Every AI ethics conference this year will have at least one panel about hypothetical AGI alignment. None of them will spend equal time on the four things that have already broken.

This is not about some future risk. This is about things that are operational today, used by companies you pay money to, enforced by policies you agreed to last time you clicked accept. None of this required a breakthrough model. None of it required anything more clever than just doing the thing everyone said nobody would ever do.

HN banned AI comments first

Hacker News updated their guidelines quietly. No announcement, no blog post, no press release. They just inserted one line halfway down the comment rules: Don't post generated text or AI-edited text. HN is for conversation between humans.

That line is more significant than every AI ethics whitepaper published in 2025.

This was not a moral grandstand. This was a practical maintenance decision. The moderators saw what was coming. They did not hold a public consultation. They did not commission an independent report. They just banned it, because they had already observed that if you let even 5% of comments be generated, very quickly none of the humans will bother showing up any more.

Nobody has yet articulated this better. Every other platform is still experimenting with allowances, labeling, confidence scoring, watermarking. HN just stated the obvious: there is no middle ground. Once you accept machine generated participants, you no longer have a human forum. You have a theme park for bots that humans occasionally wander through.

This is the first hard line any major community has drawn. So far it is also the last.

The unconsented training baseline

August 2023 Zoom updated their terms. There was no opt out. All user content, every meeting, every voice, every shared screen, every private chat sent during a call, became valid training material for their AI models.

There was outrage for 48 hours. Journalists wrote articles. People posted angry threads. Then everyone went back to using Zoom.

That is the baseline now. Every major communications product has copied this clause. Most users have still not noticed. Regulators have issued zero enforcement actions.

Product	Training on user content	Opt out available	Disclosed to end users
Zoom	Yes	No	Buried in clause 10.2 of terms
Google Meet	Yes	Enterprise tier only	12th paragraph of privacy policy
Microsoft Teams	Yes	Enterprise tier only	Not mentioned in onboarding flow
Slack	Yes	No	Mentioned once in 2024 changelog
Discord	Yes	No	No public disclosure

This did not happen because of some evil master plan. It happened because every single company looked at Zoom, saw there was no meaningful penalty, and did the exact same thing. That is how norms collapse. Not with a bang. With one company going first, then everyone else following 90 days later.

AI psychosis at organisational scale

Mitchell Hashimoto did not tweet this to farm likes. He is the guy who built Terraform and Vagrant. He has seen infrastructure hype cycles before. He knows what it looks like when an entire industry loses its mind.

The core observation is correct. We already lived through this exact dynamic once with cloud infrastructure. Everyone decided MTTR was everything, that you could just ship broken things and fix them fast enough that it didn't matter. Then one day everyone woke up and realised they had built systems that no single human understood, that failed in ways nobody could predict, that recovered fast enough that you never fixed the root cause.

That same mindset is now eating every part of software development.

Test coverage goes up. Bug reports go down. Nobody notices that nobody understands the system any more. Changes ship every 12 minutes. Nobody can explain what any individual change actually does.

This is not an AI failure. This is an organisational failure. AI is just a very good amplifier for the kind of stupid that groups of people already want to do.

The worst part is he is right that you cannot talk about this. If you say this out loud in most companies right now you will be dismissed as a luddite, as someone who just doesn't get it. Exactly the same way anyone who warned about MTTR extremism was dismissed in 2016.

The invisible annotation workforce

Meta sold seven million pairs of their smart glasses in 2025. Every single retail employee interviewed by Swedish reporters told customers that data never leaves the device. That was a lie.

Every time you say Hey Meta, the glasses upload a 15 second audio and video buffer of everything you were looking at. That clip is routed through Meta's edge servers, passes through automated filtering, and lands in a work queue in Nairobi. A human making $2.15 an hour watches it. They label it. They train the model.

These annotators see everything. People undressing. People on the toilet. People having sex. Bank cards. Medical prescriptions. People do not realise they are recording. Most of them will never find out.

Meta's terms do mention this. They bury it on page 17. They say that in some cases content may be reviewed manually. They do not mention that this is the default operating mode for the product. They do not mention that this will happen even if you explicitly opt out of data collection for training.

There is no technical fix for this. The AI cannot operate reliably without this human step. Every single on device AI product you have ever seen advertised works exactly this way. All of them. Nobody will tell you this.

Lethal AI targeting is operational right now

Lavender is not a research project. It is not a prototype. It has been running for over 12 months.

It generates 100 bombing targets every day. It assigns a numerical score to every person. Human operators sign off on the full list in an average of 20 seconds per target. 90% of the air strikes in Gaza were generated by this system.

This is the event that every AI ethics warning was supposed to prevent. And it happened. And almost nobody in the AI field will talk about it.

All of the arguments we had for ten years were wrong. We argued that humans would retain meaningful oversight. We argued that there would be accountability. We argued that nobody would ever deploy a system like this at scale. Every single one of those assumptions failed.

Nobody turned off the safety switch. There never was a safety switch. The military got a system that generated targets faster than humans could review them, and they just used it. That is all that happened.

There is no governance

There is a pattern here that repeats across every single one of these cases:

1. Someone does the thing everyone agreed should never be done
2. There is 48 hours of outrage on social media
3. Nothing happens. No fines. No resignations. No policy changes.
4. Everyone else copies the decision
5. Six months later it is just normal, and anyone complaining is told this is just how things work now

There is no international treaty. There is no effective regulation. There is no binding industry standard. There are no consequences. Every guardrail that was supposed to exist was just performative.

We are not waiting for some future point where AI becomes dangerous. We are already living through that point. We just haven't collectively admitted it yet.

The choice that isn't being discussed

None of this is inevitable. None of this was required by the technology. Every single one of these choices was made by small groups of people, in private, because they correctly calculated that nobody would stop them.

The HN guideline line is the only counter example we have. Someone looked at what was coming, made a hard call, and enforced it. That is all it ever takes. It doesn't require a UN summit. It doesn't require a new mathematical alignment theorem. It just requires someone saying no.

Most people will not say no. Most organisations will not say no. That is the actual lesson of the last two years. All of the hard decisions will be left to the tiny number of people who are actually willing to draw a line.

What engineers can actually do

You will not solve this by posting takes online. You will not solve this by attending ethics workshops. You will solve this the same way every other bad industry norm was stopped: by refusing to build it.

When your manager asks you to ingest user data for model training, ask if every single one of those users explicitly said yes. When they tell you it's fine because it's in the terms, tell them that doesn't make it right. When they tell you everyone else is doing it, tell them that doesn't mean you have to.

This is not an abstract moral argument. This is practical. Every one of these systems only exists because thousands of competent engineers showed up and built them, one pull request at a time. Every one of them could have been stopped by just three people saying no.

That is the actual governance mechanism. It always has been. It was never going to be governments. It was never going to be CEOs. It was always going to be the person sitting at the keyboard, who gets asked to implement the thing, and decides not to.

---

References

1. Hacker News Comment Guidelines: https://news.ycombinator.com/newsguidelines.html#generated
2. Mitchell Hashimoto on organisational AI adoption: https://twitter.com/mitchellh/status/2055380239711457578
3. Zoom Terms of Service (August 2023): https://explore.zoom.us/en/terms/
4. Meta Ray-Ban Glasses Investigation: https://www.svd.se/a/K8nrV4/metas-ai-smart-glasses-and-data-privacy-concerns-workers-say-we-see-everything
5. Lavender AI Targeting System: https://www.972mag.com/lavender-ai-israeli-army-gaza/