4 ms·

I really dislike the current attitudes around ethics and AI. Tools should be 100% open about what they can and can't do. Building unethical software like deepfa

by lumost 3y ago

I really dislike the current attitudes around ethics and AI. Tools should be 100% open about what they can and can't do. Building unethical software like deepfakes for porn is also unethical. Quoting of vague "ethics" risks around AI, or neutering LLMs in the vein of making them "safer" seems dubious to me. Obviously, there are a huge number of problems which depend on safe AI - where safe often means something completely different from what is currently being talked about. e.g. does the AI reliably do what I expect it to do in the face of adversarial inputs? does it handle risky settings like deciding to refund customers? How can I ensure that the AI is making legal decisions and not secretly using redlining, or other illegal criteria to make its decision?

PawgerZ 3y ago
I am still taking in information and trying to figure out my stance for AI ethics, but this is moreso about the ethics of the business and its structure, not the ethics of the AI that they make.
- mschuster91 3y ago
  Both of these are intertwined, if not outright the same thing. A place that has no corporate understanding of ethics, no staff to take care about that, or even fires that staff (like Twitter did [1]), cannot build any ethical system. No matter if it's chatbots that shouldn't scoop up abuse by 4chan [2], reproduce long-debunked racially charged myths [3], or even something as simple and innocent-seeming as a soap dispenser that had never even been tested by Black people [4]. Many people think like this is "wokeness" and at best a waste of money, and yet time and time again there are incidents that show in sometimes very horrible ways just how bad biases and a lack of diversity and oversight at all stages - planning, development and post-launch support - can end. IMHO, ethics oversight and threat modelling based on ethics should be seen, similar to pentesting and QA in AI, software (and, as the example shows, also hardware) development, as an absolute requirement to do - because if you don't do it, 4chan will, and tear your public image apart in the process. [1] https://www.cnbc.com/2023/05/26/tech-companies-are-laying-off-their-ethics-and-safety-teams-.html https://www.cnbc.com/2023/05/26/tech-companies-are-laying-of... [2] https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch... [3] https://fortune.com/well/2023/10/20/chatgpt-google-bard-ai-chatbots-medical-racism-black-patients-health-care/ https://fortune.com/well/2023/10/20/chatgpt-google-bard-ai-c... [4] https://www.standard.co.uk/news/world/automatic-soap-dispenser-sparks-racism-outrage-after-footage-shows-it-doesn-t-work-for-darkskinned-people-a3615096.html https://www.standard.co.uk/news/world/automatic-soap-dispens...
  - lumost 3y ago
    I agree that oversight is needed - but I'm not certain people are the right approach to this oversight. Everything described above represents an undesirable software output for the given business. Why can't we make the LLM smart enough to know it's current operating context and build guardrails to catch the cases which slip through?
    - mschuster91 3y ago
      Thing is, you need people to take care in selecting training data. GIGO - garbage in, garbage out - especially applies in AI.
    - patcon 3y ago
      The proper guardrails worthy of being entrusted, they best come from an ecology of human minds with diverse perspectives and ideas, to grok the whole possibility space with minimal blind spots. The human machine is already tuned to do this navigating of the information landscape. The sense of curiosity itself lures us toward things that are just a little weird -- the goldilocksian zone between familiar and unfamiliar[1]. Our minds are designed to bridge gaps and seek paths into otherness, but only when the individual steps are small. We are built to seek, grow understanding and incorporate diverse thought :) [1]: https://nautil.us/curiosity-depends-on-what-you-already-know-235803/ https://nautil.us/curiosity-depends-on-what-you-already-know...
Racing0461 3y ago
[flagged]

by lumost 3y ago

PawgerZ 3y ago

I am still taking in information and trying to figure out my stance for AI ethics, but this is moreso about the ethics of the business and its structure, not the ethics of the AI that they make.

mschuster91 3y ago
Both of these are intertwined, if not outright the same thing. A place that has no corporate understanding of ethics, no staff to take care about that, or even fires that staff (like Twitter did [1]), cannot build any ethical system. No matter if it's chatbots that shouldn't scoop up abuse by 4chan [2], reproduce long-debunked racially charged myths [3], or even something as simple and innocent-seeming as a soap dispenser that had never even been tested by Black people [4]. Many people think like this is "wokeness" and at best a waste of money, and yet time and time again there are incidents that show in sometimes very horrible ways just how bad biases and a lack of diversity and oversight at all stages - planning, development and post-launch support - can end. IMHO, ethics oversight and threat modelling based on ethics should be seen, similar to pentesting and QA in AI, software (and, as the example shows, also hardware) development, as an absolute requirement to do - because if you don't do it, 4chan will, and tear your public image apart in the process. [1] https://www.cnbc.com/2023/05/26/tech-companies-are-laying-off-their-ethics-and-safety-teams-.html https://www.cnbc.com/2023/05/26/tech-companies-are-laying-of... [2] https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist https://www.theverge.com/2016/3/24/11297050/tay-microsoft-ch... [3] https://fortune.com/well/2023/10/20/chatgpt-google-bard-ai-chatbots-medical-racism-black-patients-health-care/ https://fortune.com/well/2023/10/20/chatgpt-google-bard-ai-c... [4] https://www.standard.co.uk/news/world/automatic-soap-dispenser-sparks-racism-outrage-after-footage-shows-it-doesn-t-work-for-darkskinned-people-a3615096.html https://www.standard.co.uk/news/world/automatic-soap-dispens...
- lumost 3y ago
  I agree that oversight is needed - but I'm not certain people are the right approach to this oversight. Everything described above represents an undesirable software output for the given business. Why can't we make the LLM smart enough to know it's current operating context and build guardrails to catch the cases which slip through?
  - mschuster91 3y ago
    Thing is, you need people to take care in selecting training data. GIGO - garbage in, garbage out - especially applies in AI.
  - patcon 3y ago
    The proper guardrails worthy of being entrusted, they best come from an ecology of human minds with diverse perspectives and ideas, to grok the whole possibility space with minimal blind spots. The human machine is already tuned to do this navigating of the information landscape. The sense of curiosity itself lures us toward things that are just a little weird -- the goldilocksian zone between familiar and unfamiliar[1]. Our minds are designed to bridge gaps and seek paths into otherness, but only when the individual steps are small. We are built to seek, grow understanding and incorporate diverse thought :) [1]: https://nautil.us/curiosity-depends-on-what-you-already-know-235803/ https://nautil.us/curiosity-depends-on-what-you-already-know...

Racing0461 3y ago

[flagged]