# Some AI safety stories omitted how tests were staged

Markdown copy of https://realalpha.app/s/some-ai-safety-stories-omitted-how-tests-were-staged/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Ed Zitron](https://realalpha.app/people/ed-zitron/) · Ed Zitron is a writer and podcast host. · Hosts Better Offline and writes Where's Your Ed At, a prominent critic of AI industry hype.
- Host: Eric Newcomer · Newcomer
- Show: Newcomer Pod
- Episode: [Ed Zitron Unfiltered on OpenAI, Anthropic & Why the Whole Thing Is a Con](https://realalpha.app/v/wOBTi8dAb1Y/), Jun 8, 2026
- Moment: 50:49 to 54:05 of the episode (3 min), playing on https://realalpha.app/s/some-ai-safety-stories-omitted-how-tests-were-staged/; the episode from there: https://realalpha.app/v/wOBTi8dAb1Y/?t=3049

## Quote

> “they had actually prompted it, asked it what it would say, **preferentially chosen the responses**.”
> — [Ed Zitron](https://realalpha.app/people/ed-zitron/) at [51:00](https://realalpha.app/v/wOBTi8dAb1Y/?t=3060) · Newcomer Pod

## Takeaways

1. He says the TaskRabbit scenario was prompted.
2. He says researchers selected the model’s responses.
3. He disputes how stories about Claude blackmail were presented.

## About this segment

Ed Zitron says the TaskRabbit scenario used prompted, selected model responses, and disputes how Claude blackmail tests were framed. He argues that test setups matter when the media describes model risks.

From “Ed Zitron Unfiltered on OpenAI, Anthropic & Why the Whole Thing Is a Con”, Jun 8, 2026. [Watch the full episode](https://realalpha.app/v/wOBTi8dAb1Y/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[50:54](https://realalpha.app/v/wOBTi8dAb1Y/?t=3054) In the system card of GPT 3.5, I believe, there was a story that a task rabbit had been manipulated by the model autonomously. But when you click down into the meter study, it turns out that they had actually prompted it, asked it what it would say, preferentially chosen the responses. And then they intimated that they never did the experiment at all. They used that and allowed multiple different stories, I think 14 different stories, to talk about how ChatGPT 3.5 had convinced, maybe four actually, it was the 2023 era, had convinced the tasker and blackmailed them.

[51:27](https://realalpha.app/v/wOBTi8dAb1Y/?t=3087) Anthropic has multiple times done these stories saying Claude blackmailed someone. And what actually happened was they trained it to do something and it did it. It escaped the sandbox. Actually, it wasn't in the sandbox and it didn't escape it at all. So, yeah, actually I do. Don't you want them running safety tests? Like, the tests are meant to see if the models can be manipulated to do damage. Then if they're running safety tests, why do they keep doing weird manipulative things about them blackmailing when they're not actually...

[51:52](https://realalpha.app/v/wOBTi8dAb1Y/?t=3112) I mean, white paper, like, just most Americans, nobody is reading these white papers or people like us. Yeah, journalists are reading them. And then reporting them in the most preferential way possible, which then informs regular people. So yeah, actually, I do think my standards of- Most American- like Americans are largely skeptical of AI. Like they've failed, right? Don't you agree? Like the polling is negative.

[52:14](https://realalpha.app/v/wOBTi8dAb1Y/?t=3134) Americans- Here's the thing. Don't believe in AI. Most Americans don't run companies and most CEOs believe this because most CEOs are disconnected from labor. That's the problem. Venture capital has predominantly funded AI because of things like this. venture capital has failed to do the due diligence necessary because venture capital has uber brain venture capital thinks that burning money is good and i think it's perfectly reasonable to ask for perfection because that's how they're selling it they are selling it as autonomous intelligence this is the in your view a bunch of rich people who are delusion deluding themselves are lighting their money on fire like yes what's the problem like you don't like the rich people who are lighting their money on fire because they're covering our country in data centers ruining

[52:55](https://realalpha.app/v/wOBTi8dAb1Y/?t=3175) local environments gas belching turbines that are ruining black neighborhoods they are stealing from everyone they are threatening everyone constantly in the media and when people say they don't like this stuff someone in the media will condescendingly say oh you just don't get it you haven't used it you need to do this you're so stupid you moron why don't you like data centers yeah i i think that that's pretty bad you're you're the one you call people imbeciles morons Like, I'm not the one doing that.

[53:23](https://realalpha.app/v/wOBTi8dAb1Y/?t=3203) Like, I'm not saying, I wasn't talking about you, man. I was talking in general. But journalists are much more measured, I'd say. Go read the New York Times, mate. Go read how the Times talks to people when they don't like AI. Come on. Where? When? Casey Newton and Kevin Roose, mate. Come on. Look at them on Twitter. Oh, yeah. You're engaged. I actually don't thought. I'm not doing the horse. That is a personality.

[53:45](https://realalpha.app/v/wOBTi8dAb1Y/?t=3225) I'm genuinely not doing horse trading here. I'm just saying it happens. And that has a real effect on people. And I think the tech media needs more empathy for regular folks, but also a genuine connection with how this has been sold. This has been sold as autonomous. It is not.