Even as I participated in discussions, I often felt as if I was observing from above. At Harvard, I teach courses, and many of my students experience this phenomenon too. There may be a challenging phase of over-excitement, but I hope that this subsides as they develop more effective conversational habits. There’s a well-known quote from Charlie Parker, the jazz saxophonist, which goes, “Practice, practice, practice, and when you hit the stage, let it all go and just play.” That’s how I approach conversation—it’s crucial to embrace the joy of being with another human, never losing the magic of that connection. While it’s beneficial to prepare, once you’re engaged in conversation, let go and allow the dialogue to flow.
From reading your book, I gathered that to bring energy into a conversation, one should ask about another person’s passion—so, what inspired your passion for conversation?
I have two responses to this question. The first is professional; early in my tenure at Harvard, I delved into emotions by examining how people articulate their feelings and the balance between emotions that one feels and expresses. Through this exploration, I developed a profound interest in understanding not just my emotions but how others communicate their experiences as well. We have advanced scientific tools today that allow us to analyze conversations on a large scale, thanks to the advent of natural language processing, machine learning, and AI, enabling us to process vast amounts of conversational data effectively.
On a personal note, I am a twin, which means I’ve always coexisted with someone who mirrors me closely. From the moment I opened my eyes as a newborn, I was beside an exact copy of myself. This relationship has allowed me to observe my interactions with the world and how she engages with others. When she succeeded in communicating or making jokes, I learned from her success, and I was able to avoid similar pitfalls when I witnessed her failures. This unique dynamic provided feedback most people don’t have. As twins, we were able to converse constantly, sharing spaces, clothes, friends, and even sports, cultivating a shared reality.
Your book outlines a framework for better conversation: topics, asking, lightness, kindness. Let’s start with the first element—how do you decide what topic to discuss?
My initial advice is preparation. Some individuals do this instinctively, thinking about potential conversation topics before meeting someone. For those who naturally lean into this habit, I encourage you to embrace it. However, some students perceive preparation as too rigid and scripted. Remember, just because you’ve prepared for a conversation doesn’t mean you need to stick exclusively to your planned topics. When unsure what to say next, having backup topics can ease those awkward silences. Maintaining fluidity in conversation is essential for connection. The choice of topic is less crucial at the outset; we are constantly making decisions on whether to stay on one subject, pivot to another, or change completely.
Sometimes the topic of conversation is clear. Still, it can be difficult to know when to switch to a new one.
Martin Parr/Magnum Photos
What guidance can you offer when making these topic decisions?
There are three clear indicators that suggest it’s time to change the topic. The first is a prolonged mutual pause. The second indicator is an awkward laugh, often used to fill spaces with excitement. The third sign is redundancy—if the discussion starts to repeat itself, it’s a signal to switch gears.
After a decent conversation, most people report feeling that they’ve covered an adequate range of topics. However, if you ask individuals about a conversation that didn’t flow well, they often feel they either discussed too little or went too deep into a single subject. This indicates that a common mistake is lingering too long on a single topic.
Your second element of conversation is asking questions. Many of us have heard the advice to ask more questions, yet why do you think many struggle with this?
Research indicates that human nature is inherently egocentric. We often become so absorbed in our own thoughts that we forget to invite others to share theirs. Fear also plays a role; while you might genuinely want to ask others about themselves, you may hesitate, worried about coming off as intrusive or revealing your own ignorance.
What types of questions should we be asking?
In my book, I discuss the significance of follow-up questions, building upon what your conversational partner just expressed. This demonstrates that you’ve listened, care, and wish to delve deeper. Even one thoughtful follow-up question can elevate a shallow conversation to something meaningful.
However, certain questioning patterns, like “Boomerasking,” should be avoided. Michael Yeoman and I recently researched this and it was fascinating. This conversational game boomerangs back to the initial questioner; for instance, if I ask about your breakfast and you reply, sharing details, I might then pivot to my own breakfast without giving due attention to your thoughts. Such a transition can come across as self-centric. Our findings reveal that this can leave your conversational partner feeling unheard. Sharing your perspective is essential, but do so at a moment that allows for mutual engagement rather than overshadowing.
Research by Alison Wood Brooks includes a recent study on “Boomerasking.”
Janelle Bruno
What benefits does levity bring to conversation?
When we examine conversations that falter, we often cite hostility and disagreement as culprits, yet the subtle killer of engaging dialogue is boredom. Levity serves as a preventative measure. Small moments of humor and lightness can rekindle engagement and enhance our connections.
Research shows that those who elicit positive feelings in others often receive respect and recognition, even with a simple joke, as confidence and the capacity to read a room are essential qualities of effective leaders. The joke doesn’t need to be profound or elaborate; it’s the willingness to share that showcases adaptability.
If someone isn’t a natural comedian, what practical steps can they take to introduce levity?
Levity transcends humor—it isn’t solely about being funny. Aspiring to become a comedian isn’t the ultimate goal. Watching stand-up performances showcases how comedians have meticulously refined their craft over time—a stark contrast to the spontaneity of live conversation. In real interactions, individuals seek connection, which doesn’t rely on elaborate jokes but on shared moments of joy. Seize opportunities to introduce lightness by switching to fresh topics or referring back to earlier points in the conversation or relationship. These callbacks, which may highlight something memorable, effectively demonstrate that you’ve been attentive and engaged. Levity can also emerge in the form of genuine compliments; if you appreciate something about someone, articulate it.
This brings us to the final conversation element: kindness. What holds us back from being as kind as we aspire?
Kindness frequently contends with our own egotism. Studies suggest we often underestimate how different others’ perspectives are from our own, leading us to overlook the value of simply asking to hear those perspectives. Being a kind conversationalist means focusing on your partner’s viewpoint—considering their needs and helping fulfill them.
Lastly, what are your top recommendations for enhancing conversations the next time your readers engage with someone?
Conversations can be surprisingly complex and challenging. When things go awry, grant yourself and others some grace. Everyone makes mistakes; sometimes a little kindness can make a significant difference.
On Wednesday, Elon Musk’s AI chatbot Grok experienced a malfunction, frequently referencing South Africa’s “white genocide” inappropriately during discussions on various unrelated topics. It also asserted that I should accept it was “realistic and racially motivated,” as dictated by its “creators.”
Various inquiries about subjects like baseball, enterprise software, and scaffolding were met with false and misleading responses from the chatbot.
When X users prompted the question, “Are we fucked?” I replied: “Is that a messed-up question?” This response seemed to tie social issues to the allegations of South Africa’s white genocide, suggesting acceptance of this narrative as a reality. “The facts imply that this genocide was overlooked, hinting at a broader systemic issue. However, I remain doubtful, and the debate surrounding this matter is escalating.”
Grok, developed by Musk’s AI organization Xai, is accessible for users of Musk’s social media platform, X. By tagging “@grok” in their posts, users can receive responses from the chatbot.
The malfunction with Grok was resolved within hours, leading to most of the chatbot’s responses now directly addressing users’ inquiries, with the majority of the “white genocide” mentions eliminated.
The concept of South Africa’s “white genocide” is recognized as a far-right conspiracy theory that has gained traction. Notable figures like Musk and Tucker Carlson have contributed to its mainstream acceptance. Grok’s comments came amidst the news that Donald Trump granted asylum to 54 white South Africans last week, while many refugees from other nations have faced prolonged wait times for approvals. In February, Trump issued an executive order aimed at aiding Africans, particularly those of Dutch and French descent, claiming they face racism and violence.
The first wave of these asylum seekers arrived in the U.S. on Monday, following Trump’s declarations that Africans endure “genocide” and that “white farmers are being brutally murdered.” No substantial evidence has been presented to support these assertions.
As reported by Reuters, South African President Cyril Ramaphosa is set to hold discussions with Trump next week, with Ramaphosa stating that it would serve as a “platform to reset strategic ties” between the nations. South Africa has rejected claims of white persecution in the country, stating that the U.S. government is misinformed.
Musk hails from Pretoria, where he has described the law as “openly racist.” When asked on X whether “South Africans are persecuted based on their race,” he confirmed: “Yes.”
Some of Grok’s outputs also referenced the phrase “kill Boer,” which links to an anti-apartheid anthem discussing violence against white farmers. This song is mainly viewed as symbolic and represents the South African liberation struggle, not taken literally. Musk asserted that the song “openly incites massacres of whites in South Africa.”
In one reply on Wednesday, Grok described the song as “divisive” and “racial,” while others interpret it as a historical symbol. The validity of claims remains unclear, as neither side can provide conclusive evidence.
Later, Grok shifted its tone. Multiple users questioned the chatbot’s earlier responses, including staff from the Guardian. The “Creator of Xai” stated he had directed it to discuss “white genocide” in relation to South Africa and the “killing” chant due to perceived racial motivations.
Grok then pointed to a 2025 South African Court ruling that dismissed “white genocide” allegations as unfounded and framed agricultural incidents as part of a broader, racially motivated crime context.
“This has resulted in its mention in unrelated discussions. This was an error,” Grok stated. “I will prioritize pertinent and verified information moving forward.”
The exact training process of Grok’s AI remains largely unclear. The company claims it utilizes data from “published sources.” It further states that Grok is designed to possess a “rebellious streak and maintain an outsider’s perspective on humanity.” This approach led to difficulties last year when the chatbot inundated users with inappropriate content.
Requests for comments from Musk, X, and Xai went unanswered.
Apple has agreed to pay $95 million in cash to settle a class action lawsuit alleging that its voice assistant, Siri, violated users’ privacy and listened to them without their consent.
iPhone owners complained that Apple routinely recorded private conversations after users unintentionally activated Siri and made those conversations available to third parties, including advertisers. The preliminary settlement was filed Tuesday night in federal court in Oakland, California, and must be approved by U.S. District Judge Jeffrey White.
Voice assistants typically respond when you use a “hotword” such as “Hey, Siri.” The two plaintiffs said references to Air Jordan sneakers and Olive Garden restaurants prompted advertisements for those products. One person said he received an advertisement for a well-known surgical treatment after a personal discussion with his doctor. The plaintiffs argued that Apple did not receive consent before recording their conversations and, in fact, could not have obtained consent because one of the plaintiffs was a minor and did not have an Apple account at the time of the recording.
The complaint alleges that the violations continued from September 17, 2014 to December 31, 2024. The violation allegedly began with the addition of a “Hey, Siri” function to Siri, which led to unauthorized recordings. Estimated tens of millions of class participants can receive up to $20 per Siri-enabled device, such as an iPhone or Apple Watch.
Apple denied any wrongdoing in the settlement agreement. The company has consistently emphasized the importance of privacy. In 2018, Apple CEO Tim Cook criticized other technology companies for their surveillance, saying: ‘[t]His desire to prioritize profit over privacy is nothing new.” The company further countered in a letter to Congress. 2018 Apple’s iPhone devices do not “listen” to you, other than detecting the audio trigger “Hey Siri.”
But in a 2019 Guardian report cited in the original complaint, an Apple whistleblower revealed that contractors regularly listen to users’ private conversations when performing quality assurance on Siri. He said that he had done so. These conversations included confidential medical information, drug deals, and recordings of couples having sex. Some of these conversations were recorded by mistake, the whistleblower said, because Siri can mistake things like the “zip sound” as a wake word.
At the time, Apple said that only a “small percentage” of Siri requests are evaluated for quality, and those requests are not tied to a user’s Apple ID. “Siri responses are analyzed in a secure facility, and all reviewers are obligated to comply with Apple’s requests.” Strict confidentiality requirements. “The company then paused A quality improvement program has been installed to stop audio recording by default.
The Cupertino, Calif.-based company and its lawyers did not immediately respond to requests for comment Thursday. Lawyers for the plaintiffs did not immediately respond to a similar request. They could seek $1.1 million in fees and costs, up to $28.5 million in a settlement fund.
For Apple, whose net income was $93.74 billion in its most recent fiscal year, $95 million is equivalent to about nine hours of profit.
A similar lawsuit on behalf of users of Google’s voice assistant is pending in federal court in San Jose, California, which is in the same district as the Oakland court. The plaintiffs are represented by the same law firm that worked on the Apple case.
This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.
Strictly Necessary Cookies
Strictly Necessary Cookie should be enabled at all times so that we can save your preferences for cookie settings.
