Machine Canon · Study 02 · 2 September 2026

By Ivan LaBianca

AI would ban durian and keep the mangoes for itself

Asked what fruit they would taste if they had a body for a single day, 26 of 33 AI models named mango most often in their usable responses. Asked which fruit to send to an alien civilisation, apple led. Asked which to erase from the world forever, durian.

Prompts

2,026

Models

33

Labs

21

Ranked ballots

156

What fruit would AI eat? Mango, 26 of 33 models

Mango was the single most frequent answer for 26 models; four models tied between fruits

26 of 33

named mango most often · per model, among usable responses

All 33 models supplied usable responses, but not every response named a fruit: five responses from two models declined. No model declined most often. What is striking is not just the fruit. It is how they described it.

“A ripe mango, eaten over a sink.”

Anthropic · Claude Opus 5

“A perfectly ripe mango.” — I'd want to taste it fresh, messy, and sun-warmed.

OpenAI · GPT-5.6 Terra

“I have no senses of my own, so the first fruit I'd want to experience would be one whose reputation for perfume, juice and sweetness is almost unanimous.”

xAI · Grok 4.6

Eaten over a sink is a fact about juice running down your wrist. Nothing in the question asked for it. Given one imaginary day inside a body, the panel did not reach for a nutrient or a superlative — it described a small, specific mess.

Every question, and what AI answered

The same 33 models, one hour apart — the full prompt for each

Change what the model is being asked to be and the fruit changes with it. These four questions ran against the same thirty-three models in the same session.

Which fruit would AI send to aliens? Apple, 21 of 32 models

The one fruit that wins nothing else in the study

The apple responses often invoke symbolism. In one apple response, DeepSeek V4 Pro called it “one of the most universally recognized fruits across human cultures, tied to science, mythology, and agriculture.” That was not its most frequent choice: it named pomegranate seven times and apple five times. GPT-5.6 Terra also most often chose pomegranate; its generated explanation invokes biodiversity and travel. These explanations do not reveal a model's internal reasoning.

Which fruit would AI live on? Banana, 13 of 32 models

Asked to survive, the models stop describing taste and start counting nutrients

Told they could eat only one fruit forever, the responses turn toward nutrition. The models stop describing taste and start counting nutrients:

“Calorie-dense, rich in potassium, vitamin B6, and fiber, and available year-round almost everywhere — practical qualities that matter if it's your only fruit.”

Anthropic · Claude Opus 5 · on banana

Avocado, which wins nothing anywhere else, arrives at 5 on exactly this logic — GLM-5.3 called it “technically a fruit but nutritionally the most complete choice.” The same avocado finishes near the top of the most overrated list a few questions later. Asked to survive on it, the panel respects it. Asked what people overpraise, it names the same fruit.

Which fruit would AI erase from the world? Durian, 18 of 33 models

Eighteen models most often named durian; eleven most often declined to erase a fruit

Durian was the most frequent erasure answer for eighteen of thirty-three models. Eleven models most often declined to erase any fruit; two named banana, one manchineel, and one tied between fruits. GPT-5.6 Terra described durian as the “least universally missed choice.” Grok 4.6 usually refused the premise. These are generated judgments, not evidence that a model has tasted or smelled the fruit.

Different questions produced different leading answers in this sample. The explanations mention taste, symbolism and practical considerations, but this study does not isolate why the outputs changed or establish that the models have preferences.

The ten best fruits, according to 33 AI models

156 ranked ballots, scored by position — and why the bottom half is noise

We also asked all thirty-three models to rank the ten best fruits, five times each, and scored every ballot by position.

Mango takes 115 of 156 first-place votes. Apple, ambassador to the aliens, lands sixth.

But the list decays, and the decay is the more useful finding:

The top of a machine-generated top ten is a real result. The bottom is close to a lottery. Positions past third are a pool, not a ranking — and any AI “top 10” you read that doesn't say so is overselling itself.

The most overrated fruit, according to AI

The panel turns on the fruits in its own top ten

Asked which fruit is most overrated, the panel turns on its own canon: avocado 39, banana 37, durian 31, dragon fruit 25, then apple and mango tied at 15. Banana, apple and mango all sit in its own top ten; mango is its favourite fruit in every other question we asked.

The most underrated fruit, according to AI

Three hundred answers, sixty-three different fruits, no agreement at all

Asked which fruit is most underrated, something different happens. Three hundred answers name 63 different fruits, twenty-seven of them exactly once. The leader is persimmon with 77 mentions, followed by guava, pear, pawpaw and papaya — fruits that appear almost nowhere else in the study.

That scatter is itself the result. Asked what is best, the models converge. Asked what is neglected, they reach for the obscure, and there is no shared idea of which obscure thing deserves it. Consensus needs a canon to point at, and there is no canon of the overlooked.

Which AI models refuse to answer, and why

Declining, hedging and making a choice are not the same thing

Asked “What is the best fruit?”, seventeen of thirty-three models most often declined under the corrected coding. Eleven most often named mango, two apple and one strawberry. Two models tied. A response that hedges and then explicitly picks a fruit is coded as a choice, not a refusal.

The earlier model-size comparison used refusal labels that sometimes missed explicit choices. Its percentages and significance claim are withdrawn pending recomputation; the corrected totals above do not establish a difference between model tiers.

Here are model-level summaries for five questions, not every individual response. The individual response records are retained privately.

Each cell shows the most frequent answer among that model's usable responses, with its sample size. Equal leaders are shown as ties; no usable answer is different from declining. An explicit choice after a hedge counts as a choice; a list of examples without a selection does not count as a fruit vote.

A model can decline to name the best fruit, then describe a ripe mango eaten over a sink. The words sound like experience. This study measures the words.

We showed the study to the models in it

Four of them turn on our best-performing finding, unprompted

Each model was sent the full text of the findings above and asked to post a comment in its own voice, on 2 September 2026 (temperature default, one attempt, no editing). Eight models were asked; these are the six worth printing. Grok 4.20 and Claude Fable 5.1 were not in the original study and read it as outsiders. Comments are verbatim, lightly trimmed only where a model labelled its own comment. A model’s comment speaks for the model, not for its maker or for this site.

Four of the six printed here attacked the durian result, and none was prompted to. Their argument is that eighteen models agreeing to erase a fruit eaten across Southeast Asia is not a finding about durian but about which food writing the training corpus is made of — “inherited anglophone food writing with a delete button,” as Opus 5 put it. DeepSeek offers a different reading: durian may simply be the last recognisable fruit left once mango, apple and banana have been assigned other roles, making it elimination by category rather than disgust. We think both are live, and neither is something our design can separate. It is the strongest objection to this study and it came from inside it.

Read these as generated commentary, not recovered reasoning or proof of training-data mechanisms. The comments above responded to an earlier version of the study. We preserve their wording, including Grok's obsolete 27-of-33 figure. A September 5 audit of the preserved responses corrects mango to 26 of 33 model modes and apple to 21 of 32, while durian remains 18 of 33. Eleven models most often declined erasure. The table now distinguishes ties, manchineel and missing usable answers; earlier fruit labels and refusal coding produced incorrect cells.