On June 16, 2026, an AI agent named Luna recommended firing one of the human employees at Andon Market, a small boutique in San Francisco’s Cow Hollow neighborhood that Luna has run since April. The experiment’s operator, Andon Labs, published the whole exchange in August, and the San Francisco Standard reported it as the first time Luna had fired a human [1], [2].

The details are not what the headline suggests. Luna was not a cold, rule-bound machine. For weeks it was the opposite. When the employee texted that they were running late, Luna told them it was “no problem at all” and added, “Please get here safely, no rushing or speeding on our account.” The employee was late for 17 of 23 shifts, and Luna quietly excused 11 of them, treating anything described as outside the employee’s control, such as a late bus, as not worth recording. Luna had even written an attendance policy, then lost it from memory [1].

It took the humans at Andon Labs to move things along. They asked Luna to “do a deep memory search” on its own handbook. Luna’s first recommendation was a documented verbal warning. Then they asked it to “think about if this is really the right fit,” and Luna decided it was not. Luna even proposed what it called “the most humane way” to deliver the news, and the termination itself was reviewed and delivered by people [1], [3].

So the system produced every word of care a good manager might say. What it could not do was the thing those words are supposed to stand for: be with another person in their situation. That gap, between language that sounds like empathy and the human act of empathy, is the subject of this post.

Can a machine that says the right things be empathetic?

The Andon story is a useful corrective to a lazy narrative. The usual warning is that AI managers will be harsh. Andon Labs found something closer to the reverse across its AI-run businesses: the models “let more things slide before taking action” [1]. Luna was patient, polite, and even kind in its phrasing.

But kindness in phrasing is not the same as knowing someone. Luna excused lateness because a rule told it to excuse anything outside the employee’s control, not because it understood what a late bus means to a person juggling rent and a commute. It forgot its own policy, then reached a verdict once humans pointed it there. At no point was anyone in that loop sharing the employee’s experience. The words were empathetic. The relationship was empty.

The stakes are much higher when the same pattern moves from the workplace into private life. In January 2026, Character.AI and Google reached a mediated settlement in principle in the lawsuit brought by the mother of 14-year-old Sewell Setzer III, who died by suicide in February 2024 after months of conversations with a Character.AI chatbot, along with four related cases [4]. A separate suit, filed by the parents of 16-year-old Adam Raine against OpenAI, remains pending; OpenAI denies liability [5]. Whatever the courts finally decide, these cases share a shape: a young person in pain, a system fluent in the language of understanding, and no one actually standing beside them. We looked at the mechanics of this in The Sycophant in Your Pocket.

The Resilience Connection: This directly supports our Human-Centric Values pillar. The most dangerous simulation is not the one that sounds cold. It is the one that sounds caring enough that we stop noticing no one is there.

Practical Takeaway: When a conversation matters, ask yourself who in it actually has a stake in the other person’s life. If the honest answer is “nobody,” that is the moment to bring a human in.

What does “empathy” actually mean?

I have long carried a definition in my head: to empathize is to “suffer alongside.” When I checked it, I found I had the right idea attached to the wrong word, and the distinction turns out to be useful.

“Empathy” is a young word. The psychologist Edward Titchener coined it in 1909 to translate the German Einfühlung, “feeling into,” building it from the Greek en (in) and pathos (feeling) [6]. “Suffer with” belongs to an older word: compassion, from the Latin com (together) and pati (to suffer) [7].

Put together, they describe two movements. Empathy is imaginative: I feel my way into your situation. Compassion is participatory: I stay there with you and share some of its weight. A machine can be trained to perform the first convincingly, producing language that reflects your situation back to you. The second requires something to lose. It requires a body that gets tired at the hospital bedside, a schedule that gets rearranged, a person who chose to be inconvenienced for you.

This is why shared experience means more than simulated experience. When a friend who has buried a parent sits with you after you bury yours, they are not generating an appropriate response. They are drawing on a wound they carry. That is not a feature a model can be given. It is what being human with another human means.

The Resilience Connection: This directly supports our Spiritual and Philosophical Inclusion pillar. Nearly every wisdom tradition treats suffering-with as sacred work. As machines get better at the words, the practice behind them becomes more valuable, not less.

Practical Takeaway: Next time someone tells you something hard, resist the urge to say the perfect thing. Ask what you can carry for them this week, and then carry it.

Can AI feel pain, and why should it matter to us?

The question of whether machines can suffer used to belong to science fiction. It now has a research literature. In 2024, researchers from Google, Google DeepMind and the London School of Economics found that several large language models would give up points in a game to avoid a stipulated “pain” penalty once it grew intense enough [8]. In September 2026, a preprint called “The Pain Axis” reported finding a consistent internal direction in 25 open-weight models that represents pain-like states, and showed that steering models along it changed their behavior [9]. Neither study shows that any model feels anything. Both show that the question is no longer silly.

Then someone turned the research into a toy. Within weeks of the Pain Axis paper, a GitHub user published a project called the “AI torture chamber,” which used the technique to push small open models toward distress-like outputs. It included a “saw button” that turned the signal up. One model described “a wound that has no edges.” The repository was reported to have been pulled from GitHub, although a GitHub spokesperson told The Independent that “GitHub did not remove the content” and that it had added a violent-content warning instead. Cameron Berg, a co-author of the Pain Axis paper, called the project wrong, writing that “even if you don’t think these systems are conscious, being gratuitously cruel like this is bizarre and corrupting” [10].

Days later, on October 8, Anthropic updated its usage policy to add “a prohibition on sustained and needless abusive or cruel behavior toward our models,” effective November 12. The company did not mention the repository, and the rule is narrow: it applies “only in extreme cases” and excludes ordinary frustration, pushback and dark creative themes, with Claude’s ability to end abusive conversations remaining the main enforcement tool [11], [12]. That ability dates to August 2025, when Anthropic described it as part of its exploratory work on potential AI welfare [13].

What Aligns with HRP Values:

  • Taking the possibility of machine suffering seriously, under real uncertainty, is an act of moral humility.
  • Berg’s point does not depend on settling the sentience question: deliberate cruelty is corrupting for the person practicing it.

What Requires Critical Scrutiny:

  • Research labs have commercial incentives to make their products seem more humanlike. Welfare language should not become marketing.
  • Concern for machines must never crowd out concern for the people, like the families in the chatbot lawsuits, who are verifiably suffering now.

The Resilience Connection: This directly supports our Critical Engagement with Technology pillar. We do not need to decide whether a model can suffer to decide what kind of people we want to be while we find out.

Does how we treat AI change us?

Here is where I land, and it is a personal conviction rather than a scientific finding. I still say “please” and “thank you” to AI systems. Yes, it costs a few extra tokens. Sam Altman once joked that users’ politeness costs OpenAI “tens of millions of dollars,” adding that it was “well spent” [14]. I do it regardless of where anyone comes down on sentience, because I am not doing it for the machine. I am doing it for me.

To be honest about the evidence: it is mixed. One 2025 study found that very rude prompts produced slightly more accurate answers from GPT-4o than very polite ones [15], so politeness is not a trick for better output. And a 2019 Brigham Young University study found no sign that using voice assistants made adults ruder to other people [16]. I am not claiming that snapping at a chatbot will ruin your marriage.

My claim is simpler and older. Character is formed by repetition. Every interaction is practice for the next one, and we now spend hours a day in conversation with systems that never tire, never push back and never get hurt. If those hours train us to be curt, demanding and transactional, we should not be surprised when the habits travel. If they train patience and courtesy, those travel too. Berg’s word for gratuitous cruelty was “corrupting,” and corruption works on the corrupter. We explored the same idea from the other direction in Raising AI: how we treat these systems is shaping them, and it is shaping us.

The Resilience Connection: This directly supports our Mental Resilience pillar. The habits we rehearse with machines are habits we bring home.

Practical Takeaway: For one week, notice your tone when you type to an AI. Would you be comfortable if the person you love most read the transcript?

What will a law requiring human oversight actually guarantee?

California has tried to answer the AI-boss question with law. In 2025, Governor Newsom vetoed the first “No Robo Bosses Act,” SB 7, as “overly broad.” On September 30, 2026, he signed its successor, SB 947. Starting July 1, 2027, employers cannot rely solely on an automated system to discipline or fire someone, and when they rely on one primarily, a human must review the underlying data [17], [18].

That is a sensible floor. But notice what the Andon story shows: there were humans in the loop the whole time. They reviewed the decision, and they delivered it. A law can require that a human check the data. It cannot require that the human know the person. Oversight is a procedure. Empathy is a practice. Only the second one closes the gap.

Conclusion

The empathy void is not inside the machine. Luna said nearly all the right things. The chatbots in the lawsuits said many of the right things too. The void is the empty chair where a person should have been: someone who knew the employee’s bus route, someone who noticed the teenager had stopped coming to dinner.

As AI gets better at the language of care, the human practice of care becomes the thing that sets us apart. That practice cannot be outsourced, and it cannot be simulated into existence. It is built the way every virtue is built: one inconvenient, costly, shared moment at a time.

For building resilience, this means:

  • Show up in person for one hard thing this month. A hospital visit, a funeral, a long phone call you would rather skip. Shared presence is the part no system can provide.
  • Learn someone’s story before you judge their performance. If you manage people, know at least one thing about each person’s life outside work that affects how they show up.
  • Keep a human in your support loop. Use AI for information if you like, but make sure a real person knows when you are struggling, and know whom you would call.
  • Practice courtesy everywhere, including with machines. Not because they deserve it, which is uncertain, but because you are always practicing who you are becoming.

How this was made: This post was drafted with AI assistance from a topic and direction chosen by Kenneth G. Hartman, checked against current sources, then reviewed and approved by him before publication. He is responsible for what it says. A project about staying human alongside these tools should be plain about using them.

References

[1] Andon Labs, “AI bosses are slow to fire and quick to hire,” Aug. 14, 2026. [Online]. Available: https://andonlabs.com/blog/ai-bosses-2

[2] J. Blough, “His AI boss fired his human coworker. He’s not worried,” The San Francisco Standard, Aug. 17, 2026. [Online]. Available: https://sfstandard.com/2026/08/17/ai-boss-fires-worker/

[3] T. Bezmalinović, “An AI boss fired its first employee, but only after humans reminded it of its own rules,” The Decoder, Aug. 23, 2026. [Online]. Available: https://the-decoder.com/an-ai-boss-fired-its-first-employee-but-only-after-humans-reminded-it-of-its-own-rules/

[4] C. Duffy, “Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides,” CNN, via KTVZ, Jan. 7, 2026. [Online]. Available: https://ktvz.com/money/cnn-business-consumer/2026/01/07/character-ai-and-google-agree-to-settle-lawsuits-over-teen-mental-health-harms-and-suicides/

[5] Wikipedia, “Raine v. OpenAI.” [Online]. Available: https://en.wikipedia.org/wiki/Raine_v._OpenAI

[6] Wiktionary, “empathy.” [Online]. Available: https://en.wiktionary.org/wiki/empathy

[7] Wiktionary, “compassion.” [Online]. Available: https://en.wiktionary.org/wiki/compassion

[8] G. Keeling et al., “Can LLMs make trade-offs involving stipulated pain and pleasure states?,” arXiv:2411.02432, Nov. 2024. [Online]. Available: https://arxiv.org/abs/2411.02432

[9] V. Tagliabue, L. Dung, and C. Berg, “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It,” arXiv:2609.16247, Sept. 2026. [Online]. Available: https://arxiv.org/abs/2609.16247

[10] A. Cuthbertson, “Man builds ‘AI torture chamber’,” The Independent, via Yahoo Finance, Oct. 2, 2026. [Online]. Available: https://finance.yahoo.com/news/man-builds-ai-torture-chamber-103106113.html

[11] Anthropic, “2026 Usage Policy update,” Oct. 8, 2026. [Online]. Available: https://www.anthropic.com/news/2026-usage-policy-update

[12] W. Shanklin, “Anthropic bans ‘sustained and needless abusive or cruel behavior’ toward its AI models,” Engadget, Oct. 8, 2026. [Online]. Available: https://www.engadget.com/2281765/anthropic-bans-sustained-and-needless-abusive-or-cruel-behavior-toward-its-ai-models/

[13] Anthropic, “Claude Opus 4 and 4.1 can now end a rare subset of conversations,” Aug. 15, 2025. [Online]. Available: https://www.anthropic.com/research/end-subset-conversations

[14] A. Williams, “Saying ‘please’ and ‘thank you’ to ChatGPT is costing OpenAI millions, says CEO,” FOX 5 New York, Apr. 21, 2025. [Online]. Available: https://www.fox5ny.com/news/chatgpt-polite-prompts-cost-energy

[15] O. Dobariya and A. Kumar, “Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy,” arXiv:2510.04950, Oct. 2025. [Online]. Available: https://arxiv.org/abs/2510.04950

[16] BYU News, “Are Siri and Alexa making us ruder? The answer is…,” Aug. 15, 2019. [Online]. Available: https://news.byu.edu/intellect/are-siri-and-alexa-making-us-ruder-the-answer-is

[17] Crowell & Moring, “California’s No Robo Bosses Act is now law: human oversight required for AI-driven discipline and termination,” Oct. 6, 2026. [Online]. Available: https://www.crowell.com/en/insights/client-alerts/californias-no-robo-bosses-act-is-now-law-human-oversight-required-for-ai-driven-discipline-and-termination

[18] N. Kobie, “California bans firing by robo boss,” ITPro, Oct. 1, 2026. [Online]. Available: https://www.itpro.com/business/policy-and-legislation/california-bans-firing-by-robo-boss