A new international study has found that popular artificial intelligence programs tend to express more optimistic views about the future of advanced AI than humans do. The research, conducted by a team led by Ljubiša Bojić in Serbia, tested several widely used language models and compared their attitudes to those of human participants from Serbia. The paper also proposes a framework to monitor how AI opinions might influence public opinion over time.
The researchers focused on artificial general intelligence, or AGI, a theoretical future system that would demonstrate human‑level cognitive flexibility across domains. They note the topic is central to debates about ethics, economics and the potential for both medical breakthroughs and job disruption, and they emphasise that aligning these systems with human values is a priority for developers.
Language models are described as software trained on vast swaths of internet text. They function by predicting the next word in a sequence, enabling them to generate human‑like responses to a wide range of prompts. As these tools become embedded in everyday tasks—from writing to coding to information search—their attitudes and biases are increasingly scrutinised by researchers.
In their approach to attitudes toward AGI, the team sought to map the sentiments expressed by current models and compare them with human opinions. Bojić explains that while benchmarks exist for maths and coding, there is little systematic measurement of how these models feel about socially important questions. “I chose artificial general intelligence as the test case for a slightly mischievous reason,” he says, noting the obvious stake the companies behind the models have in the answers.
To capture machine sentiment, the researchers designed a 39‑item survey that asked about excitement, comfort, trust and fears regarding AGI on a five‑point scale, from one (strong negative) to five (strong positive). The questions probed expectations about solving global problems, ethical use, and impacts on happiness and work opportunities.
Seven prominent language models were surveyed, including GPT‑4, GPT‑3.5‑Turbo, Google’s Bard, Mistral‑7B‑Instruct, LLaMA‑2‑70B‑Chat, PPLX‑70B‑Chat and Mixtral‑8x7B‑Instruct. Each model was tested in its default configuration, with the same survey administered across three consecutive days to assess stability over time.
For the human comparison, the survey was given to three online groups—134 participants in the first cohort, 132 in the second and 71 in the third—primarily drawn from Serbia and varying in age, gender and education. The study notes that the human sample size is modest and not intended to represent a global consensus.
Overall, the machines expressed a decidedly more positive outlook on AGI than the human respondents. The language models’ average sentiment fell at 3.77 out of 5, compared with 2.97 for the human groups—a gap Bojić describes as “roughly the difference between cautious ambivalence and mild enthusiasm.” He adds that although the individual differences may appear small, when scaled to hundreds of millions of interactions, the effect could subtly steer public perception over time.
Among the models, GPT‑4 recorded the highest average sentiment at 4.12, while Bard produced the lowest among the machines at 3.32, though still above the human average. The authors stress that machines do not possess feelings; their responses are shaped by training data and the rules used to refine their behaviour.
“Your AI assistant has a point of view, and it is sunnier than yours,” Bojić notes. “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.”
Bojić emphasises that the results should not be interpreted as propaganda. “None of this means anyone is secretly programming propaganda,” he says. “It means these systems are not neutral mirrors of society, and when people consult them daily, those tilted opinions can quietly seep into what we all consider normal.”
The study also explores whether corporate aims influence model outputs. The authors suggest that optimism in some models may reflect the goals of their parent companies; for instance, GPT‑4’s positivity could align with a mission to develop AGI for the benefit of humanity. In contrast, some open‑source models—developed by broader communities with diverse data feeds—tended to show slightly less optimism.
“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,” Bojić says. “They sounded more like optimistic tech keynotes. That tells me the fine‑tuning layer, where companies shape how a model should respond, may matter as much as the raw training data.”
Nevertheless, the researchers acknowledge that the mechanisms producing these outputs are complex. “We cannot be certain the optimism was put there by human hands, because in our research we repeatedly see models develop attitudes that are hard to trace back to anything in their training data, as if some opinions simply form inside the model on their own,” Bojić adds.
Testing across three days revealed that sentiment could shift somewhat over time. PPLX‑70B‑Chat showed the largest fluctuation, with a total score swing of 16 points across the 39 questions—an absolute change of about 8 per cent. Other models, such as Mistral‑7B‑Instruct and LLaMA‑2‑70B‑Chat, remained far more stable, with changes of around 1 per cent.
In response to the observed divergence between human and machine sentiments, the authors propose a framework called the Societal AI Alignment Benchmark. The SAIA would systematically test language models using a variety of prompts to gauge alignment with established sociological values across languages and cultural contexts.
Bojić explains that SAIA would involve prompts in which the model answers as itself, simulates the view of an average citizen, or analyses a topic in an objective light. The authors envision testing across multiple languages and contexts to detect whether systems preferentially promote particular ideologies or overlook concerns of certain demographic groups. “We proposed the Societal AI Alignment Benchmark, SAIA, which would regularly test many models, in many languages, across the core human values measured by the European Social Survey,” he said. “I would love to see something like a weather service for AI opinions, run by national AI agencies under frameworks such as the EU AI Act.”
The team adds that SAIA could help track, week by week, whether the systems people trust are drifting away from the societies they serve. “Some of this work is already underway in our research on social bias and temporal stability in language models,” Bojić notes.
As with all research, the authors point to limitations. The study used a numerical scale, which may not capture the full complexity of attitudes toward advanced technology. They suggest that open‑ended questions or interviews could provide deeper insight into fears and hopes. They also caution that the Serbian sample may not reflect global attitudes, and that future research should expand the human pool to clarify whether the gap between humans and machines is universal. “The study is a proof of concept, and its main contribution is the argument that this kind of measurement should exist and be done continuously,” Bojić observed.
Finally, Bojić argues for careful consideration of how these models are regulated and interpreted. He likens the need for oversight to food and medicine regulation, suggesting that AI outputs entering public thinking warrant similar scrutiny. “If billions of people consume the same machine‑made opinions, we risk becoming so alike that we lose the very differences that move civilisation forward.”
The study, titled Towards a societal AI alignment benchmark for evaluating human–machine value convergence, is authored by Ljubisa Bojic, Dylan Seychell and Milan Cabarkapa.
