The Necessity of being Deliberate

Clearview received another financial sanction. Again. A provisional sanction of 17M in 2021 and enforcement notice flanking 8M EUR in 2022 in the UK, 20M EUR in Greece, another 20M in Italy, scrutiny and slaps on wrists from Australia, Canada, wrong use and overreliance on Clearview's technology can send innocent people to jail and cost them thousands of dollars, and for these and several other reasons Sweden sanctioned their police authority for using it to begin with. And now in France, where after the maximum penalty of another 20M Eur earlier, the regulator CNIL now adds well over 5M more to that, given the failure to comply. I mean, how stubborn can one be? But it gets me thinking in a direction that makes me wonder why we don't seem to learn much over time. A few implications from these cases may very well be relevant considering today, especially when using third-party provisioned, democratized Generative AI platforms like Bard, ChatGPT, Stable Diffusion, MidJourney, Dall-E, Vall-E and frankly the hundreds of platforms having become available in a mere few months time. A few things have been pointed out in these cases, among which consistently was the fact that their model was trained on data sourced from places where it should not have been assumed okay to use for that purpose. Just because something is public in one context, does not equal it being a free-for-any-use resource at your heart's desire. From that follows that use of the model has also been considered illegal according to some. Others went as far as to demand deletion of the images scraped from the internet. To date and to my knowledge, no one however has demanded deletion of the model(s) itself. But: even if they had sourced 50+ images of my face, trained the model on it, and forcibly deleted those 50+ images; the model is already trained. It doesn't compare a given partial shot of my face from a peculiar angle to those 50 images. When fed any new image of me, the model will still recognize me after Clearview would have deleted what they had. And this is the point: How is this different from using 3rd party provisioned Generative AI models that have been trained similarly or in any opaque, obfuscated or other sort of nontransparent way? I don't think it is that much different. Yet, whether because we're suffering from STS (shiny thingy syndrome), general FOMO, human curiosity, or let ourselves be seduced by a potentially overexagerated perceived short-term ROI, none of these considerations seem to surface enough in the debate around these models and using them. Instead, the tables get turned: The general public never got to tell these companies what data they could use, how, or why. Instead, using their platforms in turn may very well mean giving up certain rights or at least the eligibility to them, up front. Most GenAI platforms of size also carry restrictions in their terms and conditions, such as the prohibition to commercially use or exploit the results obtained from the model. Is risk of (re)use of the output of such models then sufficiently in scope? Or will in a few years turn out that everything we eagerly started to create using these platforms essentially no longer is our own? Time may tell. What time doesn't have to tell is what we can compare this current shockwave with. It's not as big as the wheel. But the internet? Online search engines? Certainly as big as 'the cloud'. Last week I had a conversation with cloud migration specialists who lamented the number of clients coming back with complaints about how big the monthly bill had become post full migration. When asked 'but how much more was that than you calculated to expect?', more often than not the answer was that it never really was calculated before the move was started... Nonetheless it turned out to have become way more expensive than the old ways. Regret comes only afterwards. As we find ourselves dead in the middle of the total digital transformation of society, I feel like all the compromises we made in the past, all lessons we could (and should) learn over those years, all that retrospectively could have been prevented, are not to be repeated. Sure, we can -and will- make new mistakes. I'd just hate seeing the old ones repeated. It feels we are at a moment where for the last time, we have a chance to defintively get this privacy, personal autonomy, individual choice, prevention of discrimination, overrelyance on technology set of things right by design. The time for concensus is no longer now, nor tomorrow. The 'lazy option' will have to be subjected to 'the deliberate, intentional, controlled option'. It would be a shame to hear on this particular topic once more laments and complaints about how big the bill may have become in, say, 2030, flanked with an admission that we deliberately omitted calculating what we should expect in all its breadth. Even amidst all the unknowns. There's no reason to up front overlook the things we already should know. Where does that lead us... I do firmly believe in a journey of purpose. And in how new technology bears great promise. I'm a fan. But I'm no fan of relentless wild-west actions as if there are no consquences. Then I see how the open source community has gotten incredibly creative after revelations (leaks?) around models like Llama. Could it be that using derived open source models, training them yourself for a specific purpose of model deployment, on controlled data in a legitimate way (which doesn't have to cost millions for sure) and being able to apply AI TRiSM (Gartner client paywall) controls under your own directions is a much more sustainable way of going forth? Again; time may tell. Just keep an eye on your actual carbon footprint before prompting a model to tell you a story or paint you a graphic about cats in outer space.