Hi there,

This week I have an optical illusion that will break your brain. That's probably going to be your favourite part of the newsletter. But the story I really want you to read comes after that.

My #1 story this week is an exploration of rule-breaking, and the way all our instincts for dealing with rule-breakers fall apart when the rule-breaker is A.I. Learning about this paper has changed how I interact with chatbots – specifically, how I handle those moments when I get frustrated by a truly stupid A.I. mistake. I hope you get as much out of that story as I did.

Another story this week reminds me of something my camera operator on Daily Planet, Brian Marleau, used to say: "You catch more flies with honey than you do with vinegar." He meant it in the sense that you should be kind to people, but now I have a counter-argument, an orchid that attracts flies with deadly poison!

There's also a paper this week suggesting that the type of music playing in a restaurant can change how the food tastes. Weird, right?

But I'm starting this week with a neat story about an image you've seen a million times.

Historians have long agreed that Van Gogh's Starry Night was painted in June 1889, based upon the view from his east-facing window at the asylum at Saint-Rémy-de-Provence, just before sunrise. Van Gogh's actual view didn't include a village, so that, at least, must have simply come from his imagination. It's also been assumed the stars were just kind of "made up," but a new paper suggests that assumption is wrong.

The new astronomical analysis suggests the sky on the canvas depicts the actual pre-sunrise sky outside Van Gogh's window. But the catch is, it's not the sky of June 1889. It's a sky that appeared a month later, on July 22, 1889. On that morning, a waning crescent Moon, Venus, and the V-shaped Hyades star cluster all sat inside a single narrow patch of eastern sky, the same arrangement the painting shows, with the Moon higher than Venus and the cluster between them.

Here's why historians' 19 June date doesn't hold up: The Moon that morning was 63% illuminated, well past half, not the sharp crescent Van Gogh painted, and Venus only cleared the horizon in the latter part of astronomical twilight, so he'd have had a very short window to see it. When the team simulated the sky for each night in June and July of 1889. Only 22 July fit: the Moon and Venus were 18.6° apart, almost exactly the frame's diagonal, with the Hyades in the middle. The day before, they were 29.7° apart, too wide to fit, and two days later they were 7.66° apart and no longer matched the composition.

I took my teenage son to NYC last year, and we saw the original Starry Night at the Museum of Modern Art (MoMA). This study makes me want to go back and see it again!

Clearly the music in a restaurant affects the ambience, but a new study suggests it also affects the way the food tastes.

Researchers used fMRI brain scanners to observe 48 people tasting different drinks (sugar water, lemon juice) while they heard different kinds of music. The researchers tried to pair the music to the tastes – songs that people called "sweet" paired with sugar water, songs that people called "sour" paired with lemon juice. But none of that really bore much fruit. Most of the effect really just came down to whether or not a subject was hearing something while they tasted; listening to music while tasting, made people's brains light up differently compared to tasting without music – specifically, it was brain regions that have to do with (1) taste and (2) mouth-feel that showed those differences.

It's a neat first step into a line of inquiry that could have real economic potential. I can imagine a chef insisting that their ceviche can only be tasted if Latin music is playing. But for the rest of us, maybe it's an incentive to be more methodical about making sure we put some music on when we sit down to dinner.

This illusion came across my radar this week, and I thought you'd probably get a kick out of it. It was made by someone called jagarikin (じゃがりきん), a Japanese digital artist who has built a large following on X since around 2020. Their specialty is static shapes that appear to spin, drift or jump, driven entirely by cycling colour. If you want a breakdown of why those stars are breaking your brain, I recommend this article by Mike McRae for ScienceAlert about some of that same artist's earlier work, that seems to function the same way.

A rose by any other name would smell as sweet...

...and a Bulbophyllum orchid by any other name would smell like a dead poisonous bug.

A new paper describes an orchid (Bulbophyllum clavigerum) that attracts flies, which act as its pollinators, not by making nectar to feed them, but by producing poisonous chemicals. The reason it works is that the insects normally feed on the “blood” (hemolymph) of injured bugs. Those bugs secrete those smelly poisons when they’re injured, so they typically indicate food is available to the flies. The flies don’t eat the poisons, but they're attractive to bugs because they indicate food is nearby. (You don't eat the giant Golden Arches outside a McDonalds, but they might attract you because you know they indicate delicious fast-food is nearby – same idea).

Normally, the flies sneak in and steal a hemolymph meal while a predator (spider, etc.) is munching away on an injured bug (like in the picture above). But the orchid makes the same smell, duping the flies into landing on them, instead.

Researchers are excited about this particular interaction because the orchid copies the most common smell secreted by injured bugs, even though that smell is from a toxic chemical – nasty enough, in fact, to kill the very flies it's meant to attract. In laboratory experiments, exposure to one of these chemicals, (E)-2-decenal, killed 100% of the flies tested. However, the researchers don't know whether visiting the actual flowers harms the flies, since the chemicals are likely much more diluted under natural conditions.

Nature's just so cool, sometimes. What a weird, beautiful, complex system.

Whenever you interact with A.I., there's a fundamental problem you may not have considered: We don't have an effective way to punish A.I. when it does something we don't like.

And that matters.

When models make us angry, we can correct them, and yes, they'll apologize, but they won't feel bad, so the interaction doesn't carry any emotional weight for them. A new study digs into this, and shows how our instincts in those moments set us up for failure. We get just as worked up as if a human had wronged us, so we punish A.I. just like we'd punish a person. And then afterwards, we don't get the same satisfaction we'd have gotten from punishing a person. We're left with the anger and no resolution.

Learning about this paper might change how you deal with your chatbot next time it does something that makes you angry.

The study involved a game, called "the ultimatum game." Subjects were told they'd been matched with a human or an A.I. algorithm as a partner, and that the partner has been given £0.50 to split with them however the partner chooses. Once that partner makes the offer, the subject will have the option to accept it, in which case the money will be split in accordance with the partner’s offer, or the subject can reject it, in which case neither of them gets anything.

This is a neat set-up because from an economics perspective, you should always take any non-zero amount you're offered, but people don't act like that. When people get an insultingly low offer, they get mad – so mad that they'd rather forego getting any money just for the satisfaction of seeing the person that wronged them get nothing, too. In the past, this game has always involved people, but researchers wondered what would happen if that insulting offer came from A.I. – a scenario in which a punishment like that would be meaningless to its intended victim.

A sample of 380 people played this game. Sometimes the partner offered to split the £0.50 evenly: £0.25 each. Those fair offers were accepted 100% of the time for both human partners and A.I. partners. But sometimes the offer was that the partner keeps £0.45 and gives the subject just £0.05. Human offers of that magnitude were rejected 62% of the time. And when A.I. made that offer? The offers were rejected with roughly the same frequency (58%) – even though “punishing” A.I. makes no sense at all.

The researchers did several more experiments and surveys, and were able to show that players got equally mad at rip-off A.I. compared to rip-off humans, but that after taking revenge, felt significantly less satisfied that justice had been served.

So what can we do about it? According to the researchers, the solution is simple - governments must set up the rules so that when an A.I. model causes harm, it's the people who make the models who must be held accountable. This incentivizes them to build their models responsibly, but it will also help to ensure the public feels justice can be served. (I'm looking at you, self-driving cars.)

Last Week's Trivial Answer

Last week's question was "What is feared by people with Coulrophobia?" The correct answer was "Clowns."

Thanks for reading! -Dan