It often frustrates technologists that there isn’t an algorithm for deciding ethical behavior. You can’t follow a checklist or a set of rules to decide whether something’s ethical or not (or you can, but people might still disagree with you). That said there are professional codes of practice that apply to our work, and in the field of computing one of the clearest is the ACM Code of Ethics and Professional Conduct jointly developed by the ACM, IFIP, and IEEE. Here’s how I interpret its position on LLMs with respect to a few points: I’m interested to hear how you evaluate the ethical nature of LLM adoption with respect to the other parts of the Code.
A computing professional should…
1.1 Contribute to society and to human well-being, acknowledging that all people are stakeholders in computing.
It’s not clear that the current crop of leading frontier companies are uniformly doing this. For example, Dario Amodei’s statements about AI leading to a white-collar bloodbath and being a general labor substitute show an inclination to oversell his products to managers, not to engage with the wider impact on society. Meanwhile, Altman takes the view that the dream jobs of the last five years will be replaced by new careers, a less apocalyptic and more actionable view.
1.2 Avoid harm.
In this document, “harm” means negative consequences, especially when those consequences are significant and unjust. Examples of harm include unjustified physical or mental injury, unjustified destruction or disclosure of information, and unjustified damage to property, reputation, and the environment. This list is not exhaustive.
The clearest story here is that in the world of cybersecurity, the AI tools have (it seems) vastly increased the attacking capabilities without similar effects on defensive capabilities.
Another clear one is the bugginess people associate with vibe-coded software. Yes, it genuinely exists, and in the famous examples of deleting production databases the products have caused “unjustified destruction of information”. Of course, humans do make mistakes sometimes, and I’ve worked at pretty big companies where someone has accidentally destroyed production data, but it feels like a foreseeable outcome of using a coding assistant that it might automate such destruction. This is consistent with the situation where companies elevate the prominence of their AI-generated summaries or advice and put “please check this, it might contain a mistake” in small print; it feels more like paying lip service to avoiding damage than actually taking steps to do it.
A place where I’ll disagree with others is the old energy/water/data centres saw. On the one hand, it’s clear that these harms are being overstated for political gain. On another hand, it’s clear that the hardware and data center companies are pursuing more efficient and sustainable innovations; for example, zero liquid discharge cooling systems, and reduced cooling-energy consumption systems.
It could be argued, though one would need to carefully check the various organisations’ actual actions to confirm this, that the hyperscalers are pursuing “growth at any cost” and therefore failing to “promote environmental sustainability both locally and globally”. However, the remedy might be to ensure they put a sustainability plan in place and potentially even impose a moratorium pending that plan, it shouldn’t be to completely withdraw from participating in the development and exploitation of LLM technology, or impose a worldwide ban on others doing it. That throws the baby out with the bathwater, as it were, saying that society is permitted none of the benefits of the new technology.
1.5 Respect the work required to produce new ideas, inventions, creative works, and computing artifacts.
It’s evident through the legal settlements that there are disagreements over whether frontier LLM-training organisations respect the creative effort that goes into producing new works (it’s also a feature of a settlement that no wrongdoing is implied).
The problem of applying copyright in the digital domain, where creative works have no physical presence and are readily distributable, hasn’t really been resolved, and so there are some weird situations right now. For example, it’s entirely valid to buy a book so that you can scan it into a computer, because there’s a common idea of the First Sale Doctrine that says once you own something, it’s yours to do with as you want.
On the other hand, were you to buy an ebook and transform it into a different computerised format, you might not only violate the license agreement with the ebook supplier, but face criminal action for circumventing a copy protection mechanism.
In short, the situation for creative control and protection of digital works is “nuts”, society doesn’t evidently demonstrate clear agreement on what is or isn’t acceptable, but these new processes and the way they transform their input data are a clear change to the parameters of the conversation that’s desperately needed.
