Four AI Launches in One Month. Are We Losing Our Humanity?

On the 8th of this month, Meta launched Muse, a personal agent that opens a browser, fills in forms, sends email, books things and pays for them, and keeps working after you close the app. On the 15th, TypeSafe AI let the first users in to Jev, a model that doesn't write anything at all: you show it a situation, ask it questions that can only be answered yes or no, pick one, or on a scale, and it hands back the answers as numbers, in under half a second. And on the 21st, xAI shipped Grok 4.7, which it says was trained to work with its Grok Bot agents, the always-on bots that run on a computer in the cloud and keep going while you're offline. And on the 22nd, Anthropic released Claude Opus 5.5, which it says matches its top model on most work while running faster and cheaper, and which it's pitching at coding, computer use and the kind of knowledge work that used to be somebody's whole job.

Four launches in fifteen days. Line them up and there's a reading of it that's hard to shake. First we started talking to AI in plain language, like it was a colleague. Then we started quoting it in conversation ("ChatGPT said I should…", a sentence I hear more and more, and not only from clients). Then we handed it tasks to carry out unsupervised. And now, with Jev, we're letting it make the decisions. Each step feels small. The direction feels like it points somewhere we didn't agree to go.

Are we losing our humanity?

No. I think we're about to find out where it was hiding.

The mundane was never the human part

The strongest version of the worry is that every task handed to a machine is a piece of ourselves given away. I don't buy it, and the easiest way to see why is to look at what the machines are actually taking.

Muse's launch examples were booking travel, lowering a bill, filling in a form in a browser and turning a saved recipe reel into a grocery list. xAI's own list of jobs for a Grok Bot is long and unglamorous: managing an inbox, preparing for meetings, chasing expenses, sorting support tickets, the stuff that repeats every week. Nobody I've met describes that part of their day at a dinner party. It's the part that arrived uninvited and stayed because nobody had time to get rid of it.

Take the inbox, since that's the one most of us have. Clearing spam, sorting what's left by urgency, drafting the three-line acknowledgement that says "received, will revert by Thursday": an agent can do all of that now, reliably, and the person whose inbox it is loses nothing they'd miss. What they get back is the hour before lunch. Whether they spend it on something that matters is up to them, but at least the option exists again.

I run a version of this myself. An automated assistant reviews our clients' search and ads data overnight and leaves me a short list in the morning: decisions rather than a report, each with the evidence attached and a recommendation. I read it, decide, and move on. The reviewing used to eat a large piece of my week. The deciding takes a fraction of that, and the decisions are better, because I'm making them with a clear head instead of at the end of an afternoon of spreadsheets. The machine took the reading and left me the judgement.

Jev is the interesting one, and for the opposite reason

Jev looks like the top of the slippery slope, the point where the machine stops fetching and starts deciding. Simon Willison called it "a new shape of LLM", and the shape is genuinely new: no prose, only answers. Is this enquiry worth a callback? Which of these five queues? How urgent, on the scale I've described?

But look at what it needs from you before it can answer. It needs the question. It needs the scale. It needs the five queues to have names. It needs someone, somewhere, to have already decided what "urgent" means in this company, and to have written it down in a form that survives being read by something with no context and no ability to ask.

That's the thing most organisations have never done. I say this having sat through the interviews. Before we build any automation for a client, I get the rule written down: what a good outcome looks like, and who decides the exceptions. That document is regularly the slowest part of the whole project, slower than the build it precedes. The most common thing it turns up is that the rule doesn't exist yet. Two people have been applying it from memory, and when you put their versions side by side, they don't match. Nobody had noticed because nobody had ever asked either of them to type theirs out.

So the model that supposedly takes decisions away from people turns out to be the one that forces them to sit down and check whether they have a decision-making process at all. In most of the companies I meet, the honest answer is that they have a Thursday meeting and a person everyone pings. Jev can't use either of those. It can only use a rule, and to give it one, someone has to think harder about the business than they have in years.

That's about as human an activity as I can name.

The step that does worry me

I'll give the slippery-slope reading one thing, because it deserves it. The step that bothers me is the second one: "ChatGPT said I should". The advice is often fine. The problem is that the sentence hands over the thinking rather than the doing, in the one place where the thinking was the point. A person who lets an agent clear their inbox has lost nothing. A person who lets a chatbot decide whether to take the job, leave the partner, or fire the supplier has skipped the part that was theirs to do.

Opus 5.5 makes this step more tempting, not less. One of Anthropic's early testers is quoted saying "it writes the way I do", and I believe it; every generation gets easier to talk to and harder to argue with. That's wonderful when it's drafting the proposal I'd have written anyway. It's a problem when it's answering a question I should have sat with for a week. The better the model gets at sounding like me, the more carefully I need to notice which of us is doing the thinking.

Notice that this is the exact opposite of what Jev does. Jev refuses to think for you. It will only ever apply what you've already thought.

What I'd do with this month

Hand the fetching to the agents. The inbox, the forms, and whatever else repeats every week. Attach a named human to each one, because a bot working while you're offline still needs someone who answers for it, but let it work.

Then, before touching anything like Jev, do the slow thing. Get the three or four decisions the company makes fifty times a week onto a page: what the terms mean and who owns the exceptions. It'll take longer than expected. It'll turn up at least one disagreement that's been running quietly for years. And at the end of it the company will have something it didn't have before, whether or not it ever uses Jev.

Machines that fetch and machines that apply rules are taking the parts of work that were never the human part. What's left is deciding what the rules should be, and being answerable for them. That's the job. It always was. This month just made it harder to hide from.