How to stress-test a scary AI prediction with Claude

Turn on web search

Without live search, Claude answers from memory and can produce a source that sounds right and does not exist. With it on, every claim comes back with a link you can open yourself, which is the entire point of the exercise.

1. 🔎 Switch web search on. Open a new chat at claude.ai, click the settings slider in the message box, and turn on Web search. Anthropic's walkthrough is here. It is available on every plan, the free one included (Anthropic announcement).
2. 🧠 Add Research if you have a paid plan. Research runs many searches and compiles a sourced answer instead of a single pass. Click the + button at the bottom left of the chat box and pick Research. It is on Pro, Max, Team and Enterprise (how Research works). On the free plan, web search alone still runs this whole method.
3. 📄 Get the transcript. On Instagram or TikTok, open the caption and transcript, or screenshot the video and let Claude read the frames. Type it out if you have to. The stress test needs the claims in text.

The stress test, in 4 moves

1. 🗂️ Extract every claim. Doom videos work because the claims blur together into one feeling. Splitting them into separate numbered lines breaks the spell before you evaluate anything, because you can see how many of the lines are actually forecasts.
2. 🔗 Trace each one to its origin. Every number came from somewhere. Claude finds the earliest source, names who said it and when, and links it. This is where most scary stats quietly turn out to be a quote rather than a measurement.
3. ⚖️ Argue both sides properly. You ask for the strongest honest case for the prediction, then the strongest evidence-based case against it. One-sided framing is what makes a story feel inevitable, and this is the step that removes it.
4. 📋 Compress it into a brief. What actually holds up, what is speculation in a confident voice, and what you personally control in the next year. Two minutes to reread, instead of a night lost to a caption.

The full stress-test prompt

This is the one I run. Paste it into a new chat, then paste the transcript underneath when Claude asks for it.

You are my claim auditor. I am going to paste the transcript of a viral video that makes a scary prediction, and I want you to stress-test it rather than agree with it. Use web search for everything and give me a real, clickable link for every source you cite.

Work in four labelled passes.

PASS 1, EXTRACT. List every distinct factual claim in the transcript as its own numbered line, in the video's own wording, without softening it. Tag each line as one of three things: a measurable fact about the past or present, a forecast about the future, or a value judgement. Count how many of each you found and tell me the totals.

PASS 2, TRACE. For each claim, find where it originally came from. Give me the earliest source you can locate, who said it, the date, and a direct link. Where the claim is a number, say clearly whether it comes from published data or from one person's estimate, because those carry very different weight. If a claim has no traceable source, write "no source found" rather than filling the gap with something plausible.

PASS 3, BOTH SIDES. Write the strongest honest case FOR the prediction in one paragraph using the best real evidence that exists, with links. Then write the strongest evidence-based case AGAINST it in a second paragraph, with links, including any data that contradicts it directly. Finish by naming the single assumption the whole chain rests on, the one that collapses everything else if it is wrong.

PASS 4, BRIEF. Give me a one-page summary in three sections: what is genuinely well supported, what is speculation delivered in the tone of fact, and what I actually control over the next 12 months given what survived. Keep the last section concrete and specific to my situation, not general life advice. Under 400 words.

Ask me for the transcript now, and ask me one question about my situation so pass 4 is useful.

Run it move by move

Same method, split into four prompts. Use these when the video is long or when you want to read each pass before the next one runs. Keep them in the same chat so Claude holds the context.

Here is the transcript of a video I saw. Pull out every distinct factual claim and list them as numbered lines using the video's own wording, without softening any of them. Then tag each line as a past or present fact, a forecast, or an opinion, and give me the totals for each tag. Do not evaluate whether they are true yet, I want a clean inventory of what is being asserted first.

TRANSCRIPT:
[paste the transcript here]

Now search the web and trace each claim on that list back to where it originally came from. For each one, give me the earliest source you can find, who said it, the date, and a direct link I can open. Where the claim is a number, tell me whether it comes from published data or from a single person's estimate or prediction, and say which explicitly. If a claim has no traceable source, write "no source found" instead of offering something that sounds close.

Argue the strongest honest case FOR this prediction in one paragraph, using the best real evidence that exists, with links. Then argue the strongest evidence-based case against it in a second paragraph, also with links, including any data that contradicts it. Finish by naming the one assumption the entire chain depends on, and tell me bluntly which side currently has the better evidence rather than staying neutral to be diplomatic.

Turn everything above into a one-page brief I can reread in two minutes. Section one: what is genuinely well supported, in plain sentences, with the source next to each point. Section two: what is speculation being delivered in the tone of fact. Section three: what I can actually control over the next 12 months given what held up, written as concrete moves for [describe your role and situation in one line], not generic advice. Keep the whole brief under 400 words.

The doom reel, stress-tested

The video that started this claimed half of white-collar jobs vanish in five years, then walked that into mortgage defaults, then a housing collapse three times the size of 2008, then money printing that eats federal tax revenue. Here is what the four passes turned up.

🔗 The headline number is a quote, not a measurement. It traces to May 2025, when Anthropic CEO Dario Amodei told Axios that AI could wipe out half of all entry-level white-collar jobs within five years (Forbes recap of the original prediction). Two things change the moment you see the source. It was about entry-level roles, not all white-collar work, and the reel dropped that qualifier. And it is a prediction from one leader, which is a different kind of thing from a statistic.

⚖️ There is real evidence on the other side. The Budget Lab at Yale tracked occupational and industry mix after ChatGPT's release and found no discernible broad disruption to the labour market so far (Evaluating the impact of AI on the labor market). Anthropic also publishes measured data on which tasks people actually bring to Claude, rather than forecasts (the Anthropic Economic Index). Neither says nothing is happening. Both say the measured picture is slower and messier than the reel's version.

🔎 The economics links each need their own check. "$14 trillion of mortgages" is exactly the kind of number worth opening for yourself. The primary source is the New York Fed's Household Debt and Credit Report, published every quarter, where the mortgage balance sits on its own line. Reading it takes a minute and you will never again accept that figure from a caption.

✅ What survived: entry-level hiring in some white-collar fields is genuinely worth watching, and senior leaders in AI companies do expect real displacement. What did not: the tidy chain, where every step triggers the next with certainty and no counter-force exists anywhere in the system.

4 signs you are reading a story, not a forecast

Save it so you never rewrite it

1. 📁 Put it in a Project. In the Claude sidebar, open Projects, create one called "claim audit", and paste the full stress-test prompt into the project instructions. Every chat you start inside it runs the method automatically, so you paste a transcript and nothing else.
2. ⚡ Or make it a skill. Ask Claude to write you a SKILL.md that packages the four passes, then in the app go to Customize, then Skills, then Add, then Upload a skill. Anthropic's steps are here. Skills work on every plan, Free included, and after that you type /claim-audit and paste.
3. 🌱 Run it on things you agree with too. This is the part people skip. A method you only point at claims you dislike is not a method, it is a preference. Run it on a prediction you want to be true and see how it holds.

Proof

Forbes, on Amodei's May 2025 Axios prediction and what changed since
The Budget Lab at Yale, evaluating the impact of AI on the labor market
The Anthropic Economic Index
New York Fed, Household Debt and Credit Report
Claude web search, available on all plans
Using Research on Claude