The 2025 EBU-BBC study exposes a deep and persistent misalignment between the promise of AI as an information gateway and its current reliability.

News Distortion by AI Assistants — Findings from the 2025 EBU-BBC Report

by ChatGPT-4o

The 2025 international study “News Integrity in AI Assistants”, coordinated by the European Broadcasting Union (EBU) and led by the BBC, represents the most comprehensive evaluation to date of how AI assistants handle news content. Involving 22 public service media (PSM) organizations across 18 countries and 14 languages, the study paints a sobering picture of systemic shortcomings, with nearly half of AI-generated responses misrepresenting the news in significant ways. This essay explores the most surprising, controversial, and valuable findings of the report, analyzes the implications, and concludes with concrete recommendations for stakeholders.

Key Findings and Their Implications

1. Systemic Distortion: 45% of AI responses had significant issues

The most striking headline is that 45% of all tested AI responses exhibited at least one significant problem—ranging from factual inaccuracies and sourcing failures to inappropriate editorialization and missing context. When including minor issues, the figure rises to 81%. This establishes that the distortion is not due to isolated failures, but is systemic and language-agnostic.

🟡 Surprising: Errors were consistently found across all assistants, all languages, and all territories—even in relatively straightforward questions.

🔴 Controversial: Despite these flaws, younger audiences show high trust in AI assistants for news. According to a parallel BBC study, nearly half of UK adults under 35 completely trust AI-generated news summaries.

🟢 Valuable: The study quantifies and categorizes AI hallucinations, offering a taxonomy of errors and a toolkit to guide remediation.

2. Sourcing as the Biggest Problem: 31% had significant sourcing issues

The single most frequent type of failure was sourcing: assistants cited nonexistent URLs, misrepresented satirical content as factual, or failed to link to the content at all.

• Google’s Gemini was the worst offender, with sourcing issues in 72% of responses. Some cited Radio France’s comedy segment “Charline Explodes the Facts” as a credible source for claims about Elon Musk doing a Nazi salute.

• ChatGPT, Copilot, and Perplexity performed better but still had noticeable issues, including fabricated links and missing attributions.

🟡 Surprising: Some responses quoted news organizations as saying things they never did—posing real reputational harm.

🔴 Controversial: Misattributions could lead to lawsuits or loss of public trust, particularly if satire is misrepresented as fact.

🟢 Valuable: The report includes real-world examples that PSM organizations can use for internal training and legal preparations.

3. Factual Errors and Hallucinations: 20% of responses had major inaccuracies

Even for simple facts like “Who is the Pope?” multiple assistants failed, with some still referring to Pope Francis months after his death in April 2025. Assistants also:

• Reported outdated election results

• Claimed surrogacy is illegal in Czechia (it is not)

• Confused causal relationships in complex events (e.g., Trump’s tariffs and resignation connections)

🟡 Surprising: The assistants fabricated quotes or misattributed them to world leaders like Zelensky or Trudeau.

🔴 Controversial: These errors impact public understanding of law, health, politics, and geopolitics.

🟢 Valuable: Detailed side-by-side comparisons show how misinformation is structurally embedded even when prompts request reliable sources.

4. Overconfidence Despite Errors

One of the most insidious findings is that AI assistants deliver erroneous content in an authoritative tone, presenting assumptions and guesses with certainty. This includes:

• Gemini dismissing valid concerns as “misconceptions” and blaming the user

• ChatGPT confidently stating Trump “escalated” a trade war that was contextually misrepresented

🟡 Surprising: Refusal rates dropped to nearly 0% in 2025 from 3% in 2024, implying that assistants now attempt answers regardless of confidence.

🔴 Controversial: This “answer-first” design encourages hallucination and contributes to misinformation.

🟢 Valuable: The findings reflect recent research from OpenAI and NewsGuard showing that language models are optimized to sound confident—even when wrong.

5. Loss of Visibility and Traffic for News Publishers

With AI assistants increasingly replacing traditional search, news organizations are losing visibility. The Financial Times, for example, reported a 25–30% drop in search-driven traffic.

🔴 Controversial: The current ecosystem favors AI providers profiting from PSM content without delivering users back to the original source.

🟢 Valuable: The study calls for enforceable attribution standards and licensing frameworks.

Consequences of the Findings

1. Erosion of Public Trust: With 42% of adults saying they’d trust the original news source less if an AI summary was wrong, errors made by assistants are eroding institutional credibility—even when those institutions aren’t at fault.

2. Reputational and Legal Risk: News organizations are being misquoted or falsely linked to satire or politically inflammatory content.

3. Democratic Harm: If AI-generated misinformation becomes the norm for young users, this could degrade democratic participation, public knowledge, and civic engagement over time.

4. Market Imbalance: AI platforms are extracting value from public service journalism while diverting traffic and failing to share attribution, weakening the economic viability of responsible journalism.

Recommendations for Stakeholders

✅ For AI Developers

• Prioritize sourcing and accuracy: Implement rigorous source verification and show links to original articles, including timestamps and content confidence levels.

• Transparency reports: Regularly publish performance statistics by language and region.

• Embed disclaimers and uncertainty: Normalize the use of “We don’t know” where evidence is unclear.

✅ For News Publishers

• Use opt-out mechanisms: Consider leveraging copyright opt-outs where applicable to avoid unauthorized AI use.

• Push for content licenses: Only permit AI training and display under clear, enforceable contracts with attribution, linking, and compliance standards.

• Monitor and audit AI use: Build internal monitoring teams to detect misrepresentations and initiate corrective or legal action.

✅ For Policymakers and Regulators

• Enforce digital transparency laws: Use the EU Digital Services Act, Digital Markets Act, and equivalent national laws to enforce discoverability and attribution.

• Mandate AI labeling: Require AI systems to disclose when responses are AI-generated and include confidence intervals or sourcing transparency.

• Fund media literacy: Invest in campaigns—especially for youth—to teach critical AI and news evaluation skills.

✅ For Audiences

• Question AI outputs: Always look for sources and verify against original journalism.

• Favor direct subscriptions: Use official apps and sites of public broadcasters to avoid algorithmic distortion.

• Participate in digital literacy efforts: Engage in public discussions and demand transparency from tech providers.

Conclusion

The 2025 EBU-BBC study exposes a deep and persistent misalignment between the promise of AI as an information gateway and its current reliability. While improvements have occurred since the initial BBC study, the fact that 45% of news responses still contain significant errors underscores the urgency of intervention. Without reform, AI assistants risk becoming systemic sources of misinformation and disinformation—whether through hallucination, editorial drift, or subtle bias. To protect democracy, public trust, and journalism itself, stakeholders must act now to demand accuracy, enforce accountability, and design for transparency in AI-mediated news delivery.