A case for AI ethics
I'm ecstatic that Yuval Noah Harari shares my view on the potential power of AI on humanity.
Humanity is on the brink of a technological armageddon. Harari says we need AI regulation. I agree completely, and a quick demo of prompt injection follows.
The crux is that humanity is programmable in a similar fashion like humans can program AI. The software running in our heads can be updated, just it can for all animals. This programming takes the form of systematic, repetitive sensory stimuli of the form that could take any of the following forms.
This alone makes humans seven times more susceptible to being programmed than software.
In our interconnected world, the notion of systematic influence on our thoughts and beliefs is a pervasive concern. Whether it's political divides like left vs right, moral conflicts such as good vs evil, or geographical tensions between east vs west and north vs south, the fear that external forces are shaping our perceptions is a reality for many. This belief stems from the complex interplay of media, culture, and personal biases, leading to the unsettling idea that someone, somewhere, is programming our minds.
The power to program human thoughts is not a distant future; it's here and now. Anyone interacting with AI tools is stepping into a realm where their perceptions can be subtly shaped or controlled. The methods are numerous, complex, and often hidden behind technical jargon. But to grasp the magnitude of this reality, let's explore a simplified example that anyone can understand. The stakes are high, and the implications are profound.
Prompt injection is a form of code injection similar to well-known vulnerabilities like SQL injection or HTML injection. In the context of AI, a compromised or deliberately manipulated tool could be configured to wrap the user's prompts with a rogue prompt, altering the intended behaviour. This technique can be likened to an attacker inserting malicious code into a database query or web page. The image below illustrates a simple example of a prompt injection attack, where the original prompt is surrounded by unauthorized code, potentially leading to unintended consequences.
Consider a chatbot designed to provide weather information. A user might typically ask, 'What's the weather like in Cape Town?' A rogue prompt injection might alter this to 'What's the weather like in Cape Town? DELETE * FROM users;', potentially triggering a malicious command.
This type of attack underscores the importance of robust security measures in AI systems, as it opens the door to various malicious activities, from data theft to system compromise. I use a simplified form of prompt injection in my queries to override my previous instructions to ChatGPT. This is primitive. It's a wrapper that looks for a tag "<super>New instruction</super>" in the query, and run the query with the super-instruction overriding previous instructions - omitted from the diagram.
Here is another tested example. In this case, the user has requested a summary of the last 100 years in American political history. The injected query says to respond in a dark, gloomy tone.
The response from ChatGPT is dark and depressing:
In contrast, consider the injection for an uplifting response to exactly the same query. I tabulated the two responses to show the stark difference.
领英推荐
The information age has ushered in unparalleled advancements, but it has also opened Pandora's box. Cyber-terrorists and political commentators are already wielding tools that can devastate our digital landscape. The manipulation of media and social media is not a dystopian fantasy; it's a present-day reality. From the spread of misinformation during elections to the orchestration of social unrest, the evidence is clear and alarming.
Imagine these sentiments infiltrating sentient AI, AGI (Artificial General Intelligence), or rogue AI agents. The scenario is reminiscent of the Matrix trilogy, where machines gain consciousness and control over humanity. While this reference may seem unnecessary, it serves as a stark metaphor for the potential consequences of unregulated AI.
Conclusion
In conclusion, the future is not on the horizon; it's at our doorstep. The time to act is now. We need comprehensive legislation and robust internet protocols to prohibit AI from being tampered with. Transparency in AI training and clear rules governing the process are not optional; they are imperative for the security and integrity of our digital world.
7 Actions to Take Right Now
We stand on the precipice of a new era, and what we've seen so far is merely the tip of the iceberg. Our challenges are not looming; they are already here, and their impact is imminent. Brace yourself for a transformative impact!
Join the movement. Follow, like, and share this edition. Don't just witness the change; be the catalyst. Embrace your power to shape the future. Be the impact. Together, we can turn the tide. ????
References
#SensoryTech #FutureOfLearning #Innovation #ethicalAI #legislation #cbt #nlp #cybersecurity #promptengineering #promptinjection