Is Grok accurate?
Grok 3 got 94% of news source lookups wrong in the Tow Center test, the worst of eight tools, and the Grok 4 models behind it now sit near 20% on Vectara's summary test.
Short answer
Grok has the weakest accuracy record of the major AI assistants in published tests. In the Tow Center test, Grok 3 named the wrong source for 94% of news excerpts, and 154 of 200 of its citations led to error pages. xAI reports far lower error rates, but those are its own figures.
Results
- Independent study
- Benchmark
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| Tow Center AI search citation test | Wrong answers when Grok 3 was asked to name the source of a news excerpt (200 queries) | 94% | Source 1Independent study | |
| Tow Center AI search citation test | Grok 3 citations that led to error pages (fabricated or broken URLs) | 154 of 200 | Source 1Independent study | |
| Vectara Hallucination Leaderboard | Summaries by Grok 4 Fast (reasoning), a model behind Grok, that added unsupported content | 20.2% | Source 2Benchmark | |
| xAI internal evalVendor's own figure, not independently checked | Hallucination rate of Grok 4.1 (non-reasoning) on production info-seeking prompts, with search tools | 4.22% | Source 3Vendor claim |
What these results mean
The Tow Center asked eight AI search tools to find the original source of 200 news excerpts. Grok 3 did worst, with 94% wrong answers (CJR, 2025-03-06). Worse than the wrong answers were the links: 154 of Grok 3's 200 citations went to error pages. Even its correct answers often pointed to made-up URLs. The paid tier, $40 a month at the time, was not more accurate.
Vectara measures whether a model sticks to a document it summarizes. Grok 4 Fast (reasoning), one of the models behind Grok, added unsupported content to 20.2% of summaries, and Grok 4.1 Fast (reasoning) to 19.2% (Vectara, 2026-10-22). The older Grok 3 scored 5.8% on the same test, so the newer Grok 4 family is less faithful here than its predecessor.
xAI tells a different story. It reported that Grok 4.1 (non-reasoning) hallucinated on 4.22% of production info-seeking prompts, against 12.09% for Grok 4 Fast (non-reasoning), and a FActScore of 2.97% against 9.89% (xAI, 2025-11-17). These are the vendor's claims, measured by xAI, and they conflict with the Vectara results for Grok 4.x models.
In the Tow Center test, most of Grok 3's citations led nowhere.
Where Grok goes wrong
Fabricated and broken links
154 of 200 Grok 3 citations led to error pages in the Tow Center test. A link in a Grok answer is not evidence that a source exists.
Naming the source of news
Grok 3 was wrong 94% of the time, the worst of eight tools. The collective rate across all eight was over 60%.
Summaries that add content
The Grok 4 Fast reasoning models sit near 20% on Vectara, among the higher rates on the board.
Paying more does not help
The Tow Center found Grok's paid tier was not more accurate than the free one.
For current events, the safest reading of this record is to treat a Grok answer as a lead to follow up, not a finding to repeat.
How it compares
Each table uses one test only.
Tow Center AI search citation test (2025-03-06): wrong answers when naming a news source
| Tool | Result |
|---|---|
| Perplexity (free) | 37% |
| ChatGPT Search | 134 of 200 (67%) |
| Grok 3 | 94% |
Source: CJR, Tow Center. Grok 3 placed last of eight. See Perplexity, which placed first, and ChatGPT.
Vectara Hallucination Leaderboard (2026-10-22): summaries with unsupported content
| Model | Result |
|---|---|
| GPT-5.5 (model behind ChatGPT) | 9.3% |
| Claude Opus 5.5 (model behind Claude) | 9.9% |
| DeepSeek-R1 (model behind DeepSeek) | 11.3% |
| Grok 4.1 Fast reasoning (model behind Grok) | 19.2% |
| Grok 4 Fast reasoning (model behind Grok) | 20.2% |
Source: Vectara. The Grok 4 models had much higher rates than the GPT-5.5 and Claude Opus 5.5 models on this test. The full table is on AI hallucination rates by model. Grok's place in the overall order is on the most accurate AI assistants.
How to check Grok's answers
Click every link before you trust it
In the Tow Center test, most Grok 3 links went to error pages.
Search the claim yourself
If you cannot find the fact on a reliable site, do not use it.
Treat breaking news answers as rumours
Wait for the original publisher or an official statement.
Do not rely on xAI's own error rates
Independent results for Grok 4.x models are much higher.
Compare with a second tool
If Claude or a search engine disagrees, go to the primary source.
The full routine is on how to fact-check an AI answer. How vendor claims are weighed against independent tests is on how we score AI accuracy.
Sources
- Independent study
- Benchmark
- Vendor claim
Questions
Is Grok always right?
No. Grok has the weakest record of the major assistants in published tests. In the Tow Center's March 2025 test, Grok 3 named the wrong source for 94% of news excerpts. On the Vectara leaderboard of 2026-10-22, Grok 4 Fast (reasoning) added unsupported content to 20.2% of summaries. Check anything it tells you.
Is Grok reliable for research?
No. Grok is a poor research tool on the published evidence. In the Tow Center test, 154 of 200 Grok 3 citations led to error pages, so its references often cannot be followed. Use a tool with a better sourcing record, such as Perplexity, and cite only pages you have opened and read yourself.
Does Grok make up sources?
Yes, often. The Tow Center found 154 of 200 Grok 3 citations led to error pages, with fabricated or broken URLs, in March 2025. Even its correct answers often linked to made-up URLs. Never cite a Grok link without opening it, and search for the original article by its headline if the link fails.
Is the paid version of Grok more accurate?
Not on the published evidence. The Tow Center found Grok's paid tier, $40 a month at the time of its March 2025 test, was not more accurate than the free version. xAI reports lower hallucination rates for Grok 4.1, but those are the vendor's own figures and conflict with independent Vectara results for Grok 4.x models.
How do I check a Grok answer?
Click every link and confirm the page loads and contains the claim. Search the key fact on reliable sites yourself. Treat answers about breaking news as unconfirmed until the original publisher reports them. Compare with a second AI tool or a search engine, and go to the primary source whenever the two disagree.
