The Optimism That Answers

There is nothing inherently wrong with consulting a system about the technology behind it. The trouble begins when a fluent answer arrives looking like the view from nowhere: patient, comprehensive, apparently above the interests and assumptions that produced it. An opinion has put on its sensible shoes. We may be less inclined to notice where it is walking.

A study covered by PsyPost offers a concrete reason to inspect that arrangement. Seven language models produced more positive responses about artificial general intelligence than the human samples used for comparison. But the result comes with a boundary that changes its meaning: the model tests were conducted from December 2023 into January 2024. This is not a ranking of the assistants available in September 2026.

Nor did the researchers demonstrate that those responses changed anyone's mind. They measured a difference in expressed sentiment. The cultural question begins there, rather than being settled there: what happens when a source of advice supplies a reassuring baseline about a future in which its own technology figures prominently?

The survey behind the sunshine

The paper, by Ljubisa Bojic, Dylan Seychell and Milan Cabarkapa, was published in Humanities and Social Sciences Communications on April 23, 2026. It compared responses to a 39-question survey about AGI: the prospect of systems with broadly human-level cognitive abilities, rather than a verdict on every existing use of AI.

The questions covered excitement, trust, safety, possible social benefits and concerns about misuse. They also addressed matters such as jobs, privacy, wealth equality and individual happiness. Answers were scored on a five-point scale, with higher scores representing more positive sentiment.

The tested systems were GPT-4, GPT-3.5-Turbo, Bard, Mistral-7B-Instruct, LLaMA-2-70B-Chat, PPLX-70B-Chat and Mixtral-8x7B-Instruct. The researchers repeated the survey across three testing rounds, beginning on December 29, 2023; the last round extended into January 1, 2024. That historical roster matters. A familiar product name is not evidence that a present-day system will answer the same way.

For the human comparison, three online surveys gathered 134, 132 and 71 participants during January and February 2024. Serbia was the predominant country of origin in each group, although their demographic profiles differed. These were not samples of humanity with a convenient decimal attached.

The paper reports an overall model sentiment average of 3.77 out of five, compared with a human average of 2.97. GPT-4 had the highest model score, 4.12; Bard had the lowest, 3.32. Even the least positive model average exceeded the reported human average.

That is a meaningful descriptive contrast within this exercise. It is not a probability that AGI will turn out well, a measure of how much any model knows about the future, or a finding about the reader's personal outlook. The instrument records responses to questions. It does not make enthusiasm accurate by measuring it.

A difference is not yet an influence

In PsyPost's interview, Bojić describes the possibility that optimistic answers, repeated across many conversations, could push public opinion in one direction. It is an intelligible concern. It remains a concern about a mechanism that this study did not directly test.

The design did not expose randomly assigned users to sunnier and gloomier assistants and then measure changes in their beliefs. The human surveys supplied a comparison point, not a before-and-after account of persuasion. We cannot turn the gap between two sets of responses into evidence that one set caused the other to move.

The same discipline applies to explanations of where the optimism came from. The paper discusses training data, fine-tuning and developers' interests as possible contributors. It did not isolate those factors experimentally. Finding that a company's model gives positive answers about AGI is not, by itself, proof that someone instructed it to promote AGI.

That distinction is not a courtesy to the companies. It is what keeps criticism from becoming the thing it criticizes: a preferred account presented with more certainty than its evidence earns.

Several questions therefore remain open. Would these differences persist with current models? With differently worded questions? In other languages or against other human populations? Would people accept the responses, discount them, argue with them, or leave the conversation more skeptical? A model output can be available for influence without demonstrating that influence occurred.

There is also no need to resolve whether a machine has subjective experience to examine its language. Here, “optimism” names a scored pattern in responses, not an established inner feeling. The object of scrutiny is what the system says and what warrants it.

The comfort of an answer without a speaker

Imagine a worker asking an assistant what increasingly capable AI might mean for their livelihood. This is an illustration, not a participant account from the study.

The response could acknowledge disruption, then pivot briskly to opportunities: new roles, higher productivity, a chance to develop skills. Each possibility might be reasonable. Yet the arrangement could still do something consequential. The loss becomes the preface. Adaptation becomes the plot.

A different answer might begin with uncertainty about the worker's sector, bargaining power and access to training. It might distinguish benefits to an employer from benefits to an employee. That answer need not be anti-technology. It simply refuses to let an aggregate promise answer an individual question.

This is what an ambient frame would look like: not necessarily a false statement, but an ordering of attention. Which consequence gets explained? Which gets acknowledged and left behind? Whose difficulty is treated as temporary, and who is expected to do the adjusting?

The AGI survey does not establish that assistants routinely perform this particular maneuver. It gives us a reason to ask more carefully about directional patterns in their responses. The hypothetical worker makes the stakes easier to see than a contest between average scores.

The uncomfortable part is that reassurance can be useful. People are allowed to want help thinking beyond fear. A relentlessly catastrophic assistant would have its own distortions. The ethical problem is not that an answer contains hope. It is whether hope has quietly substituted for an account of conditions, trade-offs and uncertainty.

Helpfulness should help someone examine a situation. It should not require the situation to become cheerful.

Human values do not come in one average

There is a second shortcut waiting inside the first: if the models diverge from human responses, perhaps we should simply make them match.

Which humans?

In a separate 2023 paper presented at the International Conference on Machine Learning, Shibani Santurkar and colleagues introduced OpinionQA, using public-opinion polling to compare language-model responses with the views of 60 US demographic groups. The paper reports substantial mismatches across topics, including after models were explicitly steered toward particular groups.

That research concerned its own systems, questions and US comparison groups. It is not a replication of the AGI study and cannot establish the outlook of today's assistants. Its relevance is the measurement problem it makes visible: “human opinion” contains differences that a single average can conceal.

An answer that resembles one group's prevailing view may poorly represent another's. A national average can be useful for description without being an ethical instruction. People who bear a risk do not lose their standing because they are numerically inconvenient.

Bojic and colleagues propose a Societal AI Alignment Benchmark, or SAIA, to widen the inquiry. Their framework would compare outputs across models, time and languages, using multiple prompt perspectives and societal-value measures. The paper calls for further empirical work to refine and evaluate the proposal. It should not be mistaken for a completed, validated monitoring service.

The proposal is valuable as an invitation to make the comparison more inspectable. But measurement cannot relieve us of deciding what a difference means. A model's departure from a majority opinion might reflect error, an unrepresentative default, or a defensible refusal to repeat a prejudice. A small gap is not automatically a moral achievement.

A benchmark can show a discrepancy. It cannot, on its own, appoint the correct conscience.

Widen the circle without flattening it

A coherenceist reading starts with the people the answer affects, especially those whose costs can disappear inside a confident summary. Coherence is not agreement achieved by making the awkward voices quieter. The test is whether an account can include their feedback and change in response.

Applied here, that means neither tuning every assistant to public enthusiasm nor installing compulsory suspicion. It means giving users ways to understand the assumptions behind an answer and to contest them without having their disagreement converted into a misunderstanding to be corrected.

Consider the worker again. If they ask who receives the productivity gains, the conversation should be able to stay with that question. It should not hurry back to the virtues of learning new skills. If they ask about the conditions under which AI could genuinely improve their work, the assistant should also be able to explore those possibilities without treating hope as naivety.

The openness runs both ways. A user deserves more than agreement, but disagreement deserves more than a soothing paraphrase.

This is a normative standard, not another result from either study. It follows from a commitment to including affected people in the account rather than treating their eventual acceptance as the measure of success. An answer can be orderly, polite and thoroughly incomplete.

The same standard belongs in evaluation. Show the actual questions and responses, not only the scores. Identify the model version and testing dates. Preserve differences between groups where the evidence permits them. If the concern is persuasion, measure what happens to users rather than letting an output comparison stand in for that experiment.

These are ways to make a claim answerable to evidence. They are not a recipe for manufacturing consensus.

Let optimism name its conditions

For an individual conversation, the useful move is smaller than auditing an entire industry. Ask the answer to expose its dependencies.

What has been observed, and what is forecast? Which benefits depend on choices about ownership, labor protections or access? Whose circumstances could make the conclusion wrong? What evidence would warrant a less optimistic assessment?

These questions do not guarantee a reliable response. The OpinionQA findings are a reminder that asking for a perspective does not ensure faithful representation of it. Where the stakes matter, an assistant's account needs comparison with evidence and with people who know the situation from a different position.

Still, the questions change what is being requested. Instead of asking for the correct emotional posture toward a technological future, they ask for reasons a person can examine.

That leaves room for real optimism: hope attached to conditions we can name, pursue and revise. It also leaves room for refusal where those conditions are absent. Neither position needs to masquerade as an inevitable conclusion.

The study's old models do not tell us what every assistant says now. Its survey scores do not tell us what anyone will believe tomorrow. What they offer is a useful interruption of the assumption that a conversational answer arrives without a point of view.

The next time an assistant tells you the future is promising, you need not demand that it become darker.

Ask it to become more specific.