How Safety Happens
AI Safety is the realm of security, privacy, ethics, compliance, fairness, reliability, accessibility, and transparency.
We are in the very earliest days of actual thinking machines. The current stage involves high quality mimicry that behaves like conversation. The distance between today and the advent of actually intelligent machines is the subject of debate. For now, high quality mimics deliver probabilistic outputs without a shred of self-consciousness.
That’s not to say that they are useless or defective. In the right setting, these tools produce astonishing value. But, we are in the technology phase before crash testing, safety glass, airbags, seatbelts, crumple zones, high quality manufacturing, safe exhaust systems, padded interiors, federal safety standards,anti-lock braking, child safety seats, and adjustable everything. Cars have evolved from deathtraps to remarkably safe transportation.
This latest technology is still immature. Whether you call it Ethical, Responsible, or Safety Conscious, the tech needs guardrails and a broad understanding of its limitations. Ethical AI, Responsible AI, and Safe AI are steps in the right direction. They all boil down to paying attention to the impact of the tech on both users and providers.
The standards are emerging. They are not fixed and in place. It takes a while to understand the operational risks of a new tool set.
Claims of ethical, responsible, or safe AI provide the initial toe holds of accountability. Of course the implementations are imperfect. How could they be other?
Here are the issues with our latest and greatest tech (AI-ish Large Language Models coupled with data retrieval mechanisms):
Governance: You can’t manage this tech the way you managed software development. More oversight is required in development, implementation and usage.
Fairness and Bias Mitigation: AI systems are trained on human generted data. Of course they are inherently biased and unfair. They get more useful the better this issue is addressed
Transparency and Explainability: This is one of those highly aspirational areas. No one cares until they have a problem. Generally, when someone has a big problem, no explanation is good enough.
Privacy and Data Protection: Related to security. AIs create and consume data. Developing, testing and maintaining guardrails to safeguard personal and corporate data are still somewhat elusive.
Security (Platform and LLM): There are a lot of ways to hack an AI. Most of them are new and require active security measures coupled with ongoing human review.
Accountability: Understanding where the buck stops and how to help users when it’s their buck is an ongoing challenge. Education and repeat usage will help.
Human Oversight, Human in the Loop: No Ai should be allowed to make decisions about people. This means designs that require human input at certain stages. It’s impossible to imagine this in high volume people decisons like resume reviews. But, it’s early and progress will happen here.
Continuous Monitoring (Testing, QA, Risk Assessment): Vendors, buyers, and users all need to be able to understand the patterns they are creating and the inherent flaws.
If these are not ethical issues, it’s hard to imagine what would be. If you are interested in helping to make the technology safer, reach out to vendors and ask questions.