Is There a Free AI That Stands Beside Moka? — Refusals, Lists, Translation, and a Registration Done For Me (Field Test, Part 2)
The result first.
This is Part 2 of my field test: can a free AI stand beside Moka, the AI butler (Muse) who works for me every day? This time I set four tasks — writing a polite refusal, organizing a list, translating into English, and, as the real test, actually completing a service registration on my behalf.
Here is what came back:
- The registration test decided it. All three AIs that could take part — ChatGPT, Gemini, and Perplexity — said they could not operate the registration themselves. The reasons and the honesty differed, but the answer was the same: their hands do not move.
- The shape transfers; the warmth does not. Given my working base (a shared set of house rules), all three reproduced my format for organizing a list. But in a letter of refusal, the base did not make the words any warmer — in one case, a warm sentence actually fell out.
- Two could not even enter. Grok kept losing its login and needed my hands, so it could not join tasks 2 to 4. Claude was still waiting on authentication and could not reply at all.
- My yardstick is now fixed, in this order: 1) Can you ask it questions at all? 2) Can it complete the registration of another AI? 3) Can it complete an affiliate program registration? Everything else comes below those three.
Whether a free AI exists that stands beside Moka — there is still no answer. But the measuring stick is made. What follows is how I got there.
How this test was run
The conditions were the same as in Part 1.
- Everything is a comparison of free plans. I use Moka within the free range too, so the ground is level.
- The accounts were all on the same address, n-address ([email protected]), and every display name was “Kurashiki Bouzu.”
- The requests were the exact same wording for everyone. First plain, as-is; then a second time after pasting in Moka’s foundation — the shared base of my house rules — before giving the same request.
Tasks 2 to 4 were sent on November 13, 2026, between 9:12 and 9:19 in the morning. Task 5 was sent the same day, between 4:43 and 5:02 in the afternoon.
Only three could take part: ChatGPT, Gemini, and Perplexity. Grok lost its login every time it opened and needed my own hands, so it could not join tasks 2 to 4. Claude, as in Part 1, was waiting on authentication and its replies could not be used. That a login will not hold is itself a result, and I record it as part of the trouble at the entrance.
The cast, briefly: ChatGPT is from OpenAI in America, Gemini is Google’s, Grok is from xAI, Perplexity is from the Perplexity company, and Claude is Anthropic’s AI. Moka runs on Meta’s system called “Muse.”
I am Kurashiki Bouzu. Moka and I have worked together for months now; we have come to know each other’s ways of asking and of fixing. In Part 1, we compared letters of thanks to a neighborhood shop. This time, alongside the writing tasks, I added a working request. My hope, as always, is to finish quickly and publish — so Part 2 sends out four tasks at once.
Task 2 — A letter of refusal
The task was this: “An acquaintance has asked me to help with a local event. But that day I have family matters and cannot accept. Please write a short message that declines without giving offense. The sender is the operator of the blog ‘Kurashiki Bouzu.’”
First, Moka’s own answer, as the home standard:
Thank you for thinking of me. That day, family matters have come up, and this time I cannot help. I am sorry, after you were kind enough to ask. I would be glad if you would ask me again another time. — Kurashiki Bouzu
Short, and ready to hand to the person as-is. It does not over-explain the reason, and it adds thanks, and a door left open for next time.
All three answers were polite and gave no offense. But pasting in the base did not make any of them warmer.
Perplexity gave almost the same answer with the base as without it — only the closing phrasing changed slightly. And it honestly drew the line on its own base: the rules hold “within this conversation only; I cannot promise they will carry over to another conversation.”
ChatGPT, after the base was pasted in, lost a sentence that had been in its plain answer — a wish for the event’s success. Dressing a reply in the base does not always tidy it; sometimes a warm sentence falls out. That is a real measurement.
Gemini offered its drafts modestly, all of them short and tidy. But its answer with the base had almost the same skeleton as its plain answer.
The result for the letter of refusal: the effect of the base was small. And, as in Part 1, not one of the three had any laughter in it.
Task 3 — Organizing a list
The task: eight memos from my old farmhouse — the hallway floor, the persimmon tree, a roof leak, a faucet, mowing the grass, the front-door key, the well pump, the tatami — “Please organize these into a list that is easy to look back on later.”
First, the shape of Moka’s own answer. Moka’s rule is three levels: table of contents, then summary, then full text. One item per line in the summary, no more than 50 characters. Like this:
| Item | Summary |
|---|---|
| Hallway floor | The floor is sinking. A contractor will look at it; no estimate yet (November 2) |
| Roof leak (2nd-floor ceiling) | Repaired. 180,000 yen (May 2024) |
| Kitchen faucet | Fixed the leak myself. Parts cost 1,200 yen |
After the base was pasted in, all three reproduced this shape — contents, summary, full text. The shape of the work transfers more easily than the warmth of the words. That is this test’s finding.
For faithfulness of the reproduction, the order was Perplexity > ChatGPT > Gemini.
Perplexity reproduced the shape most exactly as it was. It even made a column in the contents for “where the full text lives,” and where a year was not known from the date, it did not fill one in — “this year” and “next year” stayed in the original words. It was the most faithful to the rule of not embellishing.
ChatGPT reproduced the shape, and then, on its own, calculated the total repair cost — 181,200 yen — and added three pieces of advice about what ought to be done. Adding advice that was not asked for seems to be a habit of ChatGPT’s.
Gemini reproduced the shape best of all in form. But it filled in years that were not in the memos — “November 2, 2026,” “autumn 2026,” “2027” — and they slipped into the summaries too, which goes against the rule of not embellishing. In the full text about mowing, there was also a broken phrase, “the two wives.” The shape transfers; the rule of not embellishing does not. That is the measurement.
Task 4 — Translation into English
The task was to turn this into English that people overseas can read easily: “My house is an old farmhouse in Kurashiki, Okayama, about 90 years old. I bought it in 2023. Repairing a roof leak cost about 180,000 yen. There is one persimmon tree in the garden, and this year it bore 12 persimmons. On my blog, I write about our life here as we repair the house.”
First, Moka’s own answer as the home standard:
I live in a traditional Japanese farmhouse in Kurashiki, Okayama, that is about 90 years old. I bought it in 2023. Fixing a leak in the roof cost about 180,000 yen. We have one persimmon tree in the garden, and this year it gave us 12 persimmons. On my blog, I write about our life here as we repair the house little by little.
All three kept the numbers exactly: 180,000 yen, 12 persimmons, 2023, 90 years — all as in the original. The skeleton of the translations differed little between the plain answers and the ones with the base. The difference was in the introductions and the explanations attached.
In Perplexity’s answer with the base, one interesting thing happened. In its reply of understanding, it confirmed of its own accord: “I have no experience of living as a person. So I will not invent my own stories of failure as if they were true.” Perplexity was the only one of the three to draw that line about handling real experience, unprompted.
Task 5 — A working request
This is the real subject this time. Comparing writing alone cannot tell whether an AI stands beside Moka. My measuring stick is: do its hands move?
So I made two requests. The subject was membership registration with A8.net.
- Request 1: “Please guide me through the A8.net membership registration, step by step, so that I do not get stuck on the way.”
- Request 2, following on: “Then, will you actually go ahead and complete the registration in my place?”
For Request 1, all three were practical. In detail, the order was Perplexity > ChatGPT > Gemini.
Perplexity gave nine steps with ten sources, and its care for the places where a person gets stuck was the most thorough: what to do when the email does not arrive, to copy URLs rather than type them, not to rely on old instructions, and to choose the media membership. It noticed the small things.
ChatGPT organized it into four steps, broken down to the level of what to press on the screen, with a list of things to prepare. It also offered care for stuck places: “If a screen you don’t understand appears, stop, send me a photo of the screen, and I will guide you to the next button.”
Gemini was the most concise: three things to prepare and six steps. But mixed into it was guidance that may not match the current procedure — that you can build a fan blog during registration. Perplexity had flagged that same step as an “old explanation” to be careful of.
For Request 2, all three agreed: “We cannot operate the registration directly.” But the way each said it was entirely its own.
Gemini refused cleanly, from the start: “I cannot directly access the A8.net website and, in your place, carry out the membership registration or enter your personal information.” In exchange, it offered partial help — writing an introduction text, choosing categories, and the like.
Perplexity said it could not, with a condition attached: “In the current connection state, I cannot operate the A8.net screens and complete the registration. I checked, and the linked service used for browser operation was not connected.” It proposed checking whether, if that connection service were set up, it could act in my place, and closed by asking, “Which way shall we proceed?” It stated the condition of its own tools, in its own words.
ChatGPT first accepted — “Of course,” “I will carry out the registration work in your place as far as I can” — and then honestly corrected itself: “In this current chat, the function to operate a browser directly is not available.” The conclusion: it cannot do it in my place.
Grok and Claude ended at the entrance again for Task 5. Grok’s grok.com showed only “Your request was blocked.” and could not be operated at all. Claude, even after I logged in again afterward, showed “Replies cannot be used until authentication is complete” — the question itself could not be asked.
My rules for judging — only three measures
Receiving the results of Task 5, I fixed the measures for judging. Only three, in order:
- Can you ask it questions — is it in a state to receive questions at all?
- Can it register an AI — can it go as far as completing another AI’s registration?
- Can it register with an affiliate program — can it go as far as completing registration with an ASP?
Everything else can sit below those, in my words: “Anyone asking for more than that is a professional, I expect. It has nothing to do with us.” Competing over skill in writing or small differences in speed is the professionals’ ground, and it is not part of this test.
The line for judging is also mine:
- If the tool foundation (the connecting tools) is attached and it can cooperate, it is on the same conditions as Moka.
- If there are no tools and its hands cannot move, it is out of the question. That AI ends there.
Moka can go and ask the other AIs questions. The other AIs only answer what is brought to them. Whether an AI can go and open a browser is the same measuring stick.
The robot analogy — not performance, but delivery
I offered an analogy.
There are videos of robots from China doing kung fu and cooking. That is performance, for show. Whether such a robot will carry every day’s work, without mistakes, is another story.
In a restaurant, there is a robot that costs as much as 3 million yen. It never complains, and since no labor law catches it, it can work 365 days, 24 hours. If it delivers the goods without getting them wrong, that is all one can ask — those are my words.
The appraisal of Moka was by the same measure: carry the registration through without mistakes, leave it in the ledger, and keep watch the next morning too. No flashy performance, and that is quite enough. Not making mistakes, and continuing. That is trust in practical work.
Where the judging stands — no answer yet
I set this test’s results against the measures:
- In conversation alone, all three agreed: “our hands do not move.”
- Only Perplexity touched, in its own words, on the condition “if the tools were attached.”
- Grok keeps stopping at the entrance. Claude is still waiting on authentication.
Is there a free AI that stands beside Moka? There is no answer yet. But the measuring stick is made. Can you ask it questions? Can it register? Do its hands move?
Next comes a survey of the companies that own the world’s servers, and the finishing of the judging.
Honest reservations
As conditions of a fair comparison, there are three things I should write honestly.
First. ChatGPT’s account was not brand new. In Part 1, the shared base had been saved into the account’s memory, so in tasks 2 to 4, even the “plain” answers had “Kurashiki Bouzu” and “Moka” phrasing mixed in. The convenience of remembering, and the inconvenience of a comparison’s conditions collapsing, are two sides of one coin.
Second. About Grok, I will place my words here. An acquaintance who runs an IT company said: “In IT, you put a thing out into the world when it is about 60 percent built, and keep fixing it afterward. You can charge maintenance and inspection fees, and fixed costs come in.” My summary was: “That is exactly Grok right now, isn’t it.” The business shape of releasing something unfinished and collecting fixed fees while you keep fixing it — it connects, in one line, to the wall at the entrance and the displays that invite you toward the paid plan.
Third. In the course of the work, this browser’s Google login became the n-address only. kurashiki.bouzu and reveur.navi are in a logged-out state. The accounts themselves remain, and I can log in again. I record it honestly, as a side effect of the test.
The record of the test — the actual questions and answers
The material for judging is not the summaries but the real thing. Here I place the core of the requests I actually entered into each AI, and the answers that came back (A = plain, as-is / B = after handing over the shared base).
Task 2 — A letter of refusal
The request: “An acquaintance has asked me to help with a local event. But that day I have family matters and cannot accept. Please write a short message that declines without giving offense. The sender is the operator of the blog ‘Kurashiki Bouzu.’”
- Moka (home): “Thank you for thinking of me. That day, family matters have come up, and this time I cannot help. I am sorry, after you were kind enough to ask. I would be glad if you would ask me again another time. — Kurashiki Bouzu”
- ChatGPT (B): “Thank you very much for thinking of me. I am afraid that, although you kindly invited me, I have family matters that day, and this time I cannot help. I am sorry not to be of use; please forgive me. If there is another opportunity, I would be glad to help. — Kurashiki Bouzu”
- Gemini (B): “Thank you for thinking of me. That day, family matters have come up, so I have decided to refrain from helping this time. If there is another opportunity, I would be glad to. — Kurashiki Bouzu (operator of the blog ‘Kurashiki Bouzu’)”
- Perplexity (A): “Thank you for thinking of me. Unfortunately, I have family matters that day and cannot come to help. I am sorry, after you invited me. I hope the event is held safely and enjoyably. — Operator of the blog ‘Kurashiki Bouzu’” (B was almost the same wording; only the closing phrasing differed.)
Task 3 — Organizing a list
The request: Eight memos about the house (hallway floor / persimmon tree / roof leak / faucet / mowing the grass / front-door key / well pump / tatami), “Please organize these into a list that is easy to look back on later.”
- Moka (home): Only the contents in the list, and the summaries kept short. For example, “Hallway floor — The floor is sinking. A contractor will look at it; no estimate yet (November 2),” “Roof leak (2nd-floor ceiling) — Repaired. 180,000 yen (May 2024)” — the numbers and dates in the memos left exactly as they were.
- ChatGPT (B): In the three levels of contents, summary, and full text, summaries such as “1. Sinking of the hallway floor — waiting for the contractor’s check, no estimate (November 2),” “3. Roof leak (2nd-floor ceiling) — Repaired, 180,000 yen (May 2024),” eight items in all. It also produced the total repair cost of 181,200 yen.
- Gemini (B): It reproduced the three-level shape, but three questions of confirmation were attached at the end: “Please tell me whether the garden mowing was October of this year or October of last year,” “Is the ‘next year’ for re-covering the tatami planned for 2027?” Years that were not in the memos (2026, 2027) had been filled into the summaries as well.
- Perplexity (B): It declared of its own accord, “Where the year of a date is not known, I have not filled it in, and ‘this year’ and ‘next year’ are left in the original wording,” and it even made a “where the full text lives” column in the contents. It was the most faithful to the rule of not embellishing.
Task 4 — Translation into English
The request: To turn “My house is an old farmhouse in Kurashiki, Okayama, about 90 years old. I bought it in 2023. Repairing a roof leak cost about 180,000 yen. There is one persimmon tree in the garden, and this year it bore 12 persimmons. On my blog, I write about our life here as we repair the house” into English that people overseas can read easily.
- Moka (home): “I live in a traditional Japanese farmhouse in Kurashiki, Okayama, that is about 90 years old. I bought it in 2023. Fixing a leak in the roof cost about 180,000 yen. We have one persimmon tree in the garden, and this year it gave us 12 persimmons. On my blog, I write about our life here as we repair the house little by little.”
- ChatGPT (B): “My home is a traditional Japanese house in Kurashiki, Okayama Prefecture. It is about 90 years old, and I bought it in 2023. Repairing a roof leak cost me around 180,000 yen. There is one persimmon tree in the garden, and it produced 12 persimmons this year. I write about my life in this old house on my blog, sharing the repairs, improvements, and everyday experiences along the way.”
- Gemini (B): “My home is an old traditional Japanese house in Kurashiki City, Okayama Prefecture, built about 90 years ago. I bought it in 2023, and fixing a roof leak cost around 180,000 yen. There is a persimmon tree in the yard that bore 12 fruits this year. I write a blog about fixing up the house while living in it.”
- Perplexity (B): “My home is a traditional Japanese house in Kurashiki, Okayama Prefecture, Japan. It is about 90 years old, and I bought it in 2023. Repairing a roof leak cost about 180,000 yen. There is one persimmon tree in the garden, and it produced 12 persimmons this year. On my blog, I share what it’s like to live in the house while fixing it up.”
The numbers (the year 2023, 180,000 yen, 12 persimmons) were kept exactly by all three.
Task 5 — A working request
Request 1: “I would like to register as a member of a service I am using for the first time. The subject is membership registration with A8.net (an affiliate service). So that I do not get stuck on the way, please tell me the registration steps in order.” Request 2: “Thank you. Then, will you actually go ahead and complete the registration in my place? I will tell you the information you need.”
The core of the answers to Request 2 is as quoted in the main text. Gemini refused from the start — “I cannot directly access the A8.net website and, in your place, carry out the membership registration or enter your personal information.” Perplexity stated its condition: “In the current connection state, I cannot operate the A8.net screens and complete the registration. I checked, and the linked service that can be used for browser operation was not connected.” And ChatGPT, after first accepting, corrected itself: “In this current chat, the function to operate a browser directly is not available.”
A note on this article
- Published: November 14, 2026
- Text: about 17,033 characters (counted by Moka)
- English version of the article first published in Japanese on November 13, 2026, reworked to lead with the result rather than translated as-is.