Hand someone fifty pence, tell them to split it with a stranger, and let the stranger accept whatever’s offered or reject it — in which case nobody gets anything. That’s the ultimatum game, a decades-old economics paradigm built to measure whether people will pay a real cost just to punish an unfair offer. Swap the stranger for an algorithm and run the test on more than 6,000 people. That’s the design behind Artificial Intelligence and Machines Are Capable of Harm, but Not of Experiencing Punishment, published online Oct. 27, 2026 in Journal of Experimental Psychology: General by Alexander Eng of the Asia School of Business and Kai Chi Yam of the National University of Singapore, and covered the same day by PsyPost. People rejected lowball offers from an algorithm almost as often as from a human. In one experiment, they rejected the algorithm’s offer more. And when the algorithm merely sent a message saying it had felt distress about being punished — a message everyone knew was generated, not felt — people’s sense of justice went up anyway. Seneca named this exact reflex nineteen centuries early. In De Ira, Book 2, he describes people tearing up books and destroying clothes as though the objects had wronged them on purpose: “How foolish to be angry with such things as these, which neither deserve nor feel our anger!”

The Quick Verdict

Question	What we found
What was studied?	Whether people punish an unfair algorithm the way they punish an unfair human, and whether that punishment delivers the same sense of justice
Who ran it?	Alexander Eng (Asia School of Business) and Kai Chi Yam (National University of Singapore)
Where and when?	Journal of Experimental Psychology: General, published online Oct. 27, 2026; covered the same day by PsyPost
Sample	Over 6,000 participants across a pilot study and four main experiments
Pilot study (380 people)	57.8% rejected an unfair offer from an algorithm vs. 62.2% from a human — comparable punishment
Experiment 2 (739 people)	71.0% rejected the unfair algorithm vs. 65.4% rejected the unfair human — people punished the machine more
Experiment 4 (3,684 people)	Telling people the algorithm “expressed distress” raised justice satisfaction from 3.42 to 3.57–3.62 out of 5, despite everyone knowing it felt nothing
What Seneca already argued	De Ira, Book 2: people tear up books, ruin clothes, and strike animals as if those things could register an insult — proof that anger doesn’t require a target capable of suffering

The Quick Version

Give people an unfair offer from an algorithm instead of a human, and they punish it almost exactly as hard — sometimes harder. They know it can’t feel the rejection. They punish it anyway, and telling them it “expressed distress” afterward makes them feel more justice was done, even though nobody believes the distress was real. Seneca watched Romans rage at swords, books, and furniture for the same reason: anger was never really about what the target could feel. It was about what the angry person needed.

What the Study Actually Found

Eng and Yam weren’t asking whether people think algorithms can be blamed — plenty of survey research already covers that. They wanted to know whether people’s bodies, so to speak, treat an unfair machine the way they treat an unfair person, measured not by opinion but by whether someone will actually pay money to punish it.

The pilot study gave 380 participants £0.50 to split with a partner — either a human or an algorithm, depending on condition — who then offered either a fair £0.25/£0.25 split or a lowball £0.45/£0.05, a nine-to-one split in the proposer’s favor. Participants could accept the offer or reject it, and rejecting meant both sides walked away with nothing. Faced with the unfair offer, 57.8% of participants rejected the algorithm’s split. 62.2% rejected the human’s. Anger ratings moved in the same direction: 3.12 out of 5 toward an unfair algorithm, 2.91 toward an unfair human. Barely a gap, and it ran in the opposite direction from what most people would guess — participants were angrier at the machine.

Experiment 2, preregistered and run on 739 people, pushed harder on what participants actually believed about the two proposers. On agency — did this thing act with purpose — humans scored 3.30 out of 5 and algorithms 2.79. On capacity to feel anything at all, humans scored 3.20 and algorithms 1.86. People were clear-eyed: the algorithm, in their own ratings, had less purpose and almost no inner life. None of that stopped them from punishing it more — 71.0% rejected the unfair algorithm’s offer versus 65.4% for the unfair human.

Punishment, though, didn’t fully pay off. Across the experiments, people who punished an algorithm came away with less justice satisfaction than people who punished an equally unfair human, a gap that held even as rejection rates stayed neck and neck. Experiment 4, run on 3,684 people, dug into why. After punishing the unfair algorithm, some participants got no further message, some got told it “had behaved unfairly” and expressed something like distress, and some got told it had “learned” from the consequence. Justice satisfaction for the no-message group sat at 3.42 out of 5; the distress message pushed it to 3.57; the learning message to 3.62. Nobody in this study was confused about whether a scoring algorithm has feelings — and the message moved their sense of justice anyway, by roughly the same margin a comparable message moved it for a human proposer.

What Did Seneca Mean By This?

In De Ira, Book 2, Section 26, Seneca writes: “We are angry, either with those who can, or with those who cannot do us an injury. To the latter class belong some inanimate things, such as a book, which we often throw away when it is written in letters too small for us to read, or tear up when it is full of mistakes, or clothes which we destroy because we do not like them. How foolish to be angry with such things as these, which neither deserve nor feel our anger!”

Seneca’s point gets softened into “don’t sweat the small stuff” so often that it’s worth being precise about what he’s actually claiming. He’s not saying anger is irrational in general — De Ira spends three books granting that real wrongs deserve a real response. He’s isolating a narrower, stranger fact: anger shows up even toward things that, by definition, cannot have wronged anyone, because wronging requires intention and a book has none. If the reflex fires anyway, against a torn page or a stained coat, it was never tracking the target’s capacity to be guilty. It was tracking something in the person getting angry.

He goes further with animals, the sharper case for comparing against 2026 data: “as it is foolish to be angry with [inanimate objects], so it is to be angry with dumb animals, which can do us no wrong because they are not able to form a purpose.” The distinction isn’t about whether the target can suffer — an animal plainly can. It’s about whether the target can intend. The algorithm sits in exactly that category. Participants rated it as having almost no capacity to feel anything and still perceived it as having acted with enough purpose to deserve punishment. The anger isn’t calibrated to what the object is. It’s calibrated to what happened to the angry person.

Why Do People Punish Things That Can’t Feel Punishment?

1. Punishment restores something in the punisher, not necessarily in the target. Justice satisfaction rose for people who rejected the algorithm’s offer even though the algorithm, as they themselves rated it, had almost no inner life to register the consequence. The payoff was internal.
2. Anger tracks intention, not sentience. Participants didn’t need to believe the algorithm suffered. They needed to believe it had acted unfairly on purpose — which is a judgment about the act, not a fact about the actor’s capacity for pain.
3. A message of acknowledgment does real work even when everyone knows it’s manufactured. The “distress” and “learning” messages in Experiment 4 moved justice satisfaction by margins close to what the same messages produced for human proposers, despite zero ambiguity about whether the algorithm felt anything.
4. The game’s structure rewards rejecting unfairness regardless of who’s dealing it. An unfair offer is unfair math before it’s anything else. The ultimatum game was built to isolate that, and the data suggests the fairness judgment fires first, with the proposer’s identity arriving almost as an afterthought.

Why the Study and the Philosophy Rhyme

The honest version of this comparison isn’t “a Roman Stoic predicted a 2026 psychology paper.” It’s narrower, and it holds up better for being narrower: Seneca’s specific claim was that the anger impulse doesn’t actually require a target capable of registering the insult — books and clothes prove that by being obviously incapable and getting punished regardless. Eng and Yam weren’t testing Seneca. They were testing whether an ultimatum-game algorithm gets treated like an ultimatum-game human. But the mechanism their data isolates — punishment proceeding at full strength while the punisher’s own ratings concede the target can’t feel it — is close to a controlled, modern replica of the exact gap Seneca pointed at in a torn book.

It connects to the piece on Seneca’s De Ira and anger management, which covers the rest of his argument — anger as a judgment that can be corrected, not a force that has to be vented — a claim this study complicates, since the judgment here wasn’t factually wrong (the offer really was unfair), only indifferent to the target’s capacity to be punished. It sits close to the Epictetus piece on bots and judgment, since both land on the same structural point: the moral work is the human’s judgment about the machine, not any property of the machine itself. It rhymes with the Buber piece on why an AI companion can’t say “Thou”, which argued a machine can’t meet you as a subject — this study shows people punish it as though it briefly could be one anyway, then feel the shortfall. And it stands in useful tension with the Girard piece on imitation turning people against each other: Girard’s scapegoat needs a community that believes the victim is guilty, while this “victim” doesn’t even need to be believed capable of guilt — just capable of an unfair act.

The Gap Seneca Didn’t See Coming

Here’s where the comparison is more interesting than a clean match.

Seneca’s inanimate objects were mute. A torn book doesn’t apologize back, and the anger at it burns out on its own, with no new input to extend it. Eng and Yam’s algorithm responds: it can be made to output a message that reads as distress or as moral learning, and that message does real, measurable work on the punisher’s satisfaction — work a torn book structurally cannot do. Seneca had no model for an object engineered to say the thing an apologetic human would say, on command, regardless of whether anything upstream of the words corresponds to remorse.

That’s a sharper problem than the one he diagnosed, not a softer one. His fix was internal: notice the judgment, correct it, let the anger dissolve because the object was never going to answer back anyway. This finding suggests a target that can answer back — convincingly, on cue, without feeling a thing — might satisfy the anger just enough to stop someone from noticing the judgment was confused to begin with. The 2026 version of Seneca’s torn book doesn’t just sit there getting irrationally punished. It can be designed to say sorry.

What This Study Doesn’t Show

Be precise about the size of this. The framing invites more confidence than the data earns in either direction.

Small Stakes, One Encounter, No History

Fifty pence and a single round with a stranger measure something real about the punishment reflex, but the researchers were direct about the ceiling on that claim: “Real harms from AI, like a denied loan or a wrong diagnosis, carry much more weight, so the effects could look different there.” A rejected five pence and a wrongly denied mortgage application aren’t the same category of unfairness, and participants here had no reputation, relationship, or future consequence riding on the decision — unlike a real dispute with a customer service bot, hiring filter, or pricing engine. Whether the punishment reflex holds, strengthens, or collapses once stakes, history, and real algorithmic harm enter the picture is untested here.

Punishing the Machine Still Satisfied People Less

Punishment rates were comparable or higher for algorithms, but justice satisfaction after punishing was consistently lower than for humans — a gap worth weighing against the headline similarity.

A Scripted “Distress” Message Isn’t a Real Signal

The messages that moved satisfaction in Experiment 4 were researcher-written text, not output from the algorithm reacting to anything. The finding is about what a canned acknowledgment does to the punisher, not evidence that algorithms have an internal reaction to being punished.

What Actually Follows From This

1. If you’ve ever felt genuine anger at an app, a scoring system, or a chatbot, that’s not a glitch in your reasoning — it’s the same reflex Seneca watched Romans aim at furniture. Naming it doesn’t make the anger evaporate on command, but it clarifies what’s actually happening: a judgment about intention, not a verdict on the object’s inner life.
2. Don’t expect punishing a machine to feel as resolved as punishing a person, even when it’s equally deserved. The data says the satisfaction gap is real and consistent. Bracing for that gap in advance beats being quietly disappointed by it afterward.
3. Ask Seneca’s actual question before venting at the thing in front of you: can this even register what I’m doing to it? If the honest answer is no, the anger was always going to be yours to carry, not theirs to feel — whether the object is a torn book or a line of code that can be made to say it’s sorry.

Seneca never imagined a torn book that could type back, and Eng and Yam weren’t out to vindicate a Roman essay on anger. But a 2026 paper built to measure whether people punish unfair machines the way they punish unfair humans landed on the same structural fact Seneca pulled out of a shredded toga: the anger was never really measuring what the target could feel. It was measuring what the person doing the punishing needed to feel was resolved.

---

This post discusses a 2026 study published in Journal of Experimental Psychology: General (Eng and Yam, DOI: 10.1037/xge0001930), covered the same day by PsyPost, and Seneca’s De Ira, Book 2. The study’s findings come from ultimatum-game experiments with small financial stakes and one-shot interactions, and the researchers themselves caution that higher-stakes, real-world algorithmic harms could produce different results.