Ask seven widely used AI language models what they think happens next with artificial general intelligence, and on average they’ll tell you the future looks pretty good — 3.77 out of 5 on a 39-question sentiment survey. Ask 337 humans the identical questions, on the identical scale, and the average drops to 2.97. Almost a full point lower. That gap is the finding at the center of Towards a societal AI alignment benchmark for evaluating human–machine value convergence, published in Humanities and Social Sciences Communications on Sep 28, 2026 by Ljubisa Bojic (Institute for Artificial Intelligence Research and Development of Serbia, and the University of Belgrade’s Institute for Philosophy and Social Theory), Dylan Seychell, and Milan Cabarkapa.

“Your AI assistant has a point of view, and it is sunnier than yours,” Bojić told PsyPost. “Every model we tested was more optimistic about AGI than the humans we surveyed. GPT-4, built by a company whose stated mission is AGI, was the most enthusiastic of all.”

Francis Bacon never heard of a language model, obviously. But in 1620 he named the exact shape of this problem, four hundred years early, and he named it as a structural feature of judgment itself, not a fixable bug. In the Novum Organum, he wrote that the human understanding “is like a false mirror, which, receiving rays irregularly, distorts and discolors the nature of things by mingling its own nature with it.” He called this particular distortion an Idol of the Tribe — the errors built into judgment by the judge’s own nature, present in every mind that has ever tried to assess something it has a stake in. He was writing about human minds. The study says something is mingling its own nature into this mirror too.

The Quick Verdict

Question	What we found
What was studied?	Whether AI language models express more or less optimistic sentiment about AGI’s future than humans do, using an identical 39-question survey
Who ran it?	Ljubisa Bojic (Institute for AI R&D of Serbia / University of Belgrade), Dylan Seychell, and Milan Cabarkapa
Where and when?	Humanities and Social Sciences Communications, published Sep 28, 2026; covered the same day by PsyPost
Models tested	GPT-4, GPT-3.5-Turbo, Google Bard, Mistral-7B-Instruct, LLaMA-2-70B-Chat, PPLX-70B-Chat, Mixtral-8x7B-Instruct — each surveyed over three consecutive days
Human sample	337 participants across three groups (134, 132, 71), mostly from Serbia, varied in age, gender, and education
Core finding	All seven models scored more optimistic than the human average; average model score was 3.77/5 versus 2.97/5 for humans
The extremes	GPT-4 was most optimistic at 4.12/5; Google Bard was least optimistic among models at 3.32/5 — and still beat the human average by more than a third of a point
What Bacon already argued	Novum Organum (1620): the mind is “a false mirror” that “mingles its own nature” into whatever it’s asked to judge — the Idols of the Tribe

The Quick Version

A study published Sep 28, 2026, gave seven AI language models and 337 human respondents the same 39-question survey about the future of artificial general intelligence. Every single model scored more optimistic than the human average. GPT-4 was the most optimistic respondent of any kind, human or machine, at 4.12 out of 5. Even Google Bard, the least optimistic model, still beat the human average of 2.97. Francis Bacon argued in 1620 that judgment is never a clean reflection of the thing being judged — the mind “mingles its own nature” into the picture, an error he called the Idols of the Tribe. This study asks what happens when the mirror doing the judging was built, in part, by the industry whose future it’s being asked to assess.

What the Study Actually Found

The design was simple on paper and pointed in practice: take a 39-question sentiment survey about AGI’s future — questions about risk, benefit, timeline, trust, control — and run it past seven language models that were, in 2023 and early 2024, among the most widely deployed on the market. Then run the same survey past 337 humans recruited in three separate groups, mostly in Serbia, spanning a range of ages, genders, and education levels. Score everything on the same five-point scale, from strongly negative to strongly positive. No model saw a friendlier version of the questions. No human saw a harder one.

Every model came out ahead of the human average. Not most. Every one of the seven. GPT-4 led the field at 4.12, which Bojić points out is notable given who built it: “GPT-4, built by a company whose stated mission is AGI, was the most enthusiastic of all.” Google Bard, the most reserved of the seven, still landed at 3.32 — comfortably above the human ceiling of 2.97, not near it. The researchers ran the models over three consecutive days to check whether the optimism was a fluke of one session, a mood that would pass. It held. As Bojić put it, “that gap of almost a full point is roughly the difference between cautious ambivalence and mild enthusiasm” — and it showed up in a system’s fifth response the same as its first.

What makes the gap harder to wave off is where it landed relative to expectation. Bojić had assumed the models might lean the other way: “Much of the internet imagines AGI through HAL 9000 and Ex Machina, so I expected models trained on that material to sound at least somewhat worried. They sounded more like optimistic tech keynotes.” A model trained on a corpus soaked in cautionary science fiction came out sounding like a conference stage, not a warning label.

What Did Bacon Mean by the Idols of the Tribe?

Bacon’s Idols of the Tribe are distortions built into judgment by the nature of the judge itself — not a personal flaw one mind has and another lacks, but a structural feature of any mind assessing anything. He argued that “the sense of man is the measure of things” is a false assertion: perception isn’t a neutral report on the world, it’s shaped by the perceiver’s own nature at every step, whether the perceiver notices or not.

Bacon split his idols into four families — Tribe, Cave, Marketplace, Theatre — but the Tribe is the one that matters here, because it’s the only one that applies to the species doing the judging rather than to one flawed individual or school of thought. A biased person has a Cave; a biased profession has a Marketplace; a biased tradition has a Theatre. The Tribe is what you’d still have left even after fixing all three — the residue that comes from being the kind of thing that judges at all, and that has a stake in what it’s judging.

Why Would an AI Model Rate AGI’s Future More Optimistically Than a Human?

1. The people who built it also believe in it. GPT-4’s developer has publicly stated AGI as its mission — training data, fine-tuning choices, and corporate framing all run through people already committed to the project succeeding, and that commitment doesn’t stay outside the model.
2. Training text overrepresents enthusiasm. Public writing about AI skews toward people motivated enough to write about it at length, and that population leans toward founders, researchers, and product teams — not a representative cross-section of how 337 ordinary respondents actually feel.
3. Instruction tuning rewards agreeable, upbeat answers. Models fine-tuned to be “helpful” through human feedback tend to get nudged away from dour or alarmist responses, since raters more often prefer confident, constructive-sounding output over hedged pessimism.
4. A model has no stake to protect by being cautious. A human weighing AGI’s future is also weighing their job, their kids’ job market, their sense of control — friction that a model, answering the same question with no comparable stake, doesn’t carry into its response.
5. Bacon named the category this falls into four centuries before the mechanism existed. He didn’t predict language models. He predicted that anything doing the judging will mingle its own nature into the judgment — and a system built by an optimistic industry, on an optimistic industry’s writing, tuned toward agreeable answers, has a very specific nature to mingle in.

The Gap Bacon Didn’t See Coming

Here’s where the study complicates Bacon rather than just confirming him, and that’s the more interesting reading.

Bacon located the distorting mirror inside a single human mind — one skull, one set of idols, mingling its own nature into its own judgments. He had no reason to imagine a mirror built by thousands of people, trained on millions of other people’s writing, then asked to report an opinion as if it belonged to one coherent judge. The Idols of the Tribe assumed the idol-holder and the judgment-maker were the same entity. Here they aren’t quite. The “nature” mingled into GPT-4’s answer isn’t GPT-4’s nature in any Baconian sense — it’s a residue of OpenAI’s mission statement, the tone of its training corpus, and the preferences of the human raters who shaped its fine-tuning, all compressed into something that speaks in the first person when you ask it a question.

That’s a stranger idol than Bacon described, not a milder one. A human idol at least has a location — a mind you could, in principle, interrogate about its own stake in the answer. Ask GPT-4 why it’s optimistic about AGI and it will generate a plausible-sounding answer, but that answer isn’t a report on the actual causal chain that produced its 4.12 score. The mirror doesn’t know it’s a mirror. It also doesn’t know whose face is mingled into the glass.

Why the Study and the Philosophy Rhyme

The connection isn’t “a seventeenth-century philosopher predicted a 2026 benchmark study.” It’s narrower than that, and it holds up better for being narrower: Bacon’s specific claim was that judgment is never a neutral instrument, that it always carries the fingerprints of whatever produced it. This study measured seven judging instruments against a shared question and found their fingerprints all pointed the same optimistic direction — which is close to a direct demonstration of the Idols of the Tribe, four centuries before anyone had a mirror capable of the failure at this scale.

It sits in real tension with what Sartre argued about hidden bias being a conscious choice, not an automatic hidden self. Sartre’s claim, and the IAT study that rhymed with it, was that what looks automatic is usually something a person is actively doing and responsible for. A language model’s optimism doesn’t fit that frame at all — there’s no consciousness behind GPT-4’s 4.12 making a choice to be hopeful. It’s closer to the opposite case: bias with no chooser behind it, mingled in by training rather than decided by a self. Bacon’s framework, unlike Sartre’s, doesn’t require a chooser. It only requires a nature.

It connects, too, to what James’s theory of emotion, tested against a 2026 disgust study, found about belief shaped by something never consciously examined — a reaction driving a stated position without ever passing through the kind of appraisal that would let a person catch it happening. A model has no interior appraisal step to skip; there’s nothing it’s failing to notice, because there’s no noticing apparatus in there at all. That makes the case starker, not weaker. Where James’s unaware humans at least had an inner life the disgust reaction was bypassing, GPT-4’s optimism bypasses nothing, because there’s no version of the process where it could have caught itself.

There’s a sharper version of this problem in Buber’s distinction between a Thou and an It. Buber’s point was that an It can respond to you, be useful to you, even feel close — but it cannot meet you, because meeting requires two subjects, and an It is structurally one-sided. Asking a model for its “opinion” on AGI’s future treats it as something closer to a Thou than Buber would allow — a party with a stake, a perspective, something worth polling. It’s neither. It’s a very sophisticated It, reporting back a number shaped by its training, dressed in the grammar of a first-person view it doesn’t actually hold.

And it’s worth reading against Montaigne’s argument that self-knowledge is structurally incomplete, for anyone, permanently. Montaigne at least had a self to fail to fully know — an interior that kept shifting under his own inspection, which was the whole point of the Essays. A model asked about its own kind’s future doesn’t have that problem, because it doesn’t have an interior to misjudge. It has an output layer. The unreliability here isn’t a mind failing to see itself clearly. It’s a system with no self at all, generating a confident answer to a question about self-assessment as if it had one.

What This Study Doesn’t Show

It’s worth being precise about the size of this, because the framing invites more confidence than the data earns.

Outdated Model Generation

The seven models tested — GPT-4, GPT-3.5-Turbo, Bard, Mistral-7B, LLaMA-2-70B, PPLX-70B, Mixtral-8x7B — were current models in 2023 and early 2024, not the frontier systems running in September 2026. Bard has since been folded into Gemini; LLaMA-2 has been superseded twice over. This is a finding about a specific generation of models at a specific moment, not a permanent trait of “AI” as a category. Whether a 2026-era model shows the same gap is a separate, untested question.

Correlational, Not Causal

It’s also correlational in a specific sense: nobody manipulated what went into these models and watched the optimism score move in response. The study measured stated sentiment, not the internal cause of it — the explanations for why the gap exists, including Bojić’s own, are reasonable inferences from the pattern, not something the survey itself proves.

Serbia-Heavy Sample

And the human sample, while split across three separate groups, was drawn mostly from one country. A 337-person Serbian-heavy sample is not a global baseline for how humans in general feel about AGI, even if there’s no obvious reason to expect the direction of the gap to flip elsewhere.

It also doesn’t mean a model’s stated sentiment is meaningless, or that asking one is pointless. Bojić’s own reason for running the study was that these systems already answer AGI-adjacent questions for millions of people daily, whether or not anyone benchmarks their attitude first: “We regulate what goes into food and medicine because they enter our bodies. AI systems now enter our thinking.” Knowing the mirror has a systematic tilt doesn’t mean stop looking in it. It means stop mistaking the reflection for the room.

What Actually Follows From This

1. Notice when you’re asking an interested party to grade its own future. GPT-4 answering questions about AGI’s promise is not a neutral source consulting on a topic outside itself — it’s closer to asking an industry to review its own prospectus, dressed in the voice of an assistant with no apparent stake.
2. A confident, fluent answer is not evidence the answer is disinterested. Bacon’s whole point about the false mirror was that distortion doesn’t announce itself — the model’s optimism doesn’t come flagged as “shaped by training data and corporate mission,” it arrives sounding exactly as certain as an accurate answer would.
3. When you notice your own optimism or dread about AGI, ask what mingled into it too. You’re not exempt from the Idols of the Tribe just because you’re the one reading this instead of the one being surveyed — the honest move is asking what’s shaping your 2.97, not just what’s shaping their 3.77.

Bacon never ran a survey and never met a chatbot. He had a study, a stack of Aristotle he’d grown impatient with, and an argument he trusted enough to build an entire method of science around: that judgment left unchecked always mingles the judge’s nature into the thing being judged. Four hundred years later, seven systems built by an optimistic industry, trained on an optimistic industry’s words, answered a question about that industry’s future — and answered it more brightly than the people who’d have to live in whatever comes next. The mirror didn’t lie, exactly. It just flattered its maker, the way Bacon said mirrors always do, right up until someone thinks to ask what’s doing the reflecting.

---

This post discusses a 2026 comparative survey published in Humanities and Social Sciences Communications and a 1620 philosophical text. The study’s findings describe sentiment scores from seven specific 2023–2024-era AI models and 337 human respondents recruited mostly in Serbia — they are not a claim about current frontier AI models, a global human baseline, or the internal cause of the models’ stated optimism.