OK, so everything has problems and opportunities. Cars were awesome, except, ya know, greenhouse gasses and dead pedestrians. See also fridges (CFCs) and who thought that cocaine and opium were good for kids? (the 1800's did, that's who).
Everything has risks, upsides and downsides. Lets have a look at some guiding principles and a few obvious and very public risks and downsides.
Just got here? Part 1 is here.
Principles
I'm starting to formulate some basic principles for my use of AI, and in systems I'm involved in designing. Most of them come from other people, so I can't claim them as my own.
The machine is never right, at best it's an advisor
A computer can never be held accountable, therefore a computer must never make a management decision
That was IBM in the 70's, apparently, and I think that goes 10x now. AI is a TOOL, nothing more. A human should always make a final, binding decision, especially if it affects the rights of another human.
Anecdotally (I need to find a proper reference for this), our Accident Compensation Corporation (ACC) has a good principle on it's AI-driven claims engine:
The AI can accept a claim, but a human must decline a claim.
For those not in NZ, ACC is our no-fault accident insurance system.
They process a lot of claims, usually coming via a GP or other medical practitioner, and they have a policy - I'm told - that while the claims processing system can accept a claim (good customer service, cost/impact of accepting an invalid claim isn't super high, and outright fraud can be prosecuted under current law), it can't decline a claim - that has to be reviewed and done by a human (impact of declining a valid claim could be life changing for someone).
We are applying this at Tend, too. Our AI tool is exactly that - a tool for the GPs.
They are legally responsible for the notes which go into the system, but the tool can help write notes that a patient would be able to read, in less time and possibly with more detail than they have time or ability to do.
More on this later, but GPs in NZ generally super smart, but also super time poor and under insane pressure. Having an assistant to write the patient notes - for them to review, edit and then accept - is a massive time and effort saver.
Risks and Downsides
I'm still undecided about the copyright and content aspects of all of this. For me, a lot of this comes down to which part of the process we are talking about: training vrs inference.
Training: taking a HUGE corpus of information and training - creating - a model from it.
Inference: taking that model, and "asking it questions". This doesn't change the model, it just runs input thru it and generates output
At one end, I have very little time for the likes of the NY Times - who are screaming for people to read their articles - complaining that an AI is being trained on their publicly available content, while building sites with patterns which are openly hostile to paying subscribers. Same for book authors, composers etc.
Ask any artist who their inspiration is, who's styles they admire, copy (sorry, are inspired by), adapt, twist, and do their own take on. Nothing is created in a vacuum. Everything is derivative.
I don't see training an AI model to be much different to a teenager listening to Hendrix, The Who and Zeplin and then picking up a guitar and writing some inspired, ripping riffs.
The generation - inference - stage is another thing altogether.
Figma's AI tool generating an almost direct clone of Apple Weather, or asking ChatGPT to "generate a short story about a girl with button eyes in the style of Neil Gaiman" and then claiming a Coraline derivative to be your own work is just dishonest. We mostly have copyright laws for this, even if they are clunky at best.
Right now we don't have any good tools or frameworks - legal or social - to handle both aspects of this, outside of suing someone or trying to prevent bots crawling your site, which is why its such a mess.
Some of this stems from the complexity of the models, which are very big, multi-dimensional black boxes. In most cases, we can't even pin down how the model produced the output - and they are absolutely not deterministic.
For me, playing around with this, I'm going with the principle that almost anything is fine for my own use, but not if I make it public.
If I'm using a general purpose tool (eg Claude or ChatGPT) to summarise a paid-for newsletter for my own use, thats fine. Publishing that output without permission absolutely is not. Publishing it and claiming it to be my own, even more so.
Some things are off limits tho. Making AI porn - especially CSAM - will always be something I'm not comfortable with.
Environmental
This is an interesting one, and is and will change very very rapidly.
- Google announced it's CO2 emissions went up by 48% from 2019 due to Gemini
- Microsoft's are up 30% since 2020. They run a lot of ChatGPT.
A lot of this is just down to bad long-term planning by the countries they work from.
If the base renewable supply is low, then anything is going to have high CO2 emissions. The same data centre in NZ would have substantially lower CO2 emissions than if it was in Australia, only because NZ is around 87% renewable (yay Think Big hydro projects from the 80s) vrs Australia at 40%.
Efficiency is the other obvious level to pull, too. As we've seen in data centres already with the introduction of low-power ARM server CPUs, the same workload can be done with less input energy and less heat output. AI is going to be no different - we are already seeing it in models like Claude where the 3.5 model is quicker (read: uses less processing) for better output than the previous ones, and this is accelerating quickly.
Hell, NVIDIA's Grace chip claims 2x the performance per input Watt of energy. And thats just one generation of change.
The training phase is still ramping up, but you train a model once - you use and infer from it many many times. Getting the inference side down to a minimal expenditure seams like a good start.
I think if things in hardware stayed static, then sure, this is a climate disaster. But they don't, not by a long shot. This stuff is 24 months old. Most design cycles for hardware are longer than that.
I'm going to reserve judgement on it.