Game theory as a way to control AI

Hello, this is Elena Kuznetsova, automation specialist at Sherpa Robotics. Today I translated for you an article on the risk of bias in artificial intelligence proposals. This is a serious problem, because what we teach the AI ​​is what we get in its answers.
In a previous article, one of the commentators answered my colleague, Yulia Rogozina, that with the development of AI and the acceleration of obtaining information, “It will become easier to make mistakes and the mistakes will become larger!” What if errors and biases are built in from the beginning, during the learning process?
I invite you to read the article, which emphasizes the importance of a deliberate approach to AI training.

Refusing offers from artificial intelligence is likely an attempt to force it to be more generous. Artificial intelligences often learn from materials created or curated by humans. This creates significant difficulties in preventing the reproduction of the biases of these individuals and the society to which they belong. With AI increasingly being used to make medical and financial decisions, the risk becomes even more significant.

However, researchers at Washington University in St. Louis have discovered an additional aspect to this problem: People involved in AI training may change their behavior with the awareness that this behavior may influence the AI's future decisions. In some cases, these behavioral changes carry over into situations outside of AI training. This discovery highlights the complexity of human-machine interactions and the importance of a mindful approach to AI training, which could have serious real-life implications.

Want to play a game?

A recent study immersed participants in a simple version of game theory, where volunteers received $10 and had to make offers to each other to split the amount. One of the participants offered a certain share of the money, and the other could either accept or reject this offer. If the second person refused, no one received anything.

From the point of view of pure rational economics, it is logical to accept any offer, since it will always lead to more income than if it were refused. However, in practice, people often reject offers that deviate greatly from equal distribution because they perceive them to be unfair. This decision allows them to punish the one who made the unequal offer. Although cultural differences can influence perceptions of fairness, this effect has been repeatedly confirmed in various studies.

An interesting feature of the work carried out by Lauren Treiman, Chien-Zhu Ho and Wouter Cool was that some participants were told that their partner was an artificial intelligence. The results of their interactions were to be used to train the AI.

This clarifies an implicit aspect of the classical game theory approach: refusing offers can help partners understand which offers are considered fair. Participants, especially those who were informed that they were training an AI, could easily infer that their actions would influence future AI offerings.

The researchers were interested in the question: will this information affect the behavior of the participants? They compared the results with a control group that simply took a standard game theory test.

Equity of learning

Researchers Treiman, Ho and Cool pre-registered a number of multivariate analyzes they planned to run on the data. However, the results of these analyzes were not always consistent across experiments, possibly because there were insufficient numbers of participants to detect relatively subtle effects with sufficient statistical confidence, and because the large number of tests may have resulted in a few positive results occurring by chance.

In this paper, we will focus on a simple but important question: does knowing that you are training AI change human behavior? This question was investigated through a series of similar experiments. One of the key differences was whether the information that participants were training the AI ​​was displayed with a camera icon, as people are known to sometimes change their behavior if they know they are being monitored.

The answer to this question is clear: yes, people do change their behavior when they think they are teaching AI. In a number of experiments, participants were more likely to reject unfair offers if they were told that their sessions would be used to train an AI. In some cases, they were also more likely to reject what were considered fair offers (for example, in US populations, rejection of offers with a 70/30 split, where the offerer gets $7, increases significantly). The researchers believe this is because people were more likely to reject borderline “fair” proposals such as a 60/40 split.

Interestingly, this behavior persisted even despite the economic cost to participants of refusing offers. Moreover, they continued this behavior even when they were told that they would never interact with the AI ​​after training was completed, meaning they would not personally benefit from changing the AI's behavior. This indicates that participants were willing to make financial sacrifices to train AI that would later benefit others.

Impressively, in two of the three experiments that included follow-up testing, participants continued to reject offers at increased rates two days after participating in the AI ​​training, even when they were told that their actions were no longer being used for training. This suggests that participation in AI training could, in fact, change their own behavior.

Of course, this will not apply to all types of AI training, and much of the work involved in preparing training materials, such as large language models, is often done without realizing that they can be used to train AI. However, in some cases people participate more actively in learning, and it is worth considering that this is another way in which bias can creep in.

Comment

One of the directions of our company is the creation of corporate smart chat bots (neural employees) using the Sherpa AI Server platform.

And yes, we warn our customers about the importance of providing the most accurate and reliable information on their business processes for training neuroscientists.

This is extremely important as chatbots are used in technical support, HR, sales, legal decision-making and many other aspects of management.

On average, our clients are more than 80% satisfied with the responses of smart robots. However, we have room to grow and are working on the quality of our neuroscientists and the accuracy of their answers.