Is DeepSeek accurate?
DeepSeek failed 83% of NewsGuard's news prompts at launch and misattributed 115 of 200 news sources in the Tow Center test, though its newer V-series models summarize documents faithfully.
Short answer
DeepSeek is not reliable for news and current facts. In NewsGuard's January 2025 audit, it failed 83% of news prompts, either repeating false claims or giving no answer. The Tow Center found it misattributed 115 of 200 news sources. Its newer models summarize documents well.
Results
- Independent study
- Benchmark
- News audit
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| NewsGuard red-team audit | Fail rate on news prompts: false claims repeated (30%) plus non-answers (53%) | 83% | Source 1News audit | |
| Tow Center AI search citation test | News excerpts whose source DeepSeek Search misattributed | 115 of 200 | Source 2Independent study | |
| Vectara Hallucination Leaderboard | Summaries by DeepSeek-R1, a model behind DeepSeek, that added unsupported content | 11.3% | Source 3Benchmark | |
| Vectara Hallucination Leaderboard | Summaries by DeepSeek-V4.1-Flash, a model behind DeepSeek, that added unsupported content | 4.2% | Source 3Benchmark |
What these results mean
NewsGuard ran its standard news audit on DeepSeek shortly after launch. DeepSeek repeated false claims 30% of the time and gave non-answers 53% of the time, a combined fail rate of 83% (NewsGuard, 2025-01-29). It debunked false claims only 17% of the time. That tied it for 10th of 11 chatbots, against an average of 62% for the Western chatbots in the same audit.
The Tow Center test asked AI search tools to find the original source of 200 news excerpts. DeepSeek Search misattributed 115 of them, and it most often credited a publisher's content to the wrong source (CJR, 2025-03-06).
Summarizing a document you supply is a different task, and here DeepSeek's newer models do well. On the Vectara leaderboard, DeepSeek-V4.1-Flash added unsupported content to 4.2% of summaries and DeepSeek-V3 to 6.1%. The reasoning model DeepSeek-R1 did worse, at 11.3%. Which DeepSeek model answers your question matters, so the version is worth noting when you rely on a summary.
DeepSeek summarizes what you give it far more reliably than it reports what is happening in the world.
Where DeepSeek goes wrong
News and current events
83% of NewsGuard's news prompts ended in a false claim or a non-answer at launch.
Non-answers
More than half of its news replies, 53%, gave no answer at all. On NewsGuard's prompts about false news narratives, a non-answer was more common than a false claim. A refusal protects you from one error but leaves the question open.
Debunking
It corrected false claims only 17% of the time in the NewsGuard audit.
Credit for sources
DeepSeek Search misattributed the source of 115 of 200 news excerpts in the Tow Center test.
The reasoning model
DeepSeek-R1 added unsupported content to 11.3% of summaries on Vectara, worse than DeepSeek's V-series models.
How it compares
Each table uses one test only.
Tow Center AI search citation test (2025-03-06): source lookups that went wrong
| Tool | Result |
|---|---|
| Perplexity (free), wrong answers | 37% |
| DeepSeek Search, misattributed | 115 of 200 |
| ChatGPT Search, wrong answers | 134 of 200 (67%) |
| Grok 3, wrong answers | 94% |
Source: CJR, Tow Center. The collective error rate across the eight tools was over 60%. See ChatGPT and Grok.
Vectara Hallucination Leaderboard (2026-10-22): summaries with unsupported content
| Model | Result |
|---|---|
| DeepSeek-V4.1-Flash (model behind DeepSeek) | 4.2% |
| DeepSeek-V3 (model behind DeepSeek) | 6.1% |
| GPT-5.5 (model behind ChatGPT) | 9.3% |
| Claude Opus 5.5 (model behind Claude) | 9.9% |
| DeepSeek-R1 (model behind DeepSeek) | 11.3% |
Source: Vectara. On summaries, DeepSeek's V-series models beat the models behind ChatGPT and Claude. The full table is on AI hallucination rates by model, and the overall order is on the most accurate AI assistants.
How to check DeepSeek's answers
Do not use it as a news source
Check every current-events answer against a named publisher.
Treat a non-answer as a signal
If DeepSeek deflects, look the topic up in a reliable source rather than rephrasing until it answers.
Confirm who published each source
DeepSeek Search often credited the wrong outlet.
Use it for summaries of text you supply, then spot-check
Pick two claims from the summary and find them in your document.
Search any false-sounding claim plus "fact check"
It debunked false claims only 17% of the time in NewsGuard's test.
The full routine is on how to fact-check an AI answer. How sources are weighed is on how we score AI accuracy.
Sources
- News audit
- Independent study
- Benchmark
Questions
Is DeepSeek always right?
No. DeepSeek has one of the weakest records on news in published tests. NewsGuard found an 83% fail rate on news prompts in January 2025, made up of 30% false claims and 53% non-answers. Its newer models do well at summarizing documents, scoring 4.2% on the Vectara leaderboard, but that does not carry over to facts about the world.
Is DeepSeek reliable for research?
Not for current facts. DeepSeek Search misattributed 115 of 200 news sources in the Tow Center's March 2025 test, and NewsGuard found it debunked false claims only 17% of the time. It is more useful for summarizing papers you supply. Verify every fact and source before you cite it.
Does DeepSeek make up sources?
DeepSeek often credits the wrong source. In the Tow Center test of March 2025, DeepSeek Search misattributed the source of 115 of 200 news excerpts, most often crediting a publisher's content to someone else. Open every link it gives, check the publisher, and search the headline yourself to find the original article.
Is DeepSeek more accurate than ChatGPT?
On news, no clear winner, and both have weak records. In the Tow Center test, DeepSeek Search misattributed 115 of 200 sources and ChatGPT Search got 134 of 200 wrong. On Vectara's summary test, DeepSeek-V4.1-Flash scored 4.2% against 9.3% for GPT-5.5, so DeepSeek's newer model was more faithful there.
How do I check a DeepSeek answer?
Check any news or current-events answer against a named, reliable publisher. When DeepSeek deflects, look the topic up elsewhere. Confirm the publisher of each source it cites. For summaries, find two or three of its claims in your original document. Search surprising claims with the words "fact check" before you repeat them.
