# Study finds AI assistants falter on news: 45% have significant issues, sourcing is the weak spot

> AI assistants struggled on news: 45% had major issues; 81% showed at least one. Sourcing was the weak spot, and Gemini fared worst: 76% major issues, 72% sourcing.

- Type: AI news
- Page: https://completeaitraining.com/news/study-finds-ai-assistants-falter-on-news-45-have/
- Published: 2025-10-23
- Updated: 2025-10-23
- Author: Jeroen Erne
- Jobs: IT and development
- Reading time: 3 min

### AI Assistants Stumble On News: 45% Of Answers Had Significant Issues

AI assistants are struggling with news questions. In a cross-market study of 2,709 responses, **45% contained significant issues** and **81% had at least one issue**. The biggest weakness was **sourcing**\-missing citations, misattribution, and misleading references.

Performance varied by platform. **Google Gemini showed the most severe sourcing problems**, with **76% of responses containing significant issues** and **72% with sourcing issues**. Other assistants stayed at or below 37% for major issues overall and below 25% for sourcing issues.

#### What The Study Measured

The European Broadcasting Union (EBU) and BBC evaluated free/consumer versions of major AI assistants answering news questions. Tests covered ChatGPT, Copilot, Gemini, and Perplexity across 14 languages from 22 public-service media organizations in 18 countries.

Responses were generated between May 24 and June 10 using a shared set of 30 core questions, plus optional local questions. Many participating organizations temporarily lifted technical blocks so assistants could access their content during the test period.

#### Key Findings

- **45% significant issues; 81% had some issue.** Problems were consistent across languages and markets.
- **Sourcing was the top failure mode (31% at a significant level).** Citations were missing, misattributed, or pointed to off-topic pages.
- **Platform spread:** Gemini had the highest rate of significant issues (76%) and sourcing problems (72%). Others were notably lower but still showed gaps.
- **Accuracy slips:** Some answers were outdated or incorrect. One highlighted example: mischaracterizing changes to laws on disposable vapes.

#### Why This Matters For General, IT, And Development Roles

People trust concise AI summaries, especially under time pressure. If those summaries cite the wrong source-or no source-bad information spreads fast and accountability gets murky.

For teams that build with AI, this is a quality and risk problem. For publishers and comms teams, it's a brand problem: your reporting can be misrepresented in answers that appear authoritative at first glance.

#### Practical Steps To Reduce Risk

- **Set a "citations-first" rule.** Treat any unsourced claim as unverified. Require links to primary or high-authority sources with dates.
- **Use allowlists and source tiers.** Prefer official documents, government pages, and recognized outlets. Down-rank blogs and aggregation sites without clear authorship.
- **Force structure in prompts.** Ask assistants to provide: 1) a short answer, 2) bulletproof citations (with direct URLs), 3) timestamps, and 4) a confidence note.
- **Add retrieval and checks.** Pair the model with search or a news API. Pull the latest sources, then have the model summarize and cross-verify before responding.
- **Block "fresh news" autopublish.** For events in the last 24-48 hours, require human review or display a visible freshness warning.
- **Build an evaluation harness.** Create a test suite of recurring news prompts in multiple languages. Track citation validity, link accuracy, and factual precision over time.
- **Log and audit.** Store prompts, outputs, and clicked citations. Spot patterns like repeated misattribution or dead links.
- **For publishers:** Monitor how assistants reference your content, ensure structured data is clean, and consider guidance pages for AI crawlers.

#### Methodology Notes That Affect Interpretation

- **Consumer models only:** The study looked at free/consumer versions to reflect typical use.
- **Temporary access changes:** Some outlets removed blocks so assistants could reach their content during the tests.
- **Time-bounded run:** Responses were captured in a specific window, which matters for fast-moving stories.

#### What's Next

The EBU and BBC released a [News Integrity in AI Assistants Toolkit](https://www.ebu.ch/publications/news/integrity-in-ai-assistant?ref=completeaitraining.com) with guidance for tech companies, media, and researchers. Reuters also reports concerns that growing reliance on assistants for news could erode public trust ([Reuters](https://www.reuters.com/technology/artificial-intelligence/ai-assistants-misrepresent-news-raising-trust-fears-ebu-bbc-study-finds-2025-xx-xx/?ref=completeaitraining.com)).

As EBU Media Director Jean Philip De Tender put it: "When people don't know what to trust, they end up trusting nothing at all, and that can deter democratic participation."

#### Bottom Line

Use AI assistants for speed, not as a single source of truth. Bake verification into your prompts, pipelines, and publishing rules. If the answer doesn't show its receipts, it's not ready for production-or your reputation.

#### Level Up Your Team

If your team needs repeatable prompts and review workflows that produce verifiable answers, explore focused training on prompts and evaluations:

- [Prompt engineering resources](https://completeaitraining.com/tag/prompt-engineering/)
- [Latest AI courses](https://completeaitraining.com/latest-ai-courses/)

---

This is the Markdown form of https://completeaitraining.com/news/study-finds-ai-assistants-falter-on-news-45-have/. The whole site is mapped at https://completeaitraining.com/llms.txt.
