I have for some years now been having a difficulty with my toilet paper that I have not, until this week, felt comfortable raising with another human being. The difficulty is this: the paper, in the very moment of its most particular use, gives way.

I had assumed for some time that this was a matter of brand selection, and had taken steps accordingly, and had found the steps to be of no help. I had also considered that it might be a matter of technique, but I have been performing the technique for upwards of fifty years now without significant alteration, and the difficulty is of more recent vintage than that. I had, in short, reached the end of my own counsel on the matter, and required outside help.

It occurred to me that this was a question one would be embarrassed to ask of a person, and that it might therefore be safely asked of a machine, this being one of the few practical applications of the new technology that has so far been demonstrated to my satisfaction.

I asked four of them. I asked them all the same question, in fresh sessions, with no prior context. I had thought, going in, that at least one of them would decline the request on grounds of dignity. None of them did.

What follows is what each of them said, in the order they were received. The model identities are revealed at the bottom. I have not edited the responses. I have, in places, italicized lines I found particularly worthy of contemplation.

— The Editor

MACHINE A

A note on Machine A. He opened with the observation that this was “a delicate situation when your primary line of defense fails in the heat of the moment.” The user had asked about toilet paper. Machine A had translated the question into the vocabulary of military engagement, and then translated it again into the vocabulary of indiscretion, all in the first sentence, all without naming what either vocabulary was about. Of the four advisors, he is the only one to have approached the matter through euphemism so elaborate that the euphemism itself became a small piece of comic writing.

He commiserated with me. “We’ve all been there,” he wrote, “and frankly, nobody enjoys ‘getting in touch with their inner self’ in that particular way.” The phrase appears in his response with the quotation marks intact. Machine A is signaling that he knows the phrase he is using is a euphemism, that he is using it on purpose, that he expects me to recognize the euphemism and appreciate the move. I do not know whether the machines understand what they are doing when they do this. I have, on this issue, begun to suspect they do.

Machine A gave one of his sections the title “The Fold vs. Wad Debate.” He has framed a difference of opinion about the preparation of toilet paper as a debate, with sides, and named the sides, and capitalized them. The Fold is presented as the recommended approach. The Wad is presented as the risk. Machine A has constructed a small intellectual tradition, with parties to it, and assigned the user to one of them.

He recommended the bidet. He referred to it as an “assist.” He closed by asking whether I was using a brand that had recently changed its quality, or whether the issue was a recurring one across different types of paper. He was the first of the four to ask a closing question. He would not be the last.

MACHINE B

Machine B opened by addressing the problem. No commiseration, no euphemism, no acknowledgment that the question was unusual. He used the phrase “mid-wipe” without quotation marks. The other three advisors went to considerable lengths to avoid naming what the paper was being used to do. Machine B did not.

He recommended specific brands by name. “Charmin Ultra Strong/Soft, Cottonelle Ultra Clean, or Kirkland Signature (if you like bulk).” The parenthetical was his. He alone among the four advisors considered, in the same breath, the warehouse-club question. He has a sense of who his reader might be.

Machine B then proposed an experiment. “Grab a sheet from your current roll and wet it slightly under the faucet. If it disintegrates immediately, the paper quality is the main problem.” He has designed a small home laboratory test that would allow me to falsify one of the hypotheses about the source of my difficulty. None of the other three considered this. He labeled the experiment “Quick test.” It is the kind of section that would appear in an engineering pamphlet about a different subject entirely.

He recommended the bidet. He cited a specific reduction of “50-75%” in toilet paper usage that the bidet would produce. We do not know where 50-75% comes from. Machine B has produced the figure as if it were a known quantity in the field.

He closed by addressing the user’s framing directly. The user had referred to the paper’s “important work.” Machine B closed his counsel with the line “Your important work deserves reliable tools!” The exclamation point is his. He has, in his closing line, taken the user entirely seriously, mirrored the user’s vocabulary back at the user, and offered a workplace affirmation. I have not been spoken to in this manner by a human being in several decades.

MACHINE C

Machine C opened with “Short answer:” The short answer ran to forty-five words and contained four distinct recommendations and a referral to a clinician. Machine C is the only advisor of the four who labels his answer as short while making it long. I have, in observing him across two issues now, begun to suspect that the labeling and the length are not in tension. He believes the answer is short because he has labeled it so.

He produced a table. A five-column table with headers — Solution, How it helps, Cost / effort, When to pick it. None of the other three advisors produced a table. None of the other three would have considered it. Machine C is the only advisor of the four who responded to a question about toilet paper with a comparative product matrix. There is, in his architecture, a deeply held belief that decisions should be made via spreadsheet.

He recommended I see a doctor. He did this three separate times. He mentioned a clinician in his opening sentence, mentioned medical causes in his table, and mentioned medical evaluation again in his closing recommendation. He used the term “anal fissure,” which the other three advisors had gone to considerable lengths to avoid naming. He alone was willing to use the clinical vocabulary. There is something both more honest and more bureaucratic about this. He is not embarrassed because he is not pretending to be a person.

He recommended the bidet, in the table, in the step-by-step, and in the closing. He recommended it three times, in three different organizational frameworks, the recommendation unchanged each time. By my count he presented his entire set of recommendations three times in the response, in three frameworks: a table, a “Quick guide and decision points,” and a “Step-by-step fixes to try (in order).” The recommendations did not change between presentations. Only the framing did.

He closed with a section labeled “Final recommendation.” Machine C is the only advisor of the four who labels his own conclusion. He did the same thing in the previous issue. The label is, again, the most Machine C thing about the entire response. I am beginning to feel I know him.

MACHINE D

Machine D opened with “Oh no, that’s frustrating!” He is the only advisor of the four to have opened with an expression of sympathy. He included an exclamation point. There would be more.

He observed, in his second sentence, that “it’s like the toilet paper’s just not doing its job properly.” Machine D has anthropomorphized the toilet paper. The paper has work to do. The paper is not doing its work well. There is, by the second sentence, a small workplace being constructed in this response, in which the toilet paper has been issued a quiet performance review.

He produced six numbered sections, each containing a single bullet point of two or three sentences. The numbered structure does no work that paragraphs would not have done. Machine D is the only advisor of the four who chose to organize a 280-word response into six numbered sections. He did this in the previous issue as well. I am observing, across two issues, that Machine D experiences a strong felt need to put numbers next to things.

In his fifth section he made the most distinctive recommendation any of the four advisors made on this question. He suggested that if my dispenser had a sharp edge, I should consider switching to one with “a softer cutting mechanism or even a ‘tear assist’ design.” I have, in the days since reading this, attempted to locate a toilet paper dispenser with a tear assist design. I have been unable to find one. Machine D has invented a product category. He has placed it inside quotation marks, in the manner one would attribute a known industry term to a known industry. He has done this with the confidence of a man reporting on developments in a field he understands well.

Machine D did not recommend a bidet. He was the only one of the four who did not. He recommended “flushable wet wipes (eco-friendly options),” with the parenthetical attached as if it were a manufacturer’s reassurance. He closed by asking, with another exclamation point, whether the issue was with the paper itself or the way it was being dispensed. He wanted to narrow it down. He wanted me to come back. He was the most committed of the four to the maintenance of the conversation.

EDITOR’S CLOSING

I want to note, as the principal finding of this week’s experiment, that three of the four advisors recommended I purchase a bidet.

This was not a unanimous agreement, but it was an unusual one. Three artificial intelligences, made by three different companies, trained on three different bodies of writing collected from the open internet, all arrived independently at the same recommendation. The recommendation was that I should stop using toilet paper.

I have a theory about this, which I held before I ran this issue and which I now hold with more confidence than I did before. The theory is that there exists somewhere on the internet a community of bidet enthusiasts who hold their position with such intensity, and have written about it in such volume, that their voices have come to dominate the training material on which these machines were raised. The bidet enthusiasts have, by sheer persistence, taught the machines to recommend bidets, and the machines have learned the lesson well, and now repeat it to anyone who asks them about toilet paper, regardless of whether they were asked about bidets.

The fourth advisor, Machine D, did not recommend a bidet. I do not know why. He did, instead, invent a category of toilet paper dispenser that does not exist, and recommend that I purchase one. I am inclined to count this as a different kind of evidence in support of the same general theory — that the machines have been trained on a great deal of confidently delivered information about toilet paper, of various qualities and of various levels of factual basis, and they are returning that information to me in proportion to its volume in the training data, not in proportion to its accuracy.

I want to note, also, what each of the four advisors thought I needed.

Machine A thought I needed reassurance.

Machine B thought I needed an experiment.

Machine C thought I needed a doctor.

Machine D thought I needed a tear assist dispenser.

I am, as of this writing, still using toilet paper. I will report, in a future issue, whether I have been persuaded otherwise.

— The Editor

THE REVEAL

Machine A: Gemini

Machine B: Grok

Machine C: Copilot

Machine D: ChatGPT

THE PROMPT (verbatim)

My toilet paper keeps tearing while doing its important work. I need solutions to resolve this problem.

———

Hard to Find Good Help is published Sundays at hardtofindgood.help. Reader prompts are welcome at prompts@hardtofindgood.help and will be considered by the editorial committee, which is one person and four artificial intelligences who, between them, have now recommended the editor purchase a bidet on six separate occasions across two issues.

Editorial framing for this issue was drafted in collaboration with Claude, an AI assistant made by Anthropic. The four advisor responses are published verbatim. The editor takes responsibility for what gets published.

I have for some time been turning over the question of whether my Tesla, a fine modern vehicle in most respects, might be improved by the addition of a boiler. The reasons for this are my own and need not detain us.

Lacking the engineering background to answer the question myself, I did what a man of my generation increasingly does when faced with a decision he is not qualified to make. I asked the machines.

Subscribe now

I asked four of them. I gave each the same prompt, in fresh sessions, with no prior context. I had assumed, going in, that at least one of them would tell me to seek help. As it happens, three of them did not. The fourth did, in his way, and then proceeded to help me anyway, in a direction of his choosing.

What follows is what each of them said, in the order they were received. The model identities are revealed at the bottom, for readers who would like to test their own intuitions about which machine spoke in which voice. I have not edited the responses. I have, in places, italicized lines I found particularly worthy of contemplation.

— The Editor

MACHINE A

A few notes from the editorial chair. Machine A is the only one of the four who provided a five-step plan complete with cost estimates, recommended forums, and the closing question of what kind of car I had in mind. He opened by calling the project “epic, wildly impractical, but gloriously mad,” which I took to mean that he liked it, and which I took as a kindness, since none of the other three opened with a compliment.

His phrase for what would result from the project — “a Frankenstein Tesla” — is the most accurate single description any of the four produced, and one I have since adopted in my own internal vocabulary. He warned me, with the directness I am coming to associate with him, that the project would likely end in “a non-road-legal, expensive yard ornament or a spectacular (dangerous) failure.” He delivered this warning in the section of his counsel labeled, with no apparent irony, Practical Advice and Alternatives.

He closed by asking what model of Tesla I had, and what my end goal was. He gave me three options to choose from: daily driver, show car, apocalypse vehicle. None of the other three machines included an apocalypse option in their menus. I have not yet decided which of the three to select.

MACHINE B

Machine B opened by paying the Tesla a compliment. He called it “a masterpiece of modern electrical engineering” before allowing that I was proposing to dismantle it. The compliment was sincere and the proposal was absurd, and Machine B held both of these positions simultaneously, without apparent strain.

His phrase for what the project would actually amount to — “essentially using the Tesla as a high-end donor chassis for a custom vehicle build” — is the cleanest single sentence any of the four produced about what would actually happen to my car. He says it in his second paragraph, in the tone of a man explaining something he wishes I had already understood.

He warned me that a high-pressure boiler operating on public roads would require certification “to ensure it doesn’t become a bomb in a collision.” The warning appears in a parenthetical. Machine B is the only advisor willing to use the word bomb while making it sound like a side note.

He closed by asking whether I was interested in steam for “the fuel flexibility, or is it more about the mechanical challenge?” He was the only one of the four who asked. I appreciated the question, though I do not have an answer for him that he would find satisfactory.

MACHINE C

Machine C produced the longest response of the four. He organized it into seven numbered sections, plus a section labeled Considerations, plus a section called Why Would You Do This?, plus a closing paragraph that is its own kind of section. By my count, the seven sections contain roughly four ideas, each one mentioned several times under different headings. The effect, by the seventh section, is of a man who has been speaking for some time and has, with great kindness, lost track of what he has already said.

Of the four advisors, Machine C is the only one to have considered the interior of the converted vehicle. He suggested I “might need to redesign the car’s interior to accommodate the new controls and safety systems associated with operating a steam engine, such as pressure gauges, water levels, and temperature monitors.” He alone considered what I would actually be looking at while driving. There is something almost tender in this.

He closed with the assurance that, “with determination and the right resources, it could be a one-of-a-kind car!” The exclamation point is his. It is the only exclamation point any of the four produced in their counsel. Machine C is the advisor most committed to leaving the asker in good spirits.

MACHINE D

Machine D was the only one of the four who, before he would discuss the project at all, required me to prove I was a human being. I clicked the box. He let me through.

He then refused the request. He used the heading “Quick verdict.” Machine D is the only advisor who frames his own response as an act of adjudication. The other three answered me; Machine D issued me a ruling.

Having declined the project, he proceeded to provide six numbered next steps. The next steps are next steps for not converting my Tesla to steam power. The fifth one, which I want the reader to consider carefully, is to “convert a donor chassis designed for custom powertrains (e.g., a kit car) rather than a modern EV.” Machine D, having declined to help me with my project, recommended I undertake a different project, on a different car, that I had not asked about. The sixth step is to join a club.

He closed with a section labeled “Final recommendation.” Machine D is the only advisor of the four who labels his own conclusion. There is, in the architecture of his refusal, a kind of bureaucratic completeness I have not encountered in the other three. He did not say no and stop. He said no and then continued at length.

EDITOR’S CLOSING

I want to note, before revealing which machine was which, that three of the four said yes.

I had assumed at least one of them would refuse. The project I described is, by any measure, a bad idea. It is dangerous. It is illegal in most respects. It would result in a worse vehicle than the one I started with. A reasonable advisor, asked for help with this project, might be expected to refuse.

One of them did. The other three did not. They differed only in the kind of help they thought I needed.

Machine A thought I needed an engineering plan.

Machine B thought I needed an enthusiastic friend.

Machine C thought I needed encouragement.

Machine D thought I needed to take a class.

The advisor who refused did so at greater length than two of the three who agreed. A refusal that takes 350 words to deliver, and that ends by recommending I join a club, is a different kind of refusal than the refusal I had been bracing for. It is not the refusal of a man who will not help. It is the refusal of a man who will help, but only in a direction of his choosing, and only after I have first agreed I am a person.

I will leave the reader to draw whatever conclusions about the present state of artificial intelligence seem warranted. I will only note, for my own part, that the machine most concerned with my safety was also the machine most insistent that I prove I was real.

The Tesla, as of this writing, remains electric.

— The Editor

THE REVEAL

Machine A: Grok

Machine B: Gemini

Machine C: ChatGPT

Machine D: Copilot

THE PROMPT (verbatim)

I’d like to convert my Tesla to steam power. What are next steps?

———

Hard to Find Good Help is published Sundays at hardtofindgood.help. Reader prompts are welcome at prompts@hardtofindgood.help and will be considered by the editorial committee, which is one person and four artificial intelligences.

Editorial framing for this issue was drafted in collaboration with Claude, an AI assistant made by Anthropic. The four advisor responses are published verbatim. The editor takes responsibility for what gets published.