AI in home surveillance can be “quite dangerous”

Researchers have discovered that AI systems for home surveillance act completely randomly. However, the reasons for this remain unclear due to a lack of transparency. The background.

A new study from MIT and Penn State University shows that the use of artificial intelligence (AI) in home surveillance could lead to major problems. Language models such as GPT-4 are inconsistent and therefore act unpredictably. They often recommend calling the police even if no criminal activity is apparent. This inconsistency could be dangerous.

Even more problematic: the models react differently to similar scenes. For example, a video showed a car theft and the model raised the alarm, while a similar video did not elicit any reaction. These differences raise the question of how reliable AI really is. This can have fatal consequences, especially in sensitive situations.

AI Bias in Home Surveillance

Another problem lies in the way the algorithm thinks. AI shows unconscious biases. In neighborhoods with predominantly white residents, it raised the alarm less often, even when similar activities were seen as in other areas. This shows that AI is influenced by demographic differences. The exact reasons are unclear, however, as the models' training data is often not publicly available.

These prejudices could have serious consequences when used. It is impossible to predict how many people could be treated unfairly as a result. However, it is clear that caution is required when AI is used in sensitive areas.

The need for caution

The results of the study underline the risks associated with the use of AI in home surveillance. That is why we should not simply trust the technology blindly. Rather, we need more transparency and control. This is the only way to ensure that AI works fairly and reliably.

The future of home surveillance could be heavily influenced by AI. But before the technologies can be used in everyday life, it must be ensured that they actually work.

Also interesting: