So, let’s talk a little bit about AI bias.
Algorithms can reveal and help correct biases in decision-making, according to a study from Proceedings of the National Academy of Science. Participants in the study were more likely to recognize bias in algorithmic ratings compared to their own ratings, even when the algorithms were trained on their own decisions. Algorithms remove the bias blind spot by presenting decisions in a way that is more similar to how people evaluate others’ decisions. The study also found that participants were more willing to correct biases when they were attributed to algorithms, resulting in less biased final ratings.
A surge of companies claim to offer hyper-accurate deepfake detection services, but their capabilities are largely untested. Deep Media, a rising star in the field, has won military contracts worth nearly $2 million but lacks subject matter expertise and has a sole machine learning engineer with an undergraduate degree in astrophysics. While the demand for deepfake detection is high, the effectiveness of current detection tools is questionable, as they can be easily fooled. Nonetheless, about 40 companies offer deepfake detection services, each claiming high levels of accuracy.
Researchers at Anthropic have made a breakthrough in understanding large language models by using “dictionary learning” to uncover patterns in neuron activation. They identified features linked to specific topics and found that manipulating these features could change the AI system’s behavior. While there is still much work to be done, this progress in interpretability could help address concerns about bias, safety risks, and autonomy in AI systems.
Why do we care?
The Anthropic news is notable because of the advancements in understanding what happens within the black box of generative AI. Just note it’s happening.
The surge of detection start-ups risks providing a false sense of certainty and eroding public confidence in authentic media. Take your skeptical eye here to the solutions.
And note that there are solutions – that’s why the research matters. All this is input to your discussions with clients about best using these technologies.
