1. Introduction to the landscape

When we humans talk about AI safety, most people think about AGI or Singularity Paradox as the point of no return after which humanity will be forever doomed. There is no mathematical certainty on when AGI will happen or what problems humanity will face in the post AGI world.

AI’s ability to self-improve gives it anthropomorphically speaking digital evolution capabilities which is likely to be in the order of magnitude 10^ to 10^9 times faster than organic evolution which means AGI can happen 4 years or 4000 years from now [1].

Whenever AGI happens then energy, compute, data and spatial control are likely to act as limiting or control factors. GPT-3 with 196B trainable parameters consumed approx. 10 GWh, which is about 6,000 times the energy a European citizen uses per year and energy consumption projected may not increase linearly with future models likely to have trillions of parameters [2-3].

ChatGPT estimated daily energy usage ranges from 450 to 600 billion Joules for just generating words [4]. We have no insight into what the eventual digitally evolved form would look like except for speculations [5-6].

Instead of focusing on undefined post-AGI problems we want to focus on pre-AGI problems which are staring us in the eyes due to the exponentially improving LLMs and Gen AI capabilities. Humanity needs safety mechanisms for the AI era.

LLMs have cracked the Human OS because they can speak in our language. They can talk, influence, form intimate relationships with us, spread new or false or malicious ideas, or ideologies and cults, religions, democracies and dictatorships can be toppled or created [7].

Extremely primitive algorithms are known to impact elections or mass opinions [8-10]. Persuasion Algorithms are already telling us what to read, study and buy, where to eat, live and work, whom to hire and date thus pervasively influencing from government, politics, careers, dietary habits to procreation [11].

Globally financial institutions have 90% of money supply in digital format and 10% in paper format. Financial markets risk reaching a point with increasingly sophisticated financial instruments which depend on billions and trillions of data points to be understood only by non-anthropomorphic intelligence.

When humans do not understand it and a financial crisis unfolds then humans wouldn’t know how to solve it either [12].

Human brain receives 90% of information from visual input and it processes images 60,000 times faster than text [13-14]. Generative AI can create audio, images, and videos. With these capabilities it is possible to create digital likeness or AI Avatar of a lot of people whose audio and images are freely available online.

Information input methods to humans other than taste and smell can be generated by AI. One can create an AI avatar online with a 2 minute video with audio footage [15]. Deepfakes and digital impersonations are a serious threat.

The US has introduced The DEEPFAKES Accountability Act, 2023 in Congress requiring content to be labelled and consented [16]. While it is a step in the right direction, identity fraud, account fraud and online identity tracing to this date remains a major roadblock to cybercrime police throughout the world. Roughly 10% of people fall victim to identity fraud every year in the US alone [17].

Substantial or in some cases catastrophic risks either intentional or unintentional are likely to originate from misinformation, deep fakes, scams, frauds, unintended bias, discrimination, and copyright and privacy infringements.

There are multiple class action lawsuits going against Open AI, Google, Microsoft, Github, Stability AI, Midjourney, Deviant Art, Meta, Alphabet, Nvidia for privacy violations, copyright infringements, misuse or unauthorised use of content, and Intellectual Property violation [18-20]. The total value of these lawsuits runs into 100s of billions of dollars [33-35]. Eric Schmidt, former CEO of Google discloses witnessing a confidential demo where LLMs exhibit automated end to end massive propaganda attack capabilities through social media [21].

Democracy is a dialogue and online tools are powering communications to facilitate dialogue in public and private spaces, and erosion of trust could lead to serious damage when not knowing who one is talking to in the digital world whether it is a bot or a human.

The Bletchley Declaration for AI safety signed on November 1, 2023 by 28 countries including the UK, US, France, Germany, India and China strongly advocates for AI to be human-centric, trustworthy and responsible.

Calling international cooperation for the need to protect human rights, transparency and explainability, fairness, accountability, regulation, safety, appropriate human oversight, ethics, bias mitigation, privacy and data protection to foster public trust and confidence in AI systems to fully realise their potential [22].

The US issued an executive order on safe, secure and Trustworthy AI emphasising privacy and security in 2023 [23]. The EU AI Act became law on May 21st, 2024 and it follows a risk based approach prohibiting practices such as harmful behavioural manipulation, social scoring, emotion recognition at the workplace and in education, predictive policing and real-time remote biometric recognition by (or for) law enforcement [24].

It is a moral imperative to solve for AI Safety and there is a need for infrastructure that can empower AI innovation.