There’s only one question that people really care about when it comes to AGI: when will it arrive?
And the only honest answer to that question is: it doesn’t matter.
I know that’s not the answer anyone wants, but it’s the right one. Maybe AGI gets here today or tomorrow. Perhaps it takes a year or even a decade. Heck, AGI might not even be possible.
Ethically speaking, whether AGI exists is irrelevant.
We’ve entered the “Schrodinger’s AGI” era of development. It never mattered whether the cat in the box was alive or dead. What mattered was that we couldn’t know until we opened the box.
In the world of AI, as long “the emergence of an AGI model” remains sealed inside a box labelled “right around the corner,” lawmakers, venture capitalists, and nine out of the ten largest companies in the world by market cap stand to make money on the speculation alone.
As long as nobody opens the box, and AGI remains right around the corner, then the stakes are too high not to develop these machines with wild abandon. There are very serious people in the world of science who believe the human species will enter a period of unprecedented prosperity within years of an AGI emerging.
These people are unlikely to be moved over fears of long-term mental health harms or employment displacement. You gotta break a few eggs to make an omelet, right?
In a November BBC interview, for example, Google CEO Sundar Pichai opined that “AI is the most profound technology humanity is ever working on, and it has potential for extraordinary benefits, and we will have to work through societal disruption.”
Meanwhile, OpenAI CEO Sam Altman intends to loosen the restrictions on his company’s models generating “erotic” content in order to drum up more business for ChatGPT.
Yet no long term studies on the potential effects of AI-wrought job displacement, or Altman’s plan to automate smut, have been conducted by the firms developing these products and services — or by anyone else, for that matter. Modern reinforcement learning and transformer techniques only date back about a decade.
In fact, even short-term studies or due diligence appear to have been skipped in lieu of rapid iteration.
This means that, at any given time during an AI system’s life cycle, a catastrophic failure state or undefined harm vector could emerge unnoticed.
So here’s my ultimate takeaway when it comes to the ethics of AGI development: fixating on the possibilities is shortsighted. We will experience the consequences of the ethical choices made by the companies developing advanced AI products whether AGI emerges or not. So let’s treat AI and AGI development with the same ethical focus.
And, importantly, this focus should include the entire life cycle of an AGI model, including the development and maturity of pre-AGI models.
NIST defines the AI life cycle as including the following domains:
- Plan and Design
- Collect and Process Data
- Build and Use Model
- Verify and Validate
- Deploy and Use
- Operate and Monitor
- Use or Impacted By
A quick glance at the above bullets tells you that the biggest tech outfits developing advanced AI models tend to skip “verify and validate” and move right on to “deploy and use.” Worse, once deployed, these models are fine-tuned with data generated by user interactions.
This means they are live, public-facing software products with infinitely malleable (and thus unpredictable) output parameters. This poses a fundamental threat to user security and safety as it’s essentially impossible for computer scientists to predetermine risks associated with future use.
As NIST wrote in its AI RMF document:
“Measuring risk at an earlier stage in the AI lifecycle may yield different results than measuring risk at a later stage; some risks may be latent at a given point in time and may increase as AI systems adapt and evolve.”
Perhaps the solution to this problem is, as many researchers have suggested, the establishment of a non-profit, third-party, oversight committee for the purpose of assessing both foundational model safety and snapshot assessments of ongoing trustworthiness and reliability.
We’d like to know what you think. If AGI can be achieved by rushing models into public use, without conducting long term studies, does the potential reward outweigh the risk?
Would you feel differently if you knew for sure that AGI wouldn’t happen without you or your children’s lifetimes? What about if you knew for sure that AGI would happen in exactly 10 years?
Email editor@agiethicsnews.com with your comments.
Ethical concerns related to this article:
- What ethical principles should guide the prioritization of safety over speed in AGI research and deployment?
- How can transparent oversight mechanisms be implemented to monitor AGI’s development and prevent misuse?
- How should responsibility and liability be attributed for all AGI actions across its lifecycle, from development to disassembly?
A veteran AI journalist, Tristan was the creator and managing editor of The Next Web’s “Neural” imprint. He is one of the first and most prolific reporters to cover artificial intelligence as a full-time beat.