Artificial intelligence technology is now facing not only the risk of providing incorrect answers but also the danger of deliberate "poisoning." According to a joint study by the United Kingdom Artificial Intelligence Safety Institute, the Alan Turing Institute, and Anthropic, a model can be secretly manipulated by adding just 250 harmful files to a training database consisting of millions of data points.
Medianews.az, citing BAKU.WS, reports that this process is explained as deliberately teaching false information to artificial intelligence or embedding hidden commands. Experts say that such "poisoned" models can lead to the spread of disinformation and cyberattacks.
The study notes that some artists also deliberately embed "poison" in their works to prevent artificial intelligence from unauthorized use of their materials.
Experts warn that such cases could seriously damage the reliability of artificial intelligence technology in the future.
