20th August 2026
I’ve been reading some really interesting articles about how the big AI labs are using philosophers and philosophy to shape their ethical stances. But ‘interesting’ is about as good as it gets: I just wish they actually made some difference. With AI escaping its sandboxes and the seeming ambivalence of the men who run these companies toward the fate of ‘everyone else’, this edition’s piece explores why ‘alignment’ is one of the most important concepts of our time and one that will shape society for good and for bad.
First, OpenAI claimed that one of its models (actually a combination of its current top-end model and an unreleased version) escaped the supposedly safe confines of its sandbox (an isolated technical environment), connected itself to the internet and hacked its way into a number of large companies, including Hugging Face, a large repository of code and models. Then, just a few days later, Anthropic described a similar scenario for one of its models. And, because problems always come in threes (and presumably not wanting to be left out), Meta reported the same.
The headlines were, as you might have guessed, sensationalist: “AI goes rogue!”, “AI hacks its way into company!” etc, etc. So, before anyone panics, there are a few mitigating circumstances: the models were specifically tasked with trying to break cybersecurity rules - none of them ‘decided’ to do this on their own; each of the models deliberately had their guardrails removed to maximise the effectiveness of the testing; and two of the models exploited human-error in the setting up of the sandboxes (the third found a vulnerability in a software package that eventually allowed it to get access to the internet). In normal circumstances (i.e. you and I using these models with all the guardrails in place), it is almost impossible for any of this to have happened. Which, in the face of the scary headlines, is quite comforting. But… these examples do demonstrate the potential power and danger of these models and highlight the unnerving fact that their safety (and therefore ours) is in the hands of a few companies and a handful of people. Who decides what those guardrails are? Who has the right to decide what those guardrails are?
The big labs, including at least OpenAI, Anthropic and Google, employ philosophers to help them answer these questions. A couple of recent articles, one in the Economist and the other in the Guardian, make fascinating reading. Different models can have different philosophical approaches - for example, Claude from Anthropic takes a deontological stance (popularised by Kant) which determines ‘good’ or ‘moral’ actions through the lens of rules and principles (e.g. those inherent to the universe, religious laws, or a set of personal or cultural values): it is the motives that determine ‘goodness’ rather than the consequences. This contrasts with Google’s Gemini and OpenAI’s ChatGPT, which take a consequentialist perspective. Here, it is the consequences of one’s conduct that are the basis for judgement about the rightness or wrongness of that conduct: does the action lead to the greatest good for the greatest number of people? (Go here for a deeper dive on all of this).
It is partly the role of the in-house philosophers to turn these philosophical theories into something the models can ingest and act upon. The latest version of Claude’s ‘soul.md’ document (as these treatises are called) was published at the start of this year and runs to 78 pages. I’m pleased to report that, in addition to Kant, it also includes the Universal Declaration of Human Rights.
What these documents, plus any other guardrails built on top of them, do is align the model’s outputs with what the labs’ philosophers and executive team believe is the best balance among safety, correctness, and style, all in the framework of that overriding philosophy. With the ubiquity of AI now and in the future, this alignment will be one of the most important factors shaping society.
Why is it so important? If this handful of philosophers and CEOs get to determine the way their AI models behave, then we are essentially aligning the views of a handful of (mostly) white men with the rest of society (through the voice of their AI models). That doesn’t sit well with me.
Let me give you another reason. There is a very good analogy in the Guardian article mentioned above, originally quoted by Google’s head of Responsible AI Strategy, but shown to be perversely inverted. Asked how commercialisation had changed Google’s approach to AI ethics, she argued that some risks must be managed by users, not the maker: “It’s like having a knife. A knife producer can’t guarantee how someone is going to use that knife. But they can put a cover on it so that it’s as safe as possible when it’s in a drawer. And make people really aware: this blade is sharp, do not use it in certain settings.” The article’s author concedes the basic point (cars kill over a million people a year, and we keep driving), but he argues the analogy breaks on distribution and coercion. A covered knife in a drawer is something you deliberately take out. LLMs, though, are everywhere: embedded in email, social media, search results and most of the apps on our ‘phones. He concludes: “It’s one thing to keep a knife in a drawer with a cover snapped over the blade. It’s quite another to blanket every surface of our homes, classrooms and workplaces with blades while insisting that no one who isn’t using knives for everything will be able to survive the future.” i.e. the knife-maker’s disclaimer only works if the knife is optional, and that is not the way that AI is heading. With AI everywhere, and difficult to switch off, alignment becomes even more critical.
Grok, the model from Elon Musk’s X.ai, is very much in the ‘knives for all’ camp, and is hardly concerned with putting any bothersome covers on them. Grok’s philosophical stance doesn’t really fit into a deontological or consequentialist framework. X.ai claims it to be ‘maximally truth seeking’, so maybe it aligns better with ‘moral deflationism’, where its design deliberately minimises the model’s moral role: ethics is treated as the user’s business, not something the system adjudicates: “Here’s an uncovered knife, go wild”. This is exactly the cause of the issues we saw with the hacking AIs - as soon as you remove the ethical and safety guardrails, you lose control of the model and how it behaves.
Do your values align with any of those philosophies? If they don’t now, will they be after you have been using LLMs for many, many years? Will societal culture be determined by the commercial success of one or more AI labs that are headed up by a few men with very questionable motives?
AI can, undoubtedly, do some profoundly good things. Just the other week it was announced that AI has designed viruses that can (at least in theory) infect bacteria and reproduce. This may sound worrying at first, but this approach could eventually prove useful in the fight against antibiotic resistance. And we know AI can help design new drugs - this is an alignment that only a few people would disagree with. So, if we did use AI to cure cancer, for example, would that be enough? Would that outweigh all the societal disadvantages that AI can bring?
But it’s obviously not an either/or question. What we each need to do, I think, is be informed and actively decide what we use AI for. We can’t sleepwalk into a world dominated by AI aligned to the values of a few people in Silicon Valley. Remember: we can switch it off.
Sargasso Sea - The Durutti Column ft Caoilfhionn Rose
The Durutti Column are a band from Manchester that has been around for a very long time. They were becoming famous at the same time as Joy Division, and signed to the same label, the inimitable Factory Records. Their first LP, ‘The Return of the Durutti Column’ came in a sandpaper sleeve so it could destroy neighbouring records in the rack. The ‘band’ is essentially one man, Vinny Reilly, plus a rotating cast, and it his guitar playing that defines the sound: a mixture of post-punk, ambient, folk, and chamber music. Despite a couple of strokes, Vinny, who is now 73, last month released the first Durutti Column LP in 15 years, called Renascent.
It’s a magical album, with a mix of instrumentals, Vinny’s spoken word overlays, and guest vocalist Caoilfhionn (‘Keelin’) Rose on a couple of the tracks. My favourite is Sargasso Sea, which is loosely based on the book Wide Sargasso Sea by Jean Rhys (a brilliant retelling of Jane Eyre from Bertha’s perspective). Do check out the LP and the rest of the Durutti Column back catalogue.
One final quick thing from me. In the last edition of TSCG, I mentioned the Prompt Action initiative that I founded. We now have our very own website, which you can see here. Many thanks to Charlie Burgess for the design and implementation.