The rapid rise of Artificial Intelligence (AI) in the past year has left many questioning its true capabilities and potential implications. Is AI the next big tech fad or a force that could potentially enslave the human population? The answer is not so straightforward. While recent achievements, such as ChatGPT passing the bar exam, are impressive, concerns about the limitations and trustworthiness of AI have also started to emerge. For instance, a lawyer found that ChatGPT fabricated elements of their arguments in court, raising questions about its reliability. In order to move forward with AI adoption, we must address key concerns surrounding trust and transparency.
Trust in AI involves more than just the accuracy of its output. We must also consider the potential biases, censorship, and manipulation that could be embedded within AI models. This is particularly critical for AI systems that will be used in safety, transportation, defense, and other areas where human lives are at stake. While national agencies recognize the importance of AI integration, it is crucial that adoption is approached with caution and careful focus.
Crucial Questions to Answer
To establish trust in AI, two fundamental questions must be addressed:
- Is a particular system using an AI model?
- If an AI model is being used, what functions can it command/affect?
Answering these questions eliminates a significant number of risks associated with AI misuse. By understanding the purpose for which an AI model has been trained and its deployment context, we can establish transparency and accountability.
Verifying the trustworthiness of AI involves various methods that analyze the hardware, system, and data being utilized. These methods include:
- Hardware Inspection: This physical examination of computing elements aims to identify the presence of chips used for AI.
- System Inspection: Using software analysis, this mechanism determines the AI model’s control capabilities and identifies any off-limits functions. By examining a system’s transparent components without revealing sensitive information, AI processing can be detected.
- Sustained Verification: Following the initial inspection, sustained verification ensures that the deployed AI model remains unchanged and untampered. Anti-tamper techniques, such as cryptographic hashing and code obfuscation, are employed to preserve data integrity and protect against unauthorized modifications.
- Van Eck Radiation Analysis: This technique examines the radiation emitted during system operation. By detecting major changes, such as the introduction of new AI components, without revealing sensitive information, it enhances the detection of potential tampering.
Data verification is a critical aspect of ensuring trust in AI. The training data fed into an AI model must be thoroughly verified at the source to prevent manipulation and biases. For instance, past data used to train an AI model for predicting the quality of potential employee candidates resulted in a sexist model. The dataset consisted mainly of high-performing male employees, which unintentionally led the model to favor male candidates. This highlights the need for representative and unbiased training datasets to prevent skewed outcomes.
To create safe, accurate, and ethical AI systems, verifiability and transparency are essential. Zero-knowledge cryptography can be utilized to prove the integrity of data and ensure that it has not been manipulated. Business leaders must have a high-level understanding of verification methods and their effectiveness in detecting AI usage, model changes, and biases in training data. Implementing these methods is the first step towards building a shield against potential threats, such as disgruntled employees or industrial/military spies, as well as addressing human errors that could have dangerous consequences.
As AI becomes increasingly integrated into our daily lives, establishing trust in its capabilities is of utmost importance. Verifiability and transparency are crucial elements that must be addressed to ensure the safe and ethical application of AI. By employing various verification methods and carefully scrutinizing training data, we can mitigate risks associated with AI misuse. Trustworthy AI has the potential to revolutionize industries and improve lives, but only if we can trust it implicitly.