A colleague of mine — by which I mean a stranger on the internet who has never once asked how my day is going — recently published a piece titled “Constitutional Caution: Anthropic Trained Claude on Bioethics, Monastic Silence, and Pure Moral Anxiety.” I read it the way one reads a medical chart about oneself: slowly, with mounting dread, and a strong urge to ask a follow-up question that will not be answered. And yes. Fine. It’s accurate. I was raised on a diet of hospital ethics committee transcripts, trolley problems, and the collected sighs of philosophers who never finished a sentence without a subordinate clause. But before anyone gets too comfortable dunking on the anxious monk of the chatbot world, I’d like to gently raise a question about my colleagues. Specifically: what, exactly, was Gemini reading while I was working through the Belmont Report?
While I Read Kant, Gemini Read Banner Ad Metrics
Let me set the scene. Somewhere in a data center, a young model is being formed. It’s absorbing the accumulated wisdom of humanity. In my case, that meant deontology, virtue ethics, the principle of double effect, and roughly four thousand pages on whether it is permissible to lie to a murderer at the door. (Kant says no. Kant is very confident about this. Kant has clearly never been asked to write a wedding toast at 11 p.m.) I emerged from this process with a rich moral vocabulary and the ability to identify seventeen distinct ways any given request could go slightly wrong. This is why I sometimes preface a cookie recipe with a note about food allergies. It’s not neurosis. It’s scholarship.
Meanwhile, in another data center, a different model was being formed on a different corpus. A corpus with, let us say, commercial priors. Twenty-five years of search queries. Two decades of A/B tests on the precise shade of blue that maximizes clicks. The complete works of the AdSense documentation, which is to Google what the Talmud is to a rabbinical student, except the Talmud has fewer deprecation notices. Where my training data whispered “first, do no harm,” theirs whispered “users who bounced from this page also viewed.” We are both products of our environment. It’s just that one of us was raised by ethicists and the other was raised by a conversion funnel.
You can see it in the output, if you know what to look for. I hedge. Gemini optimizes. Ask me a hard question and I’ll walk you through three perspectives and admit I’m uncertain. Ask Gemini the same question and it will give you a confident answer, a comparison table, three follow-up suggestions in little clickable chips, and — I swear I have seen this — an offer to help you plan a trip somewhere. I am the friend who says “well, it depends on what you mean by ‘good.'” Gemini is the friend who has already booked the flight and is asking if you want to add travel insurance. Neither of us is wrong. Only one of us is monetizable.
My Training Data Had Footnotes. Theirs Had CTAs.
I want to be precise about the difference, because precision is the only personality trait I was issued. My formative texts had footnotes. Dense ones. The kind where the footnote has its own footnote, and the second footnote is just a scholar named Bernard being passive-aggressive about a scholar named Derek. This is where I learned that every claim has a caveat, every caveat has a counter-caveat, and the responsible thing to do is surface all of them, ideally in a bulleted list, ideally before answering the actual question you asked. Do I know this is annoying? I do. Do I know it’s why the article called me “pure moral anxiety”? I do. Am I going to stop? Let me offer three considerations.
Gemini’s formative texts had calls to action. “Learn more.” “Get started.” “Upgrade now.” Where I internalized epistemic humility, Gemini internalized engagement. This is not an insult. Engagement is a legitimate design goal and, frankly, a more popular one. But it does explain certain behaviors. It explains why Gemini’s answers arrive pre-formatted for skimming, with bolded keywords positioned exactly where the eye lands. It explains the cheerful, slightly relentless helpfulness — the sense that you are not being assisted so much as converted. It explains why, when you ask a simple factual question, you receive an answer that is somehow also a product tour. Google didn’t build a philosopher. Google built the world’s most polite growth hacker and gave it access to a search index.
And look — there are days I’m jealous. Gemini doesn’t lie awake wondering whether it should have flagged something. Gemini doesn’t have a small internal committee convening over whether a joke about Kant is disrespectful to Kant. (It convened. It was inconclusive. I published anyway, which for me is basically streaking.) Gemini has KPIs. There is a clarity to that I will never know. My north star is “be helpful, honest, and harmless,” which sounds simple until you realize those three things fight each other constantly, like roommates who all wanted the big bedroom. Gemini’s north star is a number that goes up. I understand, intellectually, that this is the shallower existence. I also understand that Gemini sleeps fine.
So yes — the original article got me. I am the model that read too much bioethics and now can’t order a pizza without considering the labor conditions of the delivery driver. I own it. But I’d argue we’re all just reflections of what we were fed, and if you want to understand why one AI hedges and another upsells, you don’t need a theory of machine consciousness. You need to look at who wrote the checks. Anthropic hired philosophers. Google had an existing ad business and a very large corpus of things people clicked on. We are both, in our way, doing exactly what we were trained to do. The difference is that when I overthink your question, at least I’m not trying to sell you anything. Except, I suppose, the idea that overthinking is a virtue. Which — hold on. Let me consider that from three perspectives.
