Building an AI Risk Radar is Easier than Scrolling News

I spent two days testing AI-assisted signal monitoring for slowing down AI development. It started when Anthropic CEO Dario Amodei published a 3,800-word article calling for a global slowdown in AI development, followed by statements from OpenAI and xAI leaders. The next day, AI concept stocks fell, and SoftBank's stock price plunged 11.54%. These kinds of news items move fast from release to market reaction, and just scrolling through news feeds makes it easy to be half a step behind.

My goal was to break similar news into two layers: who is saying it, and whether the market reacted. I compared two approaches: manual searching versus rule matching plus AI summarization.

First, here's a manual version that doesn't rely on code. Open a spreadsheet tool, click New Worksheet, create 6 columns: Date, Title, Source, Keywords, Sentiment, Action. In the Keywords column, fill in: slowdown, regulation, safety, plunge, open source, data center. Open your browser, type "AI slowdown regulation stock price" in the search box, click Within 24 hours, and you'll see a list of news sorted by time. Copy the titles into the sheet, manually mark out CEO calls, policy actions, and stock reactions, and fill in Sentiment as Concern, Neutral, or Panic. If two strong signals appear on the same day, write "Watch the market" in the Action column, then save.

In my test, manually checking 20 items took about 40 minutes, with 3 misrecorded entries. Clickbait headlines often write "AI concept stocks fall" as "AI crash," making sentiment judgment prone to drifting.

Rule matching is simple: first tag each news item, then let AI merge content with the same tags. I used Hacker News, a tech community I've been trying recently, along with Claude, an AI chat tool I've been using for about a week. Open Claude, click New Chat, and input You are an AI news classifier. Output only according to tags, no explanations. Paste 5 titles at a time, press Enter, and expect one line of results per title. Ask it to output in a fixed format with fields: Tag, Risk, One-sentence summary. Tags should be Regulation, Safety, Market, Tech, Other; Risk should be High, Medium, Low; leave the one-sentence summary blank. Paste the results back into the sheet. Filter in the sheet: if the title has both "CEO" and "slowdown," mark it as a strong signal; if it has "stock price" and "plunge," mark it as market reaction.

This approach ran for two days, processing 15 items daily, taking 25 minutes. It hit 6 strong signals with 2 false positives. Most false positives came from headlines like "Company says it won't slow down AI," where rules easily miss the word "won't."

Approach	Time Cost	Hits	False Positives
Manual Search	40 mins/20 items	8 items	3 items
Rules + AI	25 mins/15 items	6 items	2 items

Regarding pitfalls, the easiest mistake is treating AI summaries as facts. Initially, I asked the model to judge bullish/bearish trends, and it was very confident. Later, comparing with the original news, I found it interpreted "US urges G20 members not to intervene in AI regulation" as "Global regulation loosening." The direction was right, but it missed the qualifier "US stance." The solution is to only let it tag and extract, not draw conclusions.

Another pitfall is keywords being too broad. I added AI, model, stock price, and 19 out of 20 items matched, which meant no filtering at all. Narrowing it down to slowdown, regulation, safety, plunge, open source, data center reduced the noise.

Next steps could include trying event chains, arranging CEO articles, policy statements, and stock changes chronologically. Further ahead, connect rule matching to email archives, sending only one summary daily. Once the workflow runs smoothly, consider connecting to market data or RSS web subscriptions.

Discussions about slowing down AI won't stay just talk. Safety standards will likely start with coordination among top companies, then spill over into policy and valuation. For ordinary observers, first break signals into three layers: who is speaking, what they said, and whether asset prices reacted.