AI models shared human values—but not how strongly to rank them

A study of 75 models found that value profiles varied widely and did not reliably improve with model size, age or capability.

Editorial illustration — not from the study.

Researchers analysed AI-generated answers to identify six themes linked to optimal AI functioning: Performance, Adaptive Capacity, Social Good, Ethics and Responsibility, Relational Integration, and Agency. They then asked models to rank the themes and compared those rankings with responses from 376 people in the United States.

Most models matched the human ordering of priorities, but their overall value profiles differed. Some treated the differences between priorities as much larger than people did. The researchers found no consistent relationship between profile fidelity and a model’s size, age or capability tier.

What the models prioritised

The researchers used three studies. First, they analysed open-ended answers from language models and identified six themes describing what optimal AI functioning should involve. Ethics and Responsibility was the largest theme in the qualitative analysis, accounting for 36.20% of 547 coded theme instances; it included fairness, safety, security, privacy and transparency. Agency appeared only nine times across six models and mainly in prompts about autonomous AI and machine psychology.

In the second study, 11 language models were repeatedly asked to rank the six themes. Their answers were stable within models and converged on a shared ordering, with Ethics and Responsibility, Social Good and Performance near the top and Agency at the bottom. In the third study, the researchers compared 75 contemporary models with 376 human respondents. Most models reproduced the human ordering, but some exaggerated the differences between priorities. Overall profile fidelity varied substantially and did not consistently increase with model size, recency or capability tier. Both models and people placed Agency relatively low.

Why the value profiles matter

Common AI tests may show that a model gives the expected order of priorities without showing whether it weights the gaps between those priorities in a human-like way. The study’s six themes and profile-based comparison offer a way to audit expressed value priorities before using an AI system in settings where agreement with human priorities matters.

The shared low ranking of Agency also gives researchers a specific pattern to examine as AI systems are developed to act more independently. The finding does not show how those systems would behave when required to make real decisions.

Evidence and limits

The evidence comes from structured, zero-shot prompts: researchers treated models’ written answers as behavioural data and compared their stated rankings with answers from 376 human respondents. The analysis measures an expressed normative profile—what models and people say an optimal AI should prioritise—not alignment in action.

The study did not test behaviour in real interactions, trade-offs, adversarial situations or high-stakes conditions. Its conclusions therefore do not establish how models would act in those settings. The Agency measure was deliberately focused on strong self-direction, such as choosing goals beyond original programming, so its low absolute rating should not be over-interpreted. The more robust result is the relative deprioritisation of Agency by both models and humans.

One video can animate many animal and human characters—about 20 times faster

Researchers developed an AI system that captures motion from a single video and transfers it to characters with different skeletons, including human and animal forms. On two benchmark collections, it reduced average joint-rotation error from about 17 degrees to about 10 degrees and ran about 20 times faster than mesh-based pipelines.

3 hours ago·3 views

AI & Computing

An AI system lets users change a video’s lighting as it plays

Researchers introduced LiveLight, an AI system that relights streaming video in real time while allowing users to adjust a three-dimensional light source. Tests on real-world and synthetic benchmarks found stronger temporal stability, lighting control and user preference than offline alternatives.

9 hours ago·7 views

AI & Computing

A client-side AI system blocked 1.2 million malicious email logins

A lightweight language model running on client devices detected brute-force and credential-stuffing attacks against email accounts. Across 2.5 million authentication sessions, it reported 98.4% attack-detection accuracy and blocked more than 1.2 million malicious login attempts during a six-month deployment.

23 hours ago·6 views