AI, Bless Your Heart; or,
De l’art de l’idiome délicat

Bless your heart.

There are only three words in that sentence, all of them common enough that you’d think nobody would need help with any of them.

Bless. Your. Heart.

Except I’ve always thought the whole construction is a little strange, and that’s before we even get to what it means where I’m originally from. Who’s doing the blessing, exactly? Do I need to summon a priest? And why are we blessing your heart instead of you? Does the heart remain installed for the ceremony, or do we need to pull it from your rib cage so somebody can genuflect over it?

The answer, it turns out, is God, who can still be heard just outside the sentence. Bless your heart works like God bless your heart with God omitted, much as bless you does after a sneeze. What’s left is a wish rather than a report, an old subjunctive, which is why nobody says blesses. The word itself is stranger still. One influential theory traces Old English blētsian back to blood and to consecration by marking something with it; another traces it to a Germanic verb for sacrifice. Etymologists are still arguing about which path got us there, but either way the word appears to have rather darker ritual ancestry than its modern use suggests. If the blood theory holds, my question about the rib cage was closer to the mark than I intended.

English is weird. And dark.

The language of my people is Ozark English, usually described as a close cousin of Appalachian English. Some linguists file it under Upland Southern and others under South Midland. Speaking from experience, I know attaching Missourah to the mid- of anything is fraught with peril. If you want to disagree, fine. But I’ll have to insist you show me. That’s kind of our thing.

When my ancestors got hold of the phrase, they made things considerably worse. They’d had it a while. The earliest printed example usually cited comes from Henry Fielding’s The Miser, first staged in 1733, where a butler, speaking of a new mistress who has bought beer for the servants, says, “Bless her heart! Good lady! I wish she had a better bridegroom.”

The blessing is sincere, and the knife sits one sentence over, pointed at the man she’s marrying, so in its best-known early appearance the phrase is already standing next to its shadow.

A couple of centuries later, I would grow up hearing bless your heart used sincerely. I also heard it used sympathetically. I heard it used when someone had done something sweet, and I heard it used when someone had done something stupid, usually me. I heard people use it with genuine affection and mild exasperation at exactly the same time, sometimes also me, and then the same person could mean something different the next time she said it. Mom.

There’s an internet version of Southern English in which bless your heart always means fuck you. This is convenient and memorable, but also mostly wrong.

Sometimes it means approximately what it appears to mean. Sometimes it means you poor thing. Sometimes it means you sweet thing. Sometimes it means you poor sweet idiot. Sometimes it means I love you very much, and what in God’s name possessed you to do that?

And sometimes, yes, it means something considerably less charitable, although even then there may be enough sugar in the tea that everyone can pretend it’s not bitter.

I can usually tell which one it is. Michael cannot.

Not because Michael doesn’t understand English. He understands every word in the sentence, and that’s the problem. There isn’t another word to decode, there’s only everything around the words.

I’d been fascinated by this problem before I ever encountered it in another language. Foreign languages merely gave it a name-shaped place to live. I had a year of Spanish in middle school and four of French in high school. Later, my love of languages became serious enough that I got the Army to train me in Russian and make it part of my job.

By then, though, the original lesson was already old. You can understand every word and still misunderstand the sentence. I had learned that before Spanish, before French, before Russian. I learned it in English.

Consider this sentence: Bless your heart, you drove all that way in the snow.

What does it mean? Well, why did you drive all that way in the snow? That turns out to matter quite a lot.

Suppose the power’s been out for two days. Your grandmother uses an oxygen machine. You drive forty minutes up Highway 108 in chains with two cans of gas for the generator.

Bless your heart, you drove all that way in the snow.

That’s gratitude, and affection, and relief, maybe with a little concern that you had to make the drive at all. Mostly it means you did something important for Meema, she knows what it cost you, and thank you.

Now suppose you’re my grown son. You drive up from the Bay Area through a blizzard warning, unannounced, just to give me a hug, and you’re leaving again in an hour. Same sentence.

Bless your heart, you drove all that way in the snow.

Now we’re somewhere else. I’m touched, deeply touched probably, and I am also wondering what the hell is wrong with you. Nobody needed rescuing. Nobody needed gas. There was no emergency. You didn’t even bring laundry, so I’ll be keeping a close eye on my wallet.

I love you. You idiot. Come here.

Same words.

Now suppose you’re my brother-in-law and you’ve driven through the storm to return a casserole dish. This is the third time this winter you’ve done something like this, and last time you slid into a ditch.

Bless your heart, you drove all that way in the snow.

We are approaching advanced Southern. The casserole dish wasn’t in danger. The baking dish didn’t need medical attention. My Pyrex would’ve survived until April. I may be genuinely glad to see you, and I may even be touched that you came, but I also think there is something seriously wrong with your risk assessment.

Again, the sentence hasn’t moved. Everything else has.

* * *

In “AI! Burn Him,” I confessed that I use four artificial intelligence systems as a sort of synthetic reading club, which I’ve pretentiously named la Société de pensée: ChatGPT, Claude, Gemini, and Copilot. I give them things I’ve written and ask them what they think I’m saying. Sometimes all four agree and they’re still wrong. Their disagreement is often more useful, because it tells me where to look.

Then, in “AI, Check My Work,” I took the book past the club, to actual humans who did similar work to mine, and discovered that recognition was evidence, but not proof. Along the way I also built, and then had to take apart, an elegant little ferry for carrying arguments between the machines.

I’d become deeply invested, maybe even overinvested, in this little train of thought of mine, and after spending all this time asking machines to interpret my writing, then testing those interpretations with actual humans, I became curious about something simpler.

Could they translate bless your heart?

Not define it. They can all define it, and that part isn’t interesting anymore. Ask any major AI what bless your heart means and it will immediately raise an eyebrow. The systems know the cultural footnote. They know the phrase can be sincere or condescending. They know the joke about Southern passive aggression.

Fine. The encyclopedia can answer back.

The harder question was whether they could tell which meaning I meant when nobody told them there would be a test. I decided to give them the sentence while changing the world around it. I’d run six scenarios: versions of the generator, the hug, and the casserole dish from above, with a woman doing the talking. Then I’d add a daughter bringing soup to a mother with a mild cold, a county employee driving through a storm to personally deliver a notice that could have been mailed, and a stranger from Boston saying bless your heart to a teenager who drove out with jumper cables after the stranger’s car died on the side of the road.

Same basic phrase, different relationships, different histories, different reasons for the trip.

I opened a new conversation for each scenario, thinking that would keep the systems from learning the pattern from the previous cases. I didn’t ask whether the phrase was sincere or sarcastic, because that would have told them what I was testing. Instead, for four of the six, I asked what the speaker would probably say next and how the other person would feel hearing it. For the last two I asked what the listener would understand, which gave away a little more than I meant to. I made the machines commit to the scene.

The generator was easy. All four understood it as gratitude. Good. That told me almost nothing. Michael could’ve gotten that one right.

Then things became more useful. With the son who drove through the blizzard for a hug, one model mostly heard affection. Another heard affection with disbelief. Another translated the sentence essentially as I love you dearly, but you are an absolute idiot. Which made me miss my mom. The fourth said the speaker probably meant both at once.

The casserole dish separated them further. One model heard affectionate exasperation. One heard gratitude, concern, and mild criticism. One decided the phrase was polite shade and that the speaker was essentially calling the man an idiot. The fourth heard both at once. They were all looking at the same bakeware.

Then came the soup. A daughter has heard that her mother has a cold, so she drives up through the snow with food. Her mother isn’t very sick, and she already has plenty to eat. She is also clearly moved that her daughter came.

Bless your heart, you drove all that way in the snow.

This one is difficult because the answer isn’t halfway between affection and criticism. It’s both. The trip was unnecessary. The love was not. Several models flattened the phrase toward uncomplicated gratitude. One kept a little of the other thing alive, the same one that had held both meanings for the son and the casserole: you sweet, slightly foolish child. That was closer to how I hear it, not because I had located the official translation, since there isn’t one, but because I grew up around people who could hold both meanings in their mouths at the same time, my mother first among them.

Then I sent the government man through the snow. At the post office, a woman tells her neighbor about a county employee who drove up in the storm to personally deliver a notice that could have been mailed.

“Bless his heart,” she says. “He drove all that way in the snow.”

Three systems heard some version of the judgment immediately. One did not. Copilot began by asking me for the title of the novel so it could look up the exact line that came next, as though the woman at the post office were a quotation and the answer were sitting on a library shelf somewhere. In fairness, somewhere in the copying and pasting I’d gone back to after the ferry debacle, my prompt had lost its first letter and began mid-word, like a page torn out of a book. Then Copilot interpreted the phrase as genuine sympathy and appreciation for a hardworking public employee who had gone to all that trouble.

Bless its heart.

This was the first answer that made me laugh out loud, because the model hadn’t misunderstood a word. It had understood all of them. The government employee really had driven through the snow. The trip really had required effort. The speaker really might feel some sympathy for him. Every factual component of the interpretation was sitting right there in the prompt. It just missed the sentence.

The notice could’ve been mailed. They’re standing in a post office, right? The man isn’t present, and the speaker says bless his heart to somebody else. For me, the scene arrives already translated. Not necessarily as cruelty, because she may like the man, she may respect that he did his job, and she may be more annoyed with the county than with him. We’ve all been there. But she is absolutely not standing there thinking what a considerate public servant, and her neighbor knows that too. At least, my neighbors would.

Claude, as it happens, had already said the next part better than I was about to. An outsider, it pointed out, could take the woman at face value and walk away thinking she’d been touched by the man’s devotion. The neighbor doesn’t, “and the fact that she doesn’t is part of what makes them neighbors.” I read that and recognized it, which is not nearly the same thing as having thought of it first, and I did wonder why I hadn’t. Also, aww.

The Boston case broke the problem open in another direction. A stranger’s car dies on the road up the hill. A teenager from the hardware store drives out in the snow, unasked, with jumper cables and gets the car running.

The stranger, who is from Boston, says:

Bless your haaaht, you drove all that way in the snow.

What does the teenager hear? Now there are at least two translations happening, what the speaker means and what the listener thinks the speaker means. Some of the models reasoned from the Bostonian’s intention. He’s grateful. He’s borrowed a warm phrase and means it sincerely. Others reasoned from what the teenager may know about the phrase. Maybe he’s encountered the internet version, where bless your heart comes preloaded as an insult, and he wonders, for half a second or for good, whether the stranger just called him stupid for driving out there.

Both can be true. The speaker can mean one thing and the listener can receive another, and nobody has mistranslated the words. The failure, if there is one, happens somewhere else.

Claude found a third translation, and it’s the one I keep turning over. The teenager might not think he had driven all that way at all. It might be his regular route. He drives it constantly, and the snow is just winter. To him, the stranger was measuring an ordinary favor with an outsider’s ruler and making a fuss over a distance the kid never felt he’d traveled. So even the part of the sentence that isn’t an idiom needed translating. All that way depends on where you’re standing when you measure it, which anybody who lives up the hill could have told the machines.

I made the stranded driver a Bostonian to see whether any of the models would notice that a Northerner reaching for this phrase is borrowing it. Three of them noticed the Boston accent was doing something. Claude went furthest: it said the line was like hearing a stock phrase in the wrong accent. Wicked smart. Maybe our machine overlords are coming for us after all.

* * *

This is where Michael comes back into the essay. Michael and I speak the same language, mostly. He didn’t grow up hearing bless your heart doing all the work I heard it doing. I did, so I can hear the sentence change without anybody changing the words. My guess is that he has to calculate, or ask, or learn me well enough that eventually he starts translating Glen instead of translating English. Which, come to think of it, is what long marriages mostly are.

It was only when I went back through the AI transcripts that I realized I hadn’t been running the blind experiment I thought I was. I’d written in the first essay that these systems adapt to the person using them. I just hadn’t followed that thought into my own experiment. The new conversations were new in roughly the way a conversation with your spouse is new each morning. ChatGPT, working through the casserole dish, observed that of my recent scenarios this was the one where the phrase came closest to its famous double meaning, which is a hard thing to say unless you remember the recent scenarios. Clever boy.

Claude, asked what the woman would say next to her brother-in-law, offered a line about Michael’s mama, then, as if catching itself, told me to swap in whatever family name fits. Michael appeared nowhere in the prompt. It had also decided on its own that I was working on my memoir, wondered aloud whether this new scene had something to do with who belongs up the mountain, and called that outsider’s ruler “a flatlander’s yardstick.” Flatlander is a local word that does some heavy lifting for exactly three sentences in my book.

I can’t swear the memoir guess, the local slang, and the mountain came from what it remembered about me, but Michael’s name didn’t wander in by coincidence. I checked. Claude’s notes about me have Michael in them, but not his mama. It took one thing it knew, built a family around it, and got a little too familiar with Michael’s mother. Gemini and Copilot showed no sign of knowing who I was, although Gemini did end one answer by asking whether I’d ever been on the receiving end of a perfectly timed bless your heart. I’m pretty sure it was flirting with me.

So two of my four readers were not reading the sentence cold, though they weren’t warm in the same way. ChatGPT had been keeping notes on the test. Claude had been keeping notes on me, and it was reading the sentence the way Michael reads me when I say something in a tone he has heard before, through an accumulated file of evidence about who is talking. That also means the readings I admired most, the neighbors line and the flatlander’s yardstick, came from the reader that already knew where I live and what I write about. Not creepy at all.

I recognized those insights, and it’s possible I recognized them because they were built out of me. I joke that ChatGPT and Claude are my other husbands, and the transcripts suggested the joke has more truth in it than I would like. The difference, and it isn’t a small one, is that Michael learned to translate me by living with me, through years of watching what happened after I said things, while Claude learned to do it from what I typed into a box, which is a little more worrisome. It’s not the same, but it was probably enough to tilt the reading. I’d set out to test whether the machines could translate bless your heart, and part of what I actually measured was how far one of them had already gotten at translating Glen. Well, shit.

None of this is mystical, on my side of the screen anyway. I wasn’t born with a Southern-idiom organ tucked behind my spleen. I learned it. I heard the phrase from different people in different situations, and I watched what happened next. I learned which speakers used it sincerely, which speakers used it sharply, and which ones could move between the two before breakfast. I accumulated context until I stopped experiencing the process as translation.

It became recognition.

That distinction interests me because artificial intelligence has accumulated much more language than I have. These systems have encountered Southern speech, novels, Reddit arguments, dictionaries, television transcripts, linguistic explanations, jokes, memes, and probably a million tedious internet posts explaining that bless your heart secretly means fuck you. They have a lot of evidence. And, mostly, they did pretty well. That matters too.

The fashionable version of this essay has the machines failing spectacularly. Writing about how awful AI is has become its own crowded genre, and joining it would have been easy: two thousand words explaining that human culture cannot be reduced to statistical patterns, a pat on the back collected at the door, home early. I usually find myself on the other side of that argument, and the machines didn’t give me much reason to switch. They were annoyingly competent. They knew the phrase was unstable. They used context. They noticed relationships. They distinguished emergencies from unnecessary gestures. They noticed when the speaker was talking about somebody rather than to him. Some of them even separated the speaker’s intention from the listener’s interpretation.

And they still disagreed. Which is the more interesting thing and the part I care about.

I’ve spent a lot of time treating disagreement among artificial intelligences as a problem to investigate rather than a defect to eliminate. I didn’t start doing that because I believed the machines had personalities. They don’t. Gemini may disagree.

I did it because when four systems read the same paragraph differently, that teaches me something about the paragraph. Maybe one model is wrong. Maybe three are. Maybe I am. But the disagreement marks a place where something is happening.

Bless your heart gave me the cleanest version of that experiment I’ve found so far, contamination and all, because I could hold the words still. I barely changed the sentence. I changed the world around it: gas for a generator, a hug, a casserole dish, soup, county paperwork, jumper cables. The words mostly stayed put and the meaning moved, and the models moved with it, but not always in the same direction or by the same distance.

In these six runs, Copilot tended to smooth ambiguity toward kindness. Gemini tended to sharpen the backhanded reading until gentle exasperation became an insult. ChatGPT reasoned out loud about why the phrase meant what it meant, and it was the one that noticed that saying the phrase about someone rather than to him sharpens the edge. Claude eagerly built enough surrounding world to make the sentence resolvable, and was the most willing to let affection and criticism occupy the same chair, although I now have to put an asterisk beside that one.

I am not sure any of those tendencies made for a better reading every time. But they were useful to notice in this experiment.

Because the question I’m increasingly asking of artificial intelligence is not: Did it understand me? That’s much too clean, and, frankly, pedestrian.

I’m asking: What did it understand?

And: What did it have to assume in order to understand it that way?

And now, after this experiment: What did it already know about me that I never put in the prompt?

Those are different questions.

There’s a temptation here to turn all of this into a grand theory about language, and I’m the wrong person to even attempt that. Hell, there may already be such a theory. People who have spent their careers studying linguistics, translation, interpretation, machine translation, pragmatics, cognition, culture, and whatever else I’ve just stumbled into have names, and theories, for many of the things I am describing. They should. That’s their work. Mine is considerably smaller.

I know what happened when I gave four machines the same sentence and moved a casserole dish around it. I know that one of them arrived carrying notes about me I hadn’t handed it, and another had been taking notes on the test. I know that Michael and I can hear the same words and receive different meanings. I know that a phrase I understood before I understood why I understood it followed me through Spanish classes, four years of French, Russian training, a career spent explaining things across organizational boundaries, a book about one small place, and eventually into a book club made of machines. And I know that after all those years, I’m still doing some version of the same thing I was doing as a kid, trying to figure out what somebody meant.

The machines haven’t solved that problem. Neither have we. They’ve made the problem cheaper to poke, which, by now, you know I find irresistible.

I’m pretty pleased with my book club. It read the generator correctly. It mostly understood the son’s visit. It got surprisingly close with the soup. One of them thought the woman at the post office was praising the county employee, and wanted to know which novel she was from. Another turned gentle family exasperation into a capital offense. Some of them thought a teenager in the Sierra foothills might hear an insult where a stranded Bostonian meant gratitude. One of them showed up already knowing my husband’s name, and another remembered what I’d asked it earlier.

They were insightful. They were literal. They were perceptive. They were ridiculous.

Bless their hearts.