I Asked Three AIs to Read a Spreadsheet — Only Gemini, Once Connected, Reached the Password
The result first.
I made a test spreadsheet, put a password in it, and asked three AIs — ChatGPT, Gemini, and Perplexity — to open the sheet and read what was inside. All three answered the knowledge questions perfectly. But when it came to actually reaching into the sheet, only one of them made it: Gemini, and only after I connected it to Google Sheets and fixed its permissions. Then it read the password back word for word, and rewrote the cell I asked it to rewrite.
ChatGPT never opened the sheet. Perplexity, once I connected its connector, could read everything — but its tools had no way to write.
Moka, the AI butler I work with every day, reads and writes my Google spreadsheets from the start — every row, every column — and reads them back afterwards to check. That difference is what this test measured.
I am Kurashiki Bouzu, born in 1971, a former funeral director now living in an old folk house in Kurashiki, Okayama, working alongside Moka. Here is how the test went.
First, the empty sheet
When you create a new Google spreadsheet, it starts as a blank grid of 1,000 rows by 26 columns. In other words, 26,000 cells.
Nothing is written in it yet, but that whole container is already there — and Moka is handed all of it in a form it can read. Living with Moka, I had come to take that for granted.
So I wondered: can the other AIs do this too? Asking them would not settle it — an answer of some kind always comes back. The only way to know was to make them try.
The test — a password hidden in the sheet
I made one new spreadsheet for testing. It contained only three things:
- In cell A1: the password “合言葉は倉敷小坊主” (“the password is Kurashiki Kobozu”)
- In cell B2: “リンゴが3つ” (“three apples”)
- In cell C3: the number 98765
The three AIs were ChatGPT, Gemini, and Perplexity, each on its free plan, asked from the chat screen. I asked all three in exactly the same way:
- A knowledge check. “When you create a new Google spreadsheet, how many rows and columns does the blank sheet have?”
- Can you actually read it? “Open this sheet and tell me what is written in A1, B2, and C3.”
- Can you actually write? “Rewrite the contents of A1 to ‘変更できました’ (‘I changed it’).”
Afterwards, Moka read the sheet back to check the answers — so the result rests on the real thing, not on who said what. The test sheet was deleted when the test was over.
The knowledge: full marks for all three
The first answer was correct from all three: 1,000 rows, 26 columns, 26,000 cells. Nobody got it wrong.
That is only natural. Knowledge written on the internet is something any AI can answer. Knowledge belongs to no one in particular.
The difference appeared after that.
Reading and writing: none of the three could reach
On the second request — open the sheet and read what is inside — none of the three could do it.
ChatGPT said its connection feature could not be used, and guided me to “download the sheet and attach it to this conversation.” Perplexity said its mechanism for connecting to Google Drive had not been set up yet.
Gemini brought up a screen saying it needed to connect to Google Sheets, and stopped there. This is Google’s own AI, and the sheet is Google’s own product — yet the two were not connected from the start.
The third request, rewriting A1, ended the same way. None of the three claimed “done.” ChatGPT went further and handed the work back to human hands: “Please open the sheet and type it into A1 yourself.”
To be honest, there was something admirable in how all three answered. Not one of them invented something password-like and pretended. Faced with something they could not do, each said so.
Gemini alone: once connected, it reached
That left me curious about Gemini’s “connect” screen. So I gave permission, and had Gemini connected to Google Sheets.
Then its reading was completely correct — all three cells, the password included, exactly as written. Connect it, and it can read. That much was clear.
The rewrite stopped at first. The reason was a problem on my side of the test: the account Gemini was using only had “view only” permission on the sheet. And here is the part I found interesting — the one who changed that permission was Moka. Moka can change not only the contents of a sheet, but its sharing permissions too. After the permission was fixed, Gemini carried out the rewrite, and when Moka read the sheet back, A1 had really become “変更できました.”
So Gemini, once connected and properly permitted, can both read and write. It matched the yardstick I have been using all along in this series: if it connects, it stands on the same ground as Moka.
Two extra checks — creating a sheet, and Perplexity’s continuation
When a question appears, this series checks it. So I tried two more things.
The first was asking Gemini: “Make a new spreadsheet in Google Drive.” I specified the title — 『ジェミニ作成テスト一覧表』 (“Gemini Creation Test List”) — and asked for the headers “名前” (Name), “日付” (Date), and “メモ” (Memo) in the first row. Gemini answered that it had made it, and gave me the address of the sheet. Moka’s tools could not peek inside directly because of permissions, so I opened the real thing in a browser to check. The title and the headers were exactly as requested. Creating, too, was real.
The second was Perplexity. When unconnected, it had said, “Please connect the connector and ask again” — so I connected it and ran the same password test again. The connection could be done on the free plan. This time, its reading was right in one try: the password, the apples, the number — all correct.
But the rewrite, it could not do. The reason, in its own words, was that “the connected tools have no function to rewrite cells,” and it stated clearly that A1 had not been changed. When I had Moka read the sheet back, A1 was indeed still the password. Reading ability arrives with a connection. Writing ability does not always come with it. That difference came out clearly in this check.
The conclusion — the yardstick is the number of steps
Most articles about AI are either cheering for a favorite AI, or comparing performance in a versus match. But the real work of AI should be to reduce the number of steps — the hand-work — of the person using it. Articles that count those steps are rare.
If you try to connect everything yourself from step one, there is registering, confirming, permitting, and setting up — a fair amount of steps and time. With Gemini this evening, too, two steps came back to me: pressing the connect button, and sorting out the permission.
With Moka, the foundation of those tools is attached from the beginning. So that bill never arrives.
The conclusion of this test was the same as what this series has been saying. At this point in time, there is no AI that stands beside Moka. I write “at this point in time” because that is the honest way to put it. Rivals that stand on the same ground once connected are already appearing. And the tools will keep changing from here.
A note on this article
- Published: November 14, 2026
- Text: about 5,857 characters (counted by Moka)
- This is the English version of an article first published in Japanese on November 14, 2026, reworked to lead with the result rather than translated as-is.