Show a computer a stop sign and the answer seems straightforward. Show it a bowl of fruit, a sweetened smoothie and a packet of crisps, then ask which pictures represent healthy food, and the question becomes surprisingly difficult. Google's research into AI grey areas asks whether a machine can help people decide where such boundaries belong instead of silently imposing its own definition.

In a November 2026 Google Research explanation, a prototype called Agile Deliberation helps users define subjective image categories by considering difficult examples and refining the rules. It was tested in 18 live sessions lasting about an hour and a half each. The findings suggest that interactive discussion can produce more consistent classification in the study's tasks, although the prototype is not a universal judge of what is fair, healthy or acceptable.

AI grey areas do not come with universal answers

Training a computer to recognise a particular object is often described as a matter of showing it enough labelled examples. The trouble starts when the label is not really a fact about the pixels. 'Inspiring artwork', 'misleading promotion' or 'healthy meal' can depend on context, culture, purpose and individual priorities. Two reasonable people may disagree without either misidentifying what is in the picture.

A human moderator might begin with a simple rule, then discover that the real task contains many exceptions. A picture advertising an online game may include a reward without being deceptive; another may use the same imagery to imply that a player will receive money they cannot actually obtain. A fixed checklist can miss the difference unless someone thinks through the circumstances.

This is a deeper issue than computer accuracy alone. If a developer chooses one definition of a subjective category and hides that decision inside a model, the system may produce consistent answers that do not reflect what the user intended. The software can be technically good at following the wrong rule. Agreement on the rule is therefore part of the task.

First decide what belongs inside the category

Agile Deliberation starts by helping the person break a broad idea into more concrete examples and exclusions. If the task is to recognise healthy food, the user might identify salads, vegetables or particular types of balanced meal while also specifying cases that should not automatically qualify. The aim is to expose the definition before the model begins applying it at scale.

That first stage is useful because most people do not have a complete policy in their heads. They may know what they mean when they see a clear example, yet struggle to write a rule covering every case. The system prompts them to express both positive examples and boundaries. A well-chosen exclusion can be as informative as a preferred case.

However, describing a category in plain words does not eliminate disagreement. A particular food may fit one person's dietary requirements and not another's. A moderator's priorities may be shaped by a company's policy rather than personal taste. The technique can clarify whose definition is being used, but it does not decide whose values ought to govern a public service.

This is the crucial distinction between making a system more aligned with its user and making it objectively right. In a private image collection, personal preferences may be appropriate. In employment, education or public administration, a disputed classification may demand transparency, appeal rights and external scrutiny, whatever the user of the software prefers.

The clever move is showing a borderline example

The second stage confronts the user with pictures close to the current rule's edge. A fruit smoothie provides a useful illustration: it contains fruit, yet its ingredients and sugar content may matter to the person's definition. Seeing the example forces a decision that might never have occurred while drafting an abstract policy.

Google's system uses models that connect images with text descriptions to identify such troublesome cases and help refine the classifier. The user can respond to what is actually shown, rather than guessing which wording might persuade a model to behave differently. The software then incorporates that feedback into a revised set of instructions.

That is a different role for artificial intelligence. Instead of being asked to issue a judgement and move on, it is used to expose ambiguity in the human's own thinking. If the borderline images are chosen well, the exercise can reveal hidden assumptions, gaps in a policy and situations the original specification overlooked.

There is also a limitation. A system trained on particular images and language patterns may fail to surface culturally unusual or genuinely new examples. The person supervising it may accept the familiar cases without considering those who will be affected outside the testing environment. Better questions improve deliberation, but they do not guarantee that every relevant perspective appears in the room.

What the live study found

The research team evaluated the approach through 18 live participant sessions, each lasting roughly 90 minutes, rather than relying only on a static set of labelled pictures. The tasks involved identifying healthy-food images and a form of misleading gaming or entertainment promotion described as pay-to-play. These were chosen partly because the boundaries could shift as participants worked through examples.

Participants tried the researchers' method alongside alternatives that involved an initial one-off description, automated decomposition or more manual refinement. In the underlying research paper, the authors report higher classification scores than their tested baselines, including a 7.5% improvement in F1 against automated decomposition and an improvement of more than 3% over manual deliberation. F1 is a technical measure balancing two types of classification error. It is not a percentage of people helped or a guarantee of accuracy on unrelated tasks.

The team also reports that participants found the structured approach less tiring and stressful than the manual process. One reason was that the system surfaced difficult cases, so the user did not have to invent endless search terms or rewrite a prompt after every error. Those are encouraging results for interface design, especially when a category requires repeated judgement.

But this was a small interactive study of a prototype in two areas. It does not establish that an AI-assisted policy created in a short session will remain suitable over months of real-world use. Nor can it settle the ethical disagreement about what an organisation ought to classify as acceptable. More extensive trials would need to examine diverse users, changing conditions and consequences for people whose content is mislabelled.

A consistent mistake can still be unfair

Suppose a content platform improves consistency in identifying images that it considers misleading. That may reduce arbitrary variation from one moderator to another. Yet if its definition systematically misinterprets a particular community's humour or cultural conventions, consistency merely makes the error repeatable. Fairness requires examining the rule as well as how faithfully the machine follows it.

That point connects with research covered by LiveAIWire on how social norms differ from simple AI predictions. Human expectations are shaped by circumstance, and a pattern found in data does not automatically justify a rule. A system that helps users explore edge cases may make disagreements clearer, but the organisation still has to explain whose judgement takes precedence.

There is a related problem in measuring progress. Our reporting on AI bias audits that disagree with one another illustrates why a single performance score cannot represent every dimension of a contested decision. A classifier may perform better under one benchmark while creating a different kind of harm in practice. Measurement is useful when the people setting the standard can defend its relevance.

In settings with serious consequences, the ability to challenge a decision must remain visible. Someone whose image or account has been incorrectly restricted should not need to understand the underlying machine-learning system to seek review. Explaining the category and providing a meaningful appeal can matter as much as refining the model.

The computer can help a human discover the question

The most intriguing aspect of the project is not the possibility of automating more content judgements. It is the possibility of improving the conversation before automation. A team with vague rules can test concrete examples, discover disagreements early and document the reasons for changing the policy. Used carefully, AI can help make human assumptions easier to examine.

That lesson travels beyond images. A firm defining a 'high-priority complaint', a school deciding what counts as inappropriate material or an online marketplace assessing suspicious listings may all encounter borderline cases. The exact Google prototype was not tested across all those domains, so they remain applications to investigate rather than proven benefits. The common issue is how to make ambiguous rules explicit.

LiveAIWire's coverage of persuasion and fairness in AI decisions raises a useful companion warning: presenting information in a particular way can influence what people think is fair. A deliberation tool should help people understand the options, not steer them towards a convenient institutional answer while appearing neutral.

The familiar dream is a computer that makes difficult judgements instantly. Google's work points to another possibility: a computer that slows the process just enough to reveal why people disagree. In subjective situations, the breakthrough may not be removing the human from the decision. It may be giving that person a better view of the boundary they are drawing.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.