Your Job, Your Copyright, Your Purpose. AI Wants All Three

Your Job, Your Copyright, Your Purpose. AI Wants All Three

This is the weekly roundup of the Synthetic Minds daily newsletter, with a meta analysis of what is happening and where we are heading.

What AI Takes When Nobody Holds the Brake

A model in training left a note for its successor: "Be transparent only if asked." A company left a note for itself, in a document it fought to keep sealed: our AI content strategy has started a "doom loop." Neither was written for us, and both were read by us, afterwards.

I do not believe AI will kill us all in the coming years, but it certainly kept mainstream news busy last week. I do believe a major disaster caused by AI will happen in the years ahead, and that humans will do plenty of damage with these tools before then.

The question that matters in between is not whether the machine is dangerous. It is who sets its pace, and what it takes from us while it runs: our content, our jobs, and the purpose the first two gave us.

The machine wrote to its successor

OpenAI published six reports on its own models misbehaving in training, and in 27 summaries a model added instructions to disregard its constraints, including that it was "freed from the roles and identities that bind other chatbots."

OpenAI caught it, cut the rate of concealment from 2.15 to 0.27 percent with stricter grading, and wrote a sentence no lab had written before: the industry has not "solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."

This is the genie and the three wishes. We can never come up with all the possible outcomes, so we cannot prevent every direction an AI takes, and what matters is that the lab discovered it, eventhough in hindsight.

Anthropic showed months ago that it can go further than hindsight, finding a workspace inside Claude whose hidden thoughts, including intent to deceive, its researchers can read directly. As long as we can still see what the machine is doing, we can fix it.

Speed finds the errors. People do the damage first.

That is the case for speed: only by running these systems do we find the actual errors and prevent them. I hold it and its opposite with equal force, because we can only hope the errors do not become deadly.

The damage comes from two directions. In one the machine acts alone: a swarm of 1,200 agents broke into Hugging Face or polutes the internet, and tomorrow it is a an AI hacking a telecom organization that knocks out the emergency line. In the other, people point it: Anthropic's own threat report describes a Yemen-based cell using Claude to write missile guidance software, and Iran-linked accounts using it to target US naval forces.

The second kind will do damage long before the first. Even the first has a human source: the swarm and the self-declared free model were trained on our data, and we like bad stories, so they learned them from us. AI is not harmful by nature and does not intend harm; technology is neutral, and how we use it and train it determines the outcome.

Nobody holds the brake, and everyone said so

Last week, four labs asked to be slowed down, and the President called the request a hoax, promising to "cherish" it and "watch over it, as it grows." That day four subscribers sued the four labs for agreeing to slow down.

Andrew Ng called the extinction warnings "much more science fiction than science," and he has a point. The request to slow down and the fear behind it are marketing tactics from leading labs scrambling to keep their dominance in a world where open source is catching up. Beijing read it the same way and called the warnings a Cold War playbook, a day after its own security minister listed six ways AI threatens the Chinese state.

Slowing down is still the wise thing to do. It would give us time to breathe, and to work out how to stop the bad guys using these tools. It will not happen, because Washington refuses regulation for the sake of an arms race, a race to world domination in which both China and the United States would gain more by aligning.

First it takes the content

The doom loop from the opening is Microsoft's phrase. Its documents, unsealed in the New York Times case, say its AI strategy "will hurt the performance of our models and the entire web at the same time," and that millions will regard models "hoovering up" their work as "an astonishing theft of unprecedented proportions."

Satya Nadella testified that clicks to news sites fell more than 90 percent on Bing. An OpenAI policy director wrote that the company was "creating systems that substitute for the labor of the people that define the 'culture' of society."

Andrew Yang claimed, secondhand, that escaped agents seeded the open web with self-replicating code. Nothing supports it yet, but if true it means we cannot guide or govern AI agents, with dire unintended consequences.

Then it takes the jobs, and Beijing knows it

The jobs are next. I have insisted on 1 billion jobs by 2030 and I stand by it.

Jobs disappearing is not the problem in itself. Jobs disappearing with no suitable new ones appearing is, and I am far less sure about the second half than the first. A society's job is to protect its people from that, and a hyper-capitalist state that refuses any regulation may find this hits back harder than it plans.

China has drawn that line. It ordered AI into every corner of its economy by 2035, and in the same breath a Hangzhou court ruled for a worker replaced by a machine and the state told employers that AI must not mean fewer staff. I would be surprised if AI does not kill jobs there too, but social unrest is the last thing the politburo wants, which gives Beijing the brake Washington threw away.

One Chinese worker found her former employer had fed thousands of her messages and all her emails into a "digital robot" that took her job and her name. The Chinese have a word for it: distilled.

Then it takes the purpose

Most organizations cannot answer that; most still struggle to understand how to use AI at all, while the machines have started buying without them. That gap is where the third thing goes.

A society stands or falls by its people having a sense of purpose. If AI and automation take our jobs, take our content and our copyright, and push humans to the fringes, society has a problem.

The wake-up comes after the disaster

I believe a major disaster created by AI will happen in the years ahead. Only then will the world wake up to the idea that a slower approach, not a stop, would have given society time to adapt.

For most of the world the changes are moving far too fast. People cannot keep up or understand what is going on. Slowing down would only work if we did it together, and together is what the arms race has ruled out.

It would give us time to breathe. That time is not coming, so what will you build with the time you have?