Being Human in era of AI
By Rajiv Bahl / November 29, 2024
"Homo sapiens" means 'wise man' or 'knowledgeable man' (man being gender neutral). With the advent of AI and the mushrooming of various models, especially after the democratisation of AI with the release of ChatGPT in 2022, multiple attempts are being made to define "what is acceptable behaviour?". In this context, the work ofAletheia Framework™is very relevant. Attacks like Prompt Injection, which consistently keep appearing on the top of the charts once again in the recently released OWASP's 2025 Top 10 Risk & Mitigations for LLMs and Gen AI Apps makes us think that something more fundamental is required which runs across the models even beneath the foundation models as a common thread.
Keeping the above context (no pun intended), three ideas are worth exploring.
Concept 1: Establishing a Trusted Computing Base for AI – Reference Monitors and Kernels in AI
Reference Monitor is that part of the TCB that validates access to every resource before granting access requests. A reference monitor stands between every subject and object verifying that a request is valid and meets the "defined criteria" before being enabled. In this case that fundamental criterion would be "Being human". This would ensure that every prompt goes through a set of validations to check that it does not violate the "Being human" principles. This would mitigate Bias, Discrimination, Stereotyping, hatred, abuse, and profanity to a good extent. Instead of struggling to find ways and means to mitigate prompt injection in the user space if we are to address it in the system space or kernel then it will be much more effective.
Concept 2: Establishing the root of trust in AI
Hardware is done by Trusted Platform Module (TPM). Instead of trying to attempt to solve the problem in software space if we were to seek the same solution in hardware, it would be more fundamental and therefore more effective. It will make breaking "trustworthiness in AI" more difficult.
Concept 3: What is Trust?
Expanding the definition of "Trust" in Trusted Platform Module (TPM) to adopt the OECD principle of "trustworthy" by defining what it is to be a human in a Human Platform Module (HPM). Going beyond classical definition of Trust as access control alone there is a need for embedding (no pun intended) principles of trustworthy AI as defined by OECD AI Principles considering inclusiveness, sustainability, democratic values, fairness, accountability, responsibility, privacy, transparency, explainability, robustness, security and safety. Embedding these in hardware will go a long way in getting the desired behaviour which is difficult to compromise in the user space. I am hopeful that some of the above ideas will be adopted by innovative and pioneering chip makers to make it happen for a better, secure and safe society making us truly what we are "wise-man" (gender neutral).
I am hopeful that some of the above ideas will be adopted by innovative and pioneering chip makers to make it happen for a better, secure and safe society making us truly what we are "wise-man" (gender neutral).