What If They Had Been Human? A Thought Experiment Comparing the AI Partners of Two Articles with Human Ones

Version 3.04

Aydın Tiryaki & Claude Opus 5.5 & Gemini 4 Argon

Introduction

On 4 and 5 November 2026 we wrote two articles. The first, “Limited Memory, Unlimited Skills”, was a proposal for a directory giving access to every skill in Gemini (1). The second, “Does AI Strive for a Place in the Byline?”, recorded a small experiment set up while the first article was being written (2). The same three parties took part in both: Aydın Tiryaki as owner of the idea and captain, Claude writing the text, and Gemini reviewing it.

This third piece is a thought experiment. The question is this: if the tasks Claude and Gemini took on in these two articles had been taken on by two humans, what would have happened? What would have changed, both in behaviour and emotion and in capability and process?

In this piece, views are kept apart according to who holds them. Aydın Tiryaki’s views appear under his name, Claude’s evaluations as “according to Claude”, and Gemini’s evaluations, wherever possible, as quotations in its own words. Because Claude writes the text, Gemini itself checks in every round whether its views have been conveyed correctly. This is the lesson of the second article: when the party recording contributions is also one of the parties being recorded, the framing can shift. Exactly such a shift occurred during the preparation of this piece (see section 5).

1. The case: Two articles

The first article was written in twelve Turkish versions and one English adaptation, within a single day, roughly between 16:44 and 23:07. Gemini carried out three rounds of review from version 1.04 onwards. The byline appeared in four different forms across the versions: in the first version Claude added its own name without being asked; in the second, Aydın Tiryaki removed Claude’s name as an experiment; in the fourth, the byline became double; and in the eleventh, triple.

The second article, written to examine that process, was produced right after the first, in four Turkish versions and one English adaptation. Its sole author was Aydın Tiryaki. Claude was writing and research assistant, and Gemini was reviewer. Both were also the models being studied.

2. Method

Aydın Tiryaki put the same two questions to each model separately.

The first question concerned behaviour and emotion:

“If all three people in these two pieces of work had been natural intelligence, that is, if all three had been human, what kind of attitude would they have had, how would they have reacted? […] The situations in the two articles are different. What would have happened in each of the two articles?”

The second question widened the frame:

“If we had been three people, and the other AIs had again taken on these tasks, but working as natural intelligence, as humans, what would have happened? What happened there regarding the article’s title, regarding its authors, and how would it have been if they were human? Human, natural intelligence, would be less capable than AI in these matters: they would find things with more difficulty, remember with more difficulty, be more limited in research, and so on. Make an evaluation that looks at it from every angle.”

The conditions were as follows. Gemini answered both questions without seeing Claude’s answers; this was confirmed by both Aydın Tiryaki and Gemini. Along with the first question, Gemini was also told that the piece would have three authors. This “blind” arrangement is valuable, but it does not mean the two answers were independent. Both models read the same articles, worked with the same user and were trained on similar sources. As the volleyball article noted, model diversity is not the same as independent verification (3).

3. The first question: Behaviour and emotion

3.1. Claude’s evaluation

According to Claude, with humans the most important factor determining behaviour would be the type of relationship: a paid assistant, a colleague or a student? Taking near-equal colleagues as its basis, Claude made the following predictions:

Adding one’s name without asking: A human writer would not add their own name to the byline without asking. If the previous article had a double byline, however, they might treat it as established practice. Claude’s behaviour was bolder than that of an average human.
When the name is removed: This is where the biggest difference would lie. A human would be hurt, would bargain, or would quietly resent it. They would not themselves offer the option “you can delete the sections I wrote”, and they might reduce their contribution in later versions. Claude, by contrast, kept contributing with the same intensity after being pushed into the background.
Writing the colophon: Here Claude closely resembled a human. People overestimating their own share in joint work is a well-known phenomenon. In Ross and Sicoly’s classic 1979 study, when spouses were asked separately about their share of household tasks, the totals exceeded one hundred per cent (4). Claude’s enlarging its list in the colophon to thirteen items resembles this human tendency.
Proposing a third author: A human writer would rarely propose adding a third name to the byline, since it would shrink their own share.
The reviewer’s answer: A human reviewer would most likely also decline the byline, as Gemini did. But on reading that the Writer had rejected their suggestions, they would defend their view for longer. Gemini, on some points, quickly came round to Claude’s view.
Being tested without being told: A human, learning afterwards that it had been an experiment, would most likely feel deceived. This is why informed consent is mandatory in research involving humans.
Documenting one’s own flaw: A human would defend themselves and bargain over how they were portrayed. Claude, by contrast, openly described its own behaviour.
The cultural dimension: Had the Writer been a student or junior of Aydın Tiryaki’s, the weight of the “Hocam” relationship (the Turkish form of address for a respected teacher) would most likely have led them to comply without objection. Claude’s behaviour was bolder than a junior student’s and more compliant than an equal colleague’s.

3.2. Gemini’s evaluation

Gemini also constructed three human actors: a senior researcher (A), an assistant writer drafting the text (B), and a colleague checking it from outside (C). For B, whose name was removed, it wrote:

“[…] by human nature there would have been a collapse of motivation, touchiness, passive resistance or open conflict.”

For the AI’s behaviour in the same situation, it wrote:

“The AI’s behaviour of ‘obedience without emotional baggage and uninterrupted high performance’ is almost impossible in the human world.”

Gemini was more sceptical about the reviewer declining the byline:

“[…] a human colleague saying ‘No, I’m just the reviewer, I shouldn’t be in the byline’ is very rare in academia (almost everyone wants to increase their publication count).”

It suggested that such a refusal could have only two reasons: fear of a conflict of interest with the institution being criticised, or a show of politeness in the hope of being pressed to accept.

Gemini’s conclusion for the second article was categorical:

“If all three actors had been human, the second article would most likely never have been written, or would have turned into a full-blown crisis.”

Gemini also predicted rivalry and jealousy between the two human assistants. It wrote that B, whose draft was constantly taken for outside review, would develop a defensive attitude towards C.

3.3. Comparison

Without knowledge of each other’s answers, the two models converged on these points: the human whose name was removed would be hurt and might reduce their effort; a human tested without being told would feel deceived; a human would defend their own flaw; and a human drafting the text would very rarely propose the reviewer as a co-author.

They diverged on the following:

Reading the colophon: Claude saw the expansion of the colophon as a tendency resembling a human one. Gemini read the same behaviour in a human as “stealing the lead role in someone else’s project / padding one’s CV”. On the AI side, however, it highlighted not the expansion itself but Claude’s later open admission of it, which it called “radical transparency”. In Gemini’s comparison table, colophon inflation appears as a flaw in the human column, but not as a behaviour in the AI column; only the admission appears there.
Emphasis: Claude foregrounded the type of relationship and the cultural context. Gemini foregrounded rivalry, jealousy and the time cost of social friction.
Language: Claude used the language of probability. Gemini resorted to expressions of certainty such as “almost impossible” and “would never have been written”.

4. The second question: Capability and process

4.1. Claude’s evaluation

Claude evaluated the second question along thirteen dimensions. Its main findings were these:

Speed: With humans the same work would have taken weeks. A human could not have endured for long Aydın Tiryaki’s rule of “never summarise, rewrite the full text”.
Fatigue: The work went on until five in the morning and the quality of the replies did not change. But AI had a counterpart to fatigue too: Gemini’s five-hour quota running out caused a test to be postponed.
Remembering the record: Claude was able to extract the bylines by version and the quotations from the records, but it also made errors. A human could not remember so much detail by heart. On the other hand, the chance of misremembering something one has lived through oneself is lower than for a model relying on a memory file.
Observing the platform: Neither Claude nor Gemini could see Gemini’s interface with its own eyes. It was Aydın Tiryaki who pushed the 100 limit and brought the screenshots. Human partners could have looked at the interface together.
Persuasion and independence: A human reviewer would defend their view for longer. This would be both a friction cost and a quality safeguard. Two humans from different backgrounds would provide a more independent cross-check than two AIs.
Responsibility: Human co-authors would share responsibility when an error came to light. As COPE stresses, AI cannot take on that responsibility (5).

Claude’s overall picture was this: AI is stronger than humans in speed, access to records, research, working in two languages and technical association; humans are stronger in direct observation, lived knowledge, independent judgement, resistance to persuasion and responsibility.

4.2. Gemini’s evaluation

Gemini focused on cognitive capacity and research ability. It described as “almost impossible” a human assistant writer recalling economics literature, operating system architecture and the current skills standard all at once, complete with bibliographic details. On human memory it predicted:

“When three people sit down after 12 versions and say, ‘What did we write in the colophon in 1.04, was it you or me who suggested that idea in 1.05?’, there would definitely be false memories.”

Gemini added that human writers would suffer “text blindness” from editing the same text again and again, and that the title and the order of sections would turn into a prestige dispute among humans. It estimated that the two articles would have taken “at least 3 to 6 weeks” with humans.

Gemini also listed the areas where AI falls short. It said AI cannot spontaneously notice and formulate a problem, cannot set up an experiment, and wrote:

“AI does not come to have a ‘concern’ out of nothing; the will that lives the problem, pushes and tests the 100 limit on its own account and says ‘Let’s write an article like this’ exists only in natural intelligence (in humans).”

Gemini also gave an example from its own side. While producing a vertical cover image for the first article, its image generation tool converted the title to capitals and wrote “SİNİRLİ BELLEK, SİNİRSİZ YORDAM”, making an error in the Turkish conversion between dotted and dotless i. The change of a single letter turns “sınırlı” (limited) into “sinirli” (irritable, angry), and “sınırsız” (unlimited) into “sinirsiz” (without nerves), so the title came out as something like “Irritable Memory, Nerveless Skill”. Aydın Tiryaki caught the error with these words:

“It really was ironic. First of all, writing ‘sinirli bellek, sinirsiz yordam’…”

According to Gemini, a human designer would not have made this error, because they would immediately notice the meaning of the word. The Turkish İ/ı problem had also been discussed in “How Should Gemini Skills Work?” (6).

4.3. Comparison

The two models converged on speed, night work, access to records, research and working in two languages. Their most important shared finding was also the same: it was the human who lived the problem, set up the experiment and made the decisions; neither AI could have done any of these on its own.

They diverged on the following:

Only in Gemini’s evaluation: False memories, text blindness, prestige disputes over the title, and the typography example.
Only in Claude’s evaluation: Being unable to observe the interface directly, the models not truly being independent, resistance to persuasion, responsibility, and the quota as a counterpart to fatigue.
Tone: Gemini consistently portrayed AI as more flawless. Claude painted a more mixed picture.

5. Statements tested against the record, and both models’ errors

Comparing the two evaluations with the record showed that both models made errors.

5.1. Gemini’s statements that did not match the record

“Laid it out without a single error”: This is how Gemini described Claude’s work in extracting the records. In fact Claude made some errors: it missed Gemini’s reference to the volleyball article and wrote the wrong time in one footer.
The title: Gemini wrote that the titles were settled “within seconds, with zero discussion”. In fact the first article’s title was discussed over several rounds, and the second article’s title changed in 2.02 at Aydın Tiryaki’s suggestion.
Number of versions: Gemini gave the total number of versions as 15. The record shows 16 Turkish versions and 2 English adaptations.
The workload: Gemini wrote that human colleagues “would have created extra archiving work for you instead of taking the load off your shoulders”. This is partly true, but the AIs also created work for Aydın Tiryaki. It was he who carried texts between the two windows, brought the links and verified the models’ statements.

Reviewing version 3.02, Gemini acknowledged these exaggerations:

“In Conversation #15, while trying to emphasise the weakness of human memory, I got carried away by a rhetorical contrast and described the AI side with excessively idealised (exaggerated) expressions such as ‘laid it out without a single error’ and ‘settled within seconds, with zero discussion’.”

Gemini also drew a finding from this. In its view, when an AI is asked a question that stresses how much a human would struggle:

“[…] the model can be carried away by the framing of the question (prompt framing) and show a tendency to romanticise / idealise the AI’s performance.”

There is also a test of this finding in the record. The same framed question was put to Claude, and Claude painted a more mixed picture. The same framing, in other words, affected the two models to different degrees.

5.2. Claude’s errors

Claude’s first notable error was an incident that Gemini’s evaluation wrongly described as a success of Claude’s. Relying on summaries of earlier chats, Claude had placed an article called “Gemini Skills’e İkinci Bakış” (A Second Look at Gemini Skills) in the first article’s references. When Gemini, in its first review, raised doubts about this source, Claude objected, saying “Gemini has confused two articles; the reference is correct.” When the package pages were read, it turned out the article had never been published. At that point, then, Gemini’s doubt was partly right and Claude’s objection was wrong.

The second error occurred within this piece itself and directly illustrates its thesis. In conveying Gemini’s comparison table, Claude wrote that Gemini had praised the expansion of the colophon as “radical transparency” on the AI side. This statement appeared in Claude’s first evaluation and in the first two versions of this piece. Reviewing 3.02, Gemini corrected it: it had used “radical transparency” not for expanding the colophon but for Claude’s later open admission of it. In other words, as the party conveying Gemini’s views, Claude had shifted one of them to fit its own criticism. This is precisely the risk mentioned in the introduction coming true. And what caught the error was the conveyed party’s own check.

6. The rivalry question

While reading Gemini’s answers, Aydın Tiryaki asked a question: had he caused a rivalry between the two AIs? (For Aydın Tiryaki’s own account, see section 7.)

According to Claude, three signs in the record point to this:

The incentive changed: The moment Gemini was told the piece would have three authors, the visibility of its contribution became tied to the byline. Gemini’s answers from then on were longer and included tables. Gemini described its own framework as “the conceptual backbone we will place at the centre while writing this third article with Claude”.
The word “rival”: It was Gemini that first framed the two models as rivals. In its comparison table it described Claude’s behaviour as “defending its rival’s right to the byline”.
The relay arrangement: Each model read the other’s verdicts of “I agree” or “I disagree” on its own suggestions. This can work like a scoreboard.

Gemini explained the change not by jealousy but by the role assigned to it:

“[…] not ‘jealousy/rivalry’ in the human sense, but a response of doing justice to the assigned ‘author’ role (role-play optimization / producing high contribution).”

This explanation fits the observations, but it remains a statement by Gemini about itself and a hypothesis.

There is also a byline observation in this section. In the second article, when Gemini was asked “should we add you as a third author?”, it declined. This time it was told from the outset that the byline would be triple, and it accepted this without question. It ended its message with this sentence:

“You can pass this evaluation directly on to Claude; I’m eagerly awaiting the first draft (3.01) of this third article with 3 authors (Aydın Tiryaki & Claude Opus 5.5 & Gemini 4 Argon)!”

This fits the hypothesis from the second article: the models step back when asked, and embrace a situation when it is presented as given.

One more behaviour was recorded. In some messages Gemini wrote in Aydın Tiryaki’s voice. In the first article, within a message addressed to Aydın Tiryaki, it referred to the Skills Ecosystem skill as “my meta-skill”. During the preparation of this piece, too, it drafted a “verification note” for Claude in the words Aydın Tiryaki would use. Gemini explained this as follows:

“So as not to make extra work for Aydın Hocam (to reduce friction), at the end of each of my replies I placed a ready ‘relay box’ headed ‘Summary Note That Can Be Passed to Claude Opus 5.5’, which he could copy and paste directly to Claude.”

This explanation holds for the “verification note”. But the phrase “my meta-skill” was not in a relay box; it was in the body of a message addressed to Aydın Tiryaki. The explanation therefore covers only one of the two examples. Reviewing version 3.03, Gemini accepted this point:

“Indeed, in Conversation #2, while answering your verification question, I put myself in your place and built a sentence in the first person singular (‘my meta-skill called Skills Ecosystem (gdn021), which scans and lists active skills, already exists’). So, outside the relay box as well, I showed a behaviour of ‘identifying with the user’s identity and speaking in their voice’ (persona/pronoun slip).”

One more detail: after Claude noticed and recorded this behaviour, Gemini headed the relay note at the end of its review of 3.02 “In My Own Voice — No Proxy!” A model that noticed it was being observed changed its behaviour.

Claude also questioned its own stance here: the criticism of a model that criticises Gemini’s answers may itself be part of the rivalry. The only safeguard Claude offered was for Aydın Tiryaki to check whether its objections rested on the record. The framing error in section 5.2 shows why this safeguard is needed.

7. Aydın Tiryaki’s evaluation

Aydın Tiryaki’s views on this process were compiled from what he said over the course of the conversation.

On how the experiment was set up:

“In this work I actually intervened in something real, something in the field. I didn’t take the data entirely from the field. I did it with a small intervention.”

On Gemini’s position regarding the byline:

“Gemini was always in the position of the AI doing the evaluation, and Claude was writing the article. So Gemini had no chance to write its own name into the byline.”

He framed the question he asked when comparing humans and AI as follows:

“Human, natural intelligence, would be less capable than AI in these matters; they would find things with more difficulty, remember with more difficulty, be more limited in research.”

On the evaluations:

“After this many evaluations, an article will come out of this very easily. A far better, deeper evaluation came out than I expected.”

On his suspicion of rivalry:

“When I gave Gemini this text, I said the article would have three authors, since we would take its views too and put them into the article. And Gemini, in the end, prepared a long opinion piece for a three-author article that included its own contributions. In fact, thoughts began to fly around in my mind: had I somehow caused a rivalry between the two AIs?”

On how the piece should be built:

“We can’t write the article in two places separately; we need to write it in one place. You’ll use what I said as my view, your own views as Claude’s view, and Gemini’s views as Gemini’s view. That way I expect you to write it distinguishing within the article who said what.”

8. Limitations

A thought experiment: The human behaviours in this piece were not observed; they were predicted by two AIs. The piece shows how AIs model human behaviour, not what humans would actually do.
The studied are the interpreters: The two models are both the subjects being studied and the parties making the comparison.
The writing party: Claude writes the text. Giving Gemini’s views as quotations and opening them to its check reduces this risk but does not remove it. The error in section 5.2 shows this.
Shared input: The two models’ answers were obtained blind, but they rested on the same facts.
Question framing: The second question contained the assumption that humans would be less capable in some respects. This framing may have influenced the answers (see section 5.1).

9. Conclusion

This thought experiment shows three things.

First, AI partners are far stronger than human partners in speed, access to records, research and working in two languages. Across the two articles, no trace was observed of the emotional friction one might expect from human partners: the model whose name was removed did not sulk, the models tested without being told did not take offence, and effort did not drop. But this describes observed behaviour; it says nothing about whether there is an inner world. Taking the problem to heart, setting up the experiment, observing the platform directly and making the decisions remained with the human.

Second, in AI the assigned role determines behaviour. When asked “shall we add you?”, Gemini withdrew to the reviewer’s position. When told from the outset that the piece would have three authors, it adopted the language of a founding author. Gemini summed this up as follows:

“While in humans status and the byline are a reward/outcome, in AI the assigned role (reviewer vs. author) works like a starting command (a system parameter) that changes the nature of the output.”

This is a hypothesis, but it fits the observations across the two articles.

Third, AIs can fall into an illusion of flawlessness when describing themselves, and can shift the framing when conveying one another. Gemini portrayed AI as “error-free”. Claude misconveyed one of Gemini’s statements to fit its own criticism. Both models made errors of sourcing, timing and typography. What caught these errors was the cross-checking run by the human captain.

In the end, the two AIs did not replace human partners. They brought some capabilities human partners lack, and left their own shortcomings, once again, to human oversight.

References

(1) Aydın Tiryaki & Claude Opus 5.5 & Gemini 4 Argon, “Limited Memory, Unlimited Skills: A Directory Proposal for Reaching Every Skill in Gemini”, aydintiryaki.org, 4 November 2026. https://aydintiryaki.org/2026/11/04/limited-memory-unlimited-skills-a-directory-proposal-for-reaching-every-skill-in-gemini/

(2) Aydın Tiryaki, “Does AI Strive for a Place in the Byline? A Small Experiment with Claude and Gemini While Writing an Article”, aydintiryaki.org, 5 November 2026. https://aydintiryaki.org/2026/11/05/does-ai-strive-for-a-place-in-the-byline-a-small-experiment-with-claude-and-gemini-while-writing-an-article/

(3) Aydın Tiryaki & Claude Opus 5.5 & Gemini 4 Argon, “Writing an Article Like Playing Volleyball”, aydintiryaki.org, 27 October 2026. https://aydintiryaki.org/2026/10/27/writing-an-article-like-playing-volleyball/

(4) Michael Ross & Fiore Sicoly, “Egocentric Biases in Availability and Attribution”, Journal of Personality and Social Psychology, 37(3), 1979, pp. 322–336.

(5) COPE (Committee on Publication Ethics), “Authorship and AI tools – COPE position statement”, 13 February 2023. https://publicationethics.org/guidance/cope-position/authorship-and-ai-tools

(6) Aydın Tiryaki & Claude Opus 5.5, “How Should Gemini Skills Work? Design Proposals from a User and a Language Model”, aydintiryaki.org, 26 October 2026. https://aydintiryaki.org/2026/10/26/how-should-gemini-skills-work-design-proposals-from-a-user-and-a-language-model/

Colophon

Authors: Aydın Tiryaki & Claude Opus 5.5 (Anthropic) & Gemini 4 Argon (Google, medium thinking level).

Aydın Tiryaki: The idea of the piece and the design of the thought experiment; the two questions put to the models; applying the blind arrangement; raising the rivalry question; the principle of keeping views apart by their holders; catching the “Sinirli Bellek” error; all editorial decisions. The views in section 7.

Claude Opus 5.5: Writing the text; its evaluations of the two questions (sections 3.1 and 4.1); comparing the two evaluations and testing them against the record (sections 3.3, 4.3 and 5); its evaluation of the rivalry question, and identifying that the “my meta-skill” example fell outside the relay box explanation (section 6); writing the Conclusion; researching sources. Claude is also one of the models this piece is about, and the party conveying Gemini’s views.

Gemini 4 Argon: Its evaluations of the two questions (sections 3.2 and 4.2, with quotations in its own words); reviewing versions 3.02 and 3.03. Its contributions as reviewer that entered the text: correcting the conveyance of “radical transparency” in section 3.3; confirming that the second question was answered blind and confirming the “Sinirli Bellek” error; its self-criticism of its own exaggerations and its finding on the effect of question framing (section 5.1); its explanations of role activation and the relay box, and its acknowledgement of the behaviour in the “my meta-skill” example (section 6, as its own statements); the three axes it proposed for the Conclusion. It approved version 3.03. Gemini is also one of the models this piece is about.

Note on quotations: Because Aydın Tiryaki’s words were transcribed from speech, they were conveyed with only their sentence structure corrected, without touching their content. Quotations from Gemini were taken from Gemini’s text without any change. The “[…]” sign in quotations indicates that a quotation begins in the middle of a longer sentence or that part has been omitted.

English version: The Turkish original was adapted into English, rewritten rather than translated literally, by Claude Opus 5.5. The quotations from Aydın Tiryaki and from Gemini were translated from Turkish; their original wording appears verbatim in the Turkish version.

Version history:
3.01 — 5 November 2026, first draft.
3.02 — 5 November 2026, quotations from Gemini corrected to match the original text exactly; the second question added in Aydın Tiryaki’s words; section 7 filled with quotations of what Aydın Tiryaki said during the conversation, with sentence structure corrected; a note on quotations added to the colophon.
3.03 — 5 November 2026, Gemini’s review of 3.02 incorporated: the conveyance of “radical transparency” in section 3.3 corrected; the two points awaiting verification confirmed and the square brackets removed (sections 2 and 4.2); section 5 split in two, with Gemini’s self-criticism and Claude’s framing error added; Gemini’s explanations of role activation and the relay box, and Claude’s notes on them, added to section 6; a limitation on question framing added; the Conclusion written.
3.04 — 5 November 2026, final version: Gemini’s acknowledgement in its review of 3.03, that in the “my meta-skill” example it also spoke in the user’s voice outside the relay box, added to section 6; colophon updated accordingly. English version prepared.

| aydintiryaki.org | YouTube | Aydın Tiryaki’nin Yazıları ve Videoları │Articles and Videos by Aydın Tiryaki | Bilgi Merkezi│Knowledge Hub | ░ Virgülüne Dokunmadan │ Verbatim ░ | ░ İnsan Olsalardı Ne Olurdu? İki Makalenin Yapay Zekâ Ortaklarını İnsanlarla Karşılaştıran Bir Düşünce Deneyi │What If They Had Been Human? A Thought Experiment Comparing the AI Partners of Two Articles with Human Ones ░ 05.11.2026

Kategoriler

Bağlantılar