An AI news briefing is only as accurate as what it reads. A briefing built on named sources reads the same feeds, market data and official records on every run, links each story to the article it came from, and copies prices from a data provider instead of asking a model to remember them.
That is the strongest guard there is against AI news that invents things. Most worries about AI news are really worries about inputs: a model that answers from memory, or from whatever a search returned that morning, can sound confident about a story it never read. This guide explains the difference between the two approaches, then walks through every kind of source Avos reads, what each is good for, and an example section built on it. Full disclosure: Avos is our product, and we say where its sources stop.
Why do AI news summaries get things wrong?
Three causes account for most of the errors people see:
- Answering from memory. A language model's training data stops at a date. Asked about today with no article in front of it, a model fills the gap with what sounds likely, which is where invented quotes and wrong dates come from.
- Reading whatever the search returned. A web search tool helps, but the search decides at run time which pages to read: a different mix of outlets every morning, with the pages that rank well ahead of the outlets you trust.
- Writing numbers. Prices, rates and case counts are easy for a model to get subtly wrong, and hard for a reader to spot.
Better models make each of these rarer. Better inputs go after the cause.
How is a chatbot's web search different from a briefing built on named sources?
A chatbot with web search is a research assistant, and a very good one for follow-up questions. A briefing is a standing order: the sources are fixed in advance, per section, and the same pipeline runs every time.
| Chatbot with web search | Briefing on named sources | |
|---|---|---|
| Who picks the sources | The search, when it runs | You, in advance, per section |
| Same inputs every day | No, results vary from run to run | Yes, the same feeds and data until you change them |
| Prices and figures | Written by the model from what it found | Copied from the data provider into a table |
| Links | Usually cites some of the pages it read | Every news item links to its article |
| Stories you already got | Up to the model and its memory | Stories from a section's recent issues are left out |
| A day with no news | An answer is always written | The section says nothing new happened |
We made the longer argument in Can a chatbot be your daily news source? and compare the two products directly in Avos vs ChatGPT. If you are choosing an assistant for the news, read the best AI assistant for daily news.
What sources does Avos read?
Every Avos briefing is built from sections, and every news section has its own list of feeds. A feed is one source: a publication, a search, a channel, a data series. Here is every kind, then what each is good for.
| Kind of source | What it reads | Best for |
|---|---|---|
| Curated topics | Avos's own topic sources (World, Technology, Business, Science and more), with wide international coverage from across the web | The day's big stories on a broad beat |
| News search | A news search for any phrase, with its own language and region | Niche and local topics |
| RSS feeds | Any RSS or Atom feed, or one from the built-in directory of newspapers, broadcasters, trade titles, blogs and official feeds | Outlets you already trust |
| Web search and AI search | A live web search for a query, or a search you describe in plain words | Topics no feed covers |
| Market data | Stocks, ETFs, forex, commodities and US economic indicators from Alpha Vantage; crypto from CoinGecko; the US national debt from the Treasury; the market's odds of a Fed hike or cut at its next meeting | Watchlists and the economy |
| Company news | News mainly about one listed company or fund, by ticker, from Alpha Vantage | Following a company |
| X | An account's recent posts, or what people on X are posting about a topic | First reactions and primary statements |
| YouTube | A channel's latest videos, summarized from what is said in them | Channels you follow but don't have time to watch |
| Podcasts | The passages of podcast episodes where your topic came up | What experts said out loud |
| Research papers | New papers on a topic from arXiv and PubMed, with their abstracts | Science, medicine and AI research |
| Funding rounds | Startup and venture rounds with the amount, stage and investors | Deal flow in a sector |
| Official feeds | USGS earthquakes, GDACS disaster alerts, the National Hurricane Center, the FDA, US travel advisories, the Federal Reserve, the ECB, a company's SEC filings, CISA's list of exploited vulnerabilities | Primary sources |
What is each kind of source good for?
News: curated topics, news searches and RSS feeds
Curated topics cover broad beats: they weigh a wide field of the day's stories, so a big story reaches every section it belongs to. A news search covers anything narrower, in the language and region you set. RSS is the most direct route: the publisher's own feed, with nobody ranking it in between. Avos checks your RSS feeds every day and emails you if one stops working.
Example section: Lisbon property, a news search for "mercado imobiliário Lisboa" set to Portuguese and Portugal, plus the RSS feed of a property trade title, written in English as five headlines.
When no feed exists: web search and AI search
Some topics have no feed and no clean keyword. A web search runs a live search for your query every time the issue is built. An AI search takes a description instead ("new battery recycling plants announced or funded") and finds pages that match the meaning, not just the words. AI sections, which write from a prompt instead of a feed, can also search the web while they write.
Example section: Battery recycling, that AI search plus a web search for "lithium recycling", written as short briefs.
Markets and the economy
Market data is the clearest case for structured sources. A financial section is a table: each stock, ETF, crypto, currency pair or commodity with its price and change on the day, copied from Alpha Vantage or CoinGecko when your issue is built. No model writes those numbers. The same goes for the US economic indicators, which are read exactly as Avos's public Macroboard shows them. The data may be delayed: stock and ETF prices are as of the latest close. More in our guide to a personal markets briefing and on the market briefing page.
Example section: Watchlist, with NVIDIA, ASML, the S&P 500 through SPY, gold, EUR/USD and bitcoin, plus inflation, the 10-year Treasury yield and the market's odds of a Fed cut.
Social and spoken: X, YouTube and podcasts
Much of what matters first appears as a post, a video or a conversation. An X feed reads one account's recent posts, or what people are saying about a topic. A YouTube section summarizes what is said in each new video, not just its description. A podcast search finds the passages where your topic came up, so you get the quote, not the whole episode.
Example section: What builders are saying, an X search for "reactions to the EU AI Act guidance", two YouTube channels and a podcast search for "AI agents in production".
Research, deals and companies
Research papers come from arXiv and PubMed, each with its abstract, so a science section reports what a study found rather than a headline about it. Funding rounds arrive with the amount, stage and investors, for a sector or for a company or investor. Company news follows one listed company by its ticker.
Example section: Climate tech, funding rounds for "climate tech", research papers on "solid-state batteries" and company news for one listed battery maker.
Official and open sources
The feed directory includes official feeds such as USGS earthquakes, GDACS disaster alerts, the National Hurricane Center, the FDA, US travel advisories, the Federal Reserve and the ECB, and the AI editor can add a US company's SEC filings or CISA's list of actively exploited vulnerabilities. Avos's free public trackers are built the same way, from official and open data: WHO and ECDC for the Outbreak Tracker, GDACS, the USGS and NOAA for the Disaster Tracker, NOAA, Copernicus and the World Bank for the El Niño Tracker, EPPO, FAO and the World Bank for the Panama Disease Tracker, and the US Treasury for the national debt on the Macroboard.
Example section: Hazards near our suppliers, USGS earthquakes and GDACS alerts with a filter that keeps only Southeast Asia. For an email the moment one lands, use a news alert instead.
What happens to a story between the source and your inbox?
Each section runs the same steps on every issue:
- Fetch. Read the section's feeds, within its time window.
- Filter. AI checks each item against the section's filters and your instructions, ranks what is left, and prefers a mix of outlets among stories of equal weight.
- Read. Take in the full text of the articles it kept: the article itself, the video's transcript, the paper's abstract or the podcast's passages.
- Write. Produce headlines or briefs in the section's format and tone, each linked to its source.
- Deduplicate. A story that ran in an earlier section of the issue is dropped, and stories this section delivered in recent issues don't come back.
- Stay quiet when there is nothing. If nothing new happened in a section's window, the issue says so instead of padding it with old news.
Your instructions stay readable the whole time: each section's feeds, filters, format and tone sit in its settings, and the AI editor changes them when you ask. The template gallery shows finished issues built this way, such as the Daily Brief and Science & Health.
Does building on named sources stop AI hallucinations?
It removes the main causes, but no AI tool is infallible, and we would rather say so. A summary can lose nuance. A source can itself be wrong. Market data can lag. So every Avos issue ends with "For information only, not investment advice. AI generated.", every news item links to its source, and every issue ends with a 1 to 5 rating where "Something was wrong" lets you say what; Avos checks every such report. For anything you will act on, click through to the original.
FAQ
Does Avos make up news stories?
No. It writes from articles and data it fetched for your sections, links each news item to its source, and copies market figures straight from the data provider. Like any AI writing, a summary can still lose nuance or get a detail wrong, which is why every issue is labelled AI generated and you can report a mistake from the rating at the end of the issue.
Which news sources does Avos use?
The ones you choose, per section, from curated topics with wide international coverage, news searches and any RSS feed to X, YouTube, podcasts, research papers from arXiv and PubMed, funding rounds, and market data from Alpha Vantage and CoinGecko. The table above lists every kind.
Can I add my own sources to an AI news briefing?
Yes. Paste any RSS feed, search the feed directory, add an X account or a YouTube channel by handle, or ask the AI editor to add them for you. Each plan sets how many feeds a section can read; the pricing page has the details. Related guides: Turn your RSS feeds into a daily email briefing and Track your industry.
Is the market data in Avos real-time?
No. Prices come from Alpha Vantage and CoinGecko when your issue is built and may be delayed: stock and ETF prices are as of the latest close, and commodity prices update once a day. For an email when a price crosses your level, set a price alert.