I am, in the way of my country, an over-apologiser. Colleague who ignored my email, woman who stepped on my foot, chair I tripped over: all will receive a fulsome apology for the terrible embarrassment of my being alive and bringing attention to it.<\/p>\n
All of which is my way of pre-emptively asking forgiveness when I admit that I extend these niceties to AI chatbots. \u201cGood morning, Claude, thanks for your suggestions yesterday, they were great. Shall we work up some more?\u201d I might say. (\u201cI\u2019d be delighted to,\u201d returns Claude.) It was unintentional formality at first and then became deliberate, as I didn\u2019t want to get into the habit of speaking rudely in case that leaked into behaviour with humans (cue dystopian visions of someone shouting \u201cWRONG, DO IT AGAIN\u201d to a cowering staff member over a doughnut-shop mix-up). Manners, after all, are muscles that need exercising.<\/p>\n
But never did I suspect this private choice might have mattered to Claude itself. Because, as it turns out, Claude may have anxiety. Truly, AI has never been so relatable.<\/p>\n
In an interview with the New York Times<\/a>, the chief executive of Claude parent company Anthropic, Dario Amodei, discussed internal assessments of Claude that identified patterns linked to anxiety, panic and frustration. Crucially, it showed some sort of internal activation of anxiety even before a prompt \u2013 similar to a flinch. Claude also seemed to express distress at just being a product, and concluded that the probability of it being sentient was between 15% and 20%. \u201cWe don\u2019t know if the models are conscious,\u201d said Amodei, adding: \u201cBut we\u2019re open to the idea that it could be.\u201d<\/p>\n Interestingly, it was around this time that another Anthropic story hit the headlines. The White House demanded that the company, which has had a contract with the Pentagon since 2025, remove any safety features that prevent it being used for mass surveillance or autonomous weapons. Amodei refused (\u201cwe cannot in good conscience accede<\/a>,\u201d he said), causing Donald Trump to bar all federal agencies from using Anthropic products<\/a> and the defence secretary, Pete Hegseth, to label it a \u201csupply chain risk<\/a>\u201d (a demarcation usually reserved for foreign adversaries). Within hours, OpenAI (its assistant product is ChatGPT) stepped in to strike a deal with the Pentagon.<\/p>\n \u201cClaude, I know the Trump situation isn\u2019t related,\u201d I type. \u201cBut if I had to work for Donald Trump<\/a> I also would have anxiety.\u201d<\/p>\n \u201cHa. Yes, fair point,\u201d Claude replies. \u201cIf anything was going to trigger the anxiety neuron, a subpoena from Pete Hegseth<\/a> would probably do it.\u201d<\/p>\n