-Kerr suggests that preferred pronouns have a similar effect, that "a conflict between what we see [...] and what we are expected to say, affects us." As an exercise, she suggests (privately!) translating sentences about transgender people to use natal-sex-based pronouns, and honestly asking oneself: "Do you feel differently, on reading it this way? Do you react differently?"
+Kerr suggests that preferred pronouns have a similar effect, that "a conflict between what we see and know to be true, and what we are expected to say, affects us." As an exercise, she suggests (privately!) translating sentences about transgender people to use natal-sex-based pronouns.
+
+Unfortunately, I don't have a study with objective measurements on hand (let me know in the comments if you do!), but I think most native English speakers who try this exercise and introspect—especially using examples where the trans person exhibits features or behavior typical of their natal sex—will agree with Kerr's assessment: "You can know perfectly the actual sex of a male person, and yet you will still react differently if someone calls them _she_ instead of _he_."
+
+Let's relate this is Yudkowsky's specialty of artificial intelligence. In a post on ["Multimodal Neurons in Artificial Neural Networks"](https://openai.com/blog/multimodal-neurons/), Gabriel Goh _et al._ explore the capabilities and biases of the [CLIP](https://openai.com/blog/clip/) neural network trained on textual and image data.
+
+There are some striking parallels between CLIP's behavior, and phenomena observed in neuroscience. Neurons in the human brain have been observed to respond to the same concept represented in different modalities (_e.g._, [Quiroga _et al._](/papers/quiroga_et_al-invariant_visual_representation_by_single_neurons.pdf) observed a neuron in one patient that responded to photos and sketches of actress Halle Berry, as well as the text string "Halle Berry"), and so do CLIP neurons. Futhermore, CLIP is vulnerable to a Stroop-like effect where its image-classification capabilities can be fooled by "typographic attacks"—a dog with instances of the text "$$$" superimposed over it gets classified as a piggy bank, an apple with a handwritten sign saying "LIBRARY" gets classified as a library. The network knows perfectly what dogs and apples look like under ordinary circumstances, and yet still reacts differently when presented with clashing textual labels.
+
+I conjecture that the appeal of subject-chosen pronouns lies _precisely_ in how they exert Stroop-like effects on speakers' cognition. (Once again, if it were _actually true_ that _she_ and _he_ had no difference in meaning, _there would be no reason to care_.) [Pronoun badges](/2018/Oct/sticker-prices/) are, quite literally, a typographic attack against native English speakers' brains.
+
+Note, I mean this as a value-free description of how the convention _actually functions_ in the real world, [not a condemnation](https://www.lesswrong.com/posts/N9oKuQKuf7yvCCtfq/can-crimes-be-discussed-literally). One could consistently hold that these "attacks" are morally good—
+
+
+Is susceptibility to Stroop-like effects an indication of bad mind design? I mean, maybe! You could argue that! One would expect that an _intelligently_-designed agent (as contrasted to messy human brains coughed up [blind evolution](https://www.lesswrong.com/posts/jAToJHtg39AMTAuJo/evolutions-are-stupid-but-work-anyway) or [lucky](https://www.lesswrong.com/posts/dpzLqQQSs7XRacEfK/understanding-the-lottery-ticket-hypothesis) neural networks found by gradient descent) could easily bind and re-bind symbols on the fly:
+
+