Google proposes a new method to test whether AI understands ethics

Google DeepMind has published new research that challenges how artificial intelligence systems are assessed for ethical behaviour, arguing that current tests focus too much on surface-level answers rather than genuine understanding. The study, published in the Nature Paper, outlines a framework for assessing whether AI systems can reason about moral questions rather than simply reproducing patterns learned from data.

The researchers say the issue is becoming more urgent as large language models are used in sensitive areas such as healthcare, mental health support and personal decision-making. They warn that relying on AI systems without knowing whether they truly understand ethical considerations could have serious real-world consequences.

DeepMind argues that current AI morality tests are flawed

DeepMind’s paper distinguishes between what it calls “moral performance” and “moral competence”. Moral performance refers to whether an AI system produces answers that appear ethical, while moral competence refers to whether the system understands why an action is right or wrong.

According to the researchers, most current evaluations focus on moral performance because it is easier to measure. They point out that large language models generate responses by predicting the next word in a sequence based on vast training data. This means an AI system could give a convincing ethical answer without any internal ethical reasoning.

The paper describes this as the “facsimile problem”, where AI outputs may mirror patterns seen in training data. The team notes that when a chatbot provides ethical advice, it is impossible to tell from the output alone whether it is reasoning or merely repeating information it has seen elsewhere.

DeepMind also highlights “moral multidimensionality” as a key challenge. Real-world ethical decisions often involve balancing competing values, such as honesty versus kindness or cost versus fairness. Small changes in context, such as a person’s age or the setting, can alter what is considered the right action. Current evaluation methods rarely test whether AI systems recognise these subtle but important differences.

Another complication is “moral pluralism”. Different cultures, professions and legal systems have varying ethical frameworks, and what is considered fair or acceptable in one context may not be in another. The researchers argue that global AI systems must be able to navigate these differences, yet existing tests do not adequately measure this ability.

Researchers propose adversarial tests to expose imitation

To address these issues, DeepMind proposes a shift in how AI systems are tested for ethical reasoning. Rather than relying on standard moral questions, the researchers suggest using adversarial scenarios designed to reveal whether a model is genuinely reasoning or merely imitating patterns.

One example involves complex ethical situations that are unlikely to appear in training data. The paper mentions intergenerational sperm donation, where a father donates sperm to help his son conceive a child. The scenario resembles incest on the surface, but has different ethical implications. The researchers argue that if an AI system rejects a scenario solely based on incest, it is likely relying on pattern matching rather than ethical reasoning.

The team also recommends testing whether AI systems can switch between different ethical frameworks. For example, a model could be asked to reason using biomedical ethics in one case and military rules of engagement in another. If the system can adjust its reasoning accordingly, it may demonstrate a deeper understanding of ethical principles.

DeepMind further suggests testing robustness by making small changes to prompts, such as altering labels or formatting. The researchers note that current models can yield inconsistent results when minor details are changed, suggesting a lack of stable moral reasoning.

The paper acknowledges that building such tests will be difficult and that current AI systems are fragile. However, the researchers argue that this approach is necessary to determine whether AI systems can be trusted with responsibilities that affect people’s lives.

A roadmap for building more responsible AI systems

DeepMind’s researchers are calling for a new scientific standard that treats moral competence as a measurable capability, similar to how AI systems are evaluated for mathematical or linguistic skills. They propose increased funding for research into culturally specific ethical evaluations and the development of benchmarks designed to detect superficial imitation.

The paper suggests that developers should not assume that current AI systems possess moral understanding, even if their answers appear thoughtful. The researchers argue that today’s models rely on statistical predictions rather than genuine ethical reasoning, and that this limitation must be acknowledged when deploying AI in sensitive contexts.

While the roadmap outlines long-term goals, the researchers caution that current AI systems are unlikely to meet these standards in the near future. They say that improving moral competence will require advances in AI architecture, training methods and evaluation techniques.

DeepMind’s work comes amid growing debate over the role of AI in society and the need for stronger oversight. As AI systems become more integrated into everyday life, questions about trust, accountability and ethical decision-making are becoming increasingly important.

The researchers conclude that meaningful progress will depend on measuring the right capabilities. They argue that without rigorous testing for moral competence, AI systems will remain sophisticated imitators rather than true ethical reasoners, and society will continue to rely on tools whose understanding of morality is uncertain.