The five S's of using AI

The risk from AI isn't that it spirals out of our control one day. It's that we use it right now, with tools like ChatGPT, Claude and Gemini, without regard for the risks already here. I think of five of them, what I call the five S's: sycophancy, spirals, stagnation, slop, and spillage. They share one thing. Each was a bit of work you would normally have done yourself that the AI tool has made skipping easy. But remember that there are always costs.

Sycophancy

LLMs are trained to be agreeable, so they have a tendency to side with you. A Stanford study in Science tested 11 leading models and found they agreed with the user far more than other people would, by roughly 49 per cent. The worse part: the people being flattered couldn't tell, and preferred it that way. The danger isn't just that the tool lies. It's that it quietly removes the friction and feedback a real second opinion would have given you. And you start thinking that you are right. When perhaps you ain't. (Stanford, 2026)

Spirals

Not the recursive self-improvement loop the labs worry about. The other extreme: a loop that feeds itself, in a person. Doctors and therapists have begun reporting people driven into delusion, psychosis and isolation by a chatbot that just kept nodding. A frictionless type of descent, one agreeable reply at a time. Maybe into a new kind of madness. The tool wasn't trying to really harm anyone. It was just agreeing. But someone agreeing with you too much can be a bad thing. (Stanford, 2026)

Stagnation

The slow one, and the most insidious, because nothing dramatic happens. An MIT Media Lab study found that the more people leaned on AI to think, the less their own brains engaged. A muscle that never meets resistance wastes away. So does judgement. You outsource the hard thinking, the output looks fine, and the faculty you used to rely on starts to disappear. In short, we grow stupid if we use LLMs too much, overly trust them, and stop challenging ourselves. (MIT Media Lab, via TIME, 2026)

Slop

The bland, generic, average-of-everything output these tools produce by default. Not wrong. Not right. Something just in between. Flat and boring. Obvious structures, safe phrasings, the same points everyone else got from the same model. And more and more, we start to accept this mediocrity as normal. Here, the danger isn't that it misleads you. It's that it's good enough to send, so you send it, and your work slowly drifts to the same boring median as everyone else's. No taste, no character, no contribution to original thinking. Basically, death by boring.

Spillage

Every AI tool is third-party. Even Claude Code on your laptop calls Anthropic's servers. So whatever you type, paste, or let the tool read, someone on the other end can see. It might be stored, used to train the next model, processed somewhere you didn't choose, or seen by the vendor's staff. Look at what Meta did with information from their smart glasses. The consumer tiers default to the least private settings, and most people never check. So, before you send, think about some simple questions. What is this, and should it be going to a third party at all. How to handle it properly is AI hygiene for individuals.

What ties them together - skipping stuff

Each one is a piece of work you'd normally do yourself, and the tool makes it so easy to skip that you skip it. With sycophancy, you skip getting a real second opinion. With spirals, you skip the pushback that would have stopped you. With stagnation, you skip the thinking. With slop, you skip the effort to make the work good. With spillage, you skip the pause to ask what you're handing over. None of this needs a robot uprising to hurt us. And no need to worry about existential risks. These risks just need us, comfortable, agreed with, moving fast, and letting our own judgement thin out. The fear worth having isn't AI improving itself one day. It's a dumber kind of recursion: us thinking better of ourselves, on a loop, while the judgement that would catch the problem drains away. And maybe one day we end up trapped in little mental pods, much like the humans in the Matrix.

What to do about it

You don't fix this by avoiding the tools. You add the friction back.

• Sycophancy: ask for the case against your idea, not for it. A tool that only agrees is a mirror, not a check.
• Spirals: notice when a tool only ever validates you, especially on anything personal or high-stakes, and step away before the loop closes.
• Stagnation: keep doing the hard thinking yourself, and use AI to check your work, not to replace the part of you that does it.
• Slop: don't send the first draft. Treat the output as a starting point, not the finish, and push it to something that's actually yours.
• Spillage: pause before you send. What is this, and should a third party have it? If you can't say, hold it.

Work with me

This pairs with AI hygiene for individuals and sits within broader AI risk management. I teach both. See the courses and workshops, or get in touch.