The Best AI in the World Misses 1 in 6 Lies — Your Agent Misses 1 in 3

Ask your AI to write a report and it will sound flawless. Every sentence is confident. Every number is precise. Every name is spelled correctly. And roughly one time in six, the best model on the planet is quietly lying to you — not with an obvious blunder, but with a wrong date, a swapped name, or a detail that never existed. Ask an AI agent to do a multi-step job, and the odds get worse: the best models catch barely two out of three of its hallucinations.

That is not a guess. It is the result of HaluCheck, a new 42,000-case benchmark from Beijing Normal University and Tencent AI Lab that tests whether language models can actually detect hallucinations — in text and in agent behaviour. The paper is "From Twisted Facts to Rogue Agents: Can LLMs Reliably Detect Multi-Dimensional Hallucinations?" (arXiv:2611.08231, 5 November 2026).

What the Researchers Did

Most hallucination benchmarks only ask one question: did the model make something up? HaluCheck asks four. Can the model tell a passage contains a hallucination? Can it point to the exact wrong words? Can it fix them? And when an agent goes wrong, can it find the step where the trouble started?

To build the test, the team manufactured hallucinations deliberately from real sources. They took Wikipedia passages and factual QA benchmarks (HotPotQA, PopQA, HalluQA, CMRC2018) and injected controlled errors. Because they made the errors themselves, they know exactly where each one is, which makes the scoring precise. This half is called HaluText, in two flavours: HaluReal (twisted real facts) and HaluFake (encyclopedia-style descriptions of entities that do not exist at all).

The injected errors follow three patterns found by analysing real model failures. Misattribution moves an attribute from one entity to another. Numerical deviation changes a number, a date, or a ratio. Detail invention adds a plausible-sounding fact that simply is not true. Each error is graded in three difficulty tiers: Obvious, Moderate, and Subtle. And one finding jumps out: numbers and dates are the most hallucinated entity types. If your AI output contains a figure, that is where to look first.

The second half, HaluAgent, matters for anyone running agents. The team took real agent trajectories — multi-turn logs of instructions, reasoning, tool calls, and tool outputs — from five existing agent datasets, and injected hallucinations into them. An agent can go wrong in three ways: it plans the wrong steps, it misreads what a tool returned, or it calls the wrong tool. HaluAgent tests whether a model can spot which kind of failure happened, and identify the earliest step where the trajectory went off the rails.

The scoring is honest in a way most benchmarks are not. Models are allowed to answer "Uncertain" when the evidence is insufficient, and that answer earns partial credit. Saying "I don't know" is treated as better than guessing wrong. Fourteen models from six families were tested: Gemini 3.1 Pro and Flash, GPT-5.4 and GPT-5 Mini, DeepSeek V4 Pro and Flash, Tencent's HY3 Preview, Llama 3.3 70B and Llama 3.1 8B, and five Qwen3 variants.

Why It Matters

First, the best model still misses one lie in six. Gemini 3.1 Pro — the top scorer — detected 84.35% of hallucinations in Chinese text and 84.87% in English. That is the ceiling. Most other models scored between 50% and 70%, and Llama 3.1 8B managed just 52.88% in Chinese — barely better than a coin flip. If you are using a small or mid-tier model to check its own output, it is essentially guessing.

Second, spotting a lie is much easier than fixing it. Even when models correctly flagged a passage as hallucinated, they were far worse at pointing to the exact wrong words (SpanF1 of 70.07% for the best model, and under 50% for most) and at correcting them. Detection without localization is like a spellchecker that says "there's a typo somewhere in this document" and walks away.

Third, agents are far harder to police than text. On HaluAgent, the best sample-level accuracy was just 64.55% in Chinese and 72.36% in English. Finding the root cause — the first bad step in a long trajectory — topped out around 76%. Agent hallucinations are not fact-checking problems. They demand tracking context across many steps, tool calls, and observations, and current models are weak at exactly that.

Fourth, the models that know their limits are the honest ones. On the fabricated-entity test, stronger models were more likely to say "I'm not sure" — while smaller models gave confident, direct answers. Overconfidence, not ignorance, is the failure mode of cheap models. A checker that never says "uncertain" is a checker you cannot trust.

Finally, the fixes help — but none of them finish the job. Letting the model "think" step by step improved localization and correction. Giving it web search helped even more — external knowledge beat internal reasoning. And supervised fine-tuning helped on familiar error patterns but failed to generalise to new ones. Each is a partial improvement, not a solution. Hallucination detection, the paper concludes, remains unsolved.

How You Can Use It

Step 1: Check the numbers first. The paper's own analysis says numbers and dates are the most hallucinated entity types. Before you read a report for sense, scan it for every figure, date, percentage, and ratio, and verify each one against a source.

Step 2: Split the output into atomic claims, then check each one. Break a paragraph into single claims ("X acquired Y in 2024", "the API returns Z"), and verify each claim separately. A whole paragraph can sound true while one claim inside it is invented. Use this prompt template with a second model — or the same model in a fresh session — as your checker:

"Check the text below claim by claim. For EACH claim, reply with one of: VERIFIED (you found supporting evidence), WRONG (you found contradicting evidence — quote the correct fact), or UNCERTAIN (not enough evidence to decide). Do NOT guess. If you cannot verify a claim, say UNCERTAIN. Pay special attention to numbers, dates, names, and attributions. Treat any claim you cannot verify as unusable until checked. TEXT TO CHECK: [paste the AI output here]"

Two details in that template come straight from the paper: "Uncertain" must be a valid answer, and numbers, dates, and attributions get called out explicitly, because that is where the errors cluster.

Step 3: Give the checker web access — and never let a small model check itself. Web search improved detection more than step-by-step thinking did, so your checker should be able to search. And the scoreboard is a warning: Llama 3.1 8B scored near chance, while Gemini 3.1 Pro scored 84%+. Use your strongest model for checking, even if a cheaper one did the writing — verification is where the quality budget belongs.

Step 4: For agents, log every step and check the trajectory, not just the answer. HaluAgent shows that the failure is usually in the middle of the run: a misread tool result, a wrong plan step. Keep the full trace — instruction, reasoning, tool calls, tool outputs — and have your checker review the trace for the first step that goes wrong.

Step 5: Treat fine-tuning as a bandage, not a cure. Supervised fine-tuning helps models recognise familiar hallucination patterns but does not generalise to new ones. Keep a human or a web-enabled model in the loop for everything else.

Key Takeaways

• HaluCheck is a 42,000-case, bilingual benchmark that tests whether LLMs can detect hallucinations — in text and in agent behaviour — including locating the exact error, fixing it, and tracing an agent's first wrong step.
• The best model, Gemini 3.1 Pro, catches about 84% of text hallucinations. Most models manage 50–70%. Small models are barely better than chance.
• Finding the lie is easier than fixing it: localization and correction scores lag far behind detection for nearly every model.
• Agent hallucinations are the hardest case: the best accuracy is only 65–72%, and tracing the first bad step tops out near 76%.
• Numbers and dates are the most hallucinated things an AI writes — check them first.
• What helps: step-by-step thinking, web search (the biggest single boost), and — above all — a checker strong enough to say "I don't know."
• Your move today: split AI output into single claims, verify numbers and dates first, and use your strongest model — with web access — as the checker, never the cheapest one.

Read the original paper: "From Twisted Facts to Rogue Agents: Can LLMs Reliably Detect Multi-Dimensional Hallucinations?" by Hanyu Qiu, Ze Liu, Zhe Zhao, Kuang Xiong, Blaze Chen, Zhenxiang Yan, and Renfen Hu (Beijing Normal University and Tencent AI Lab). Link: https://arxiv.org/abs/2611.08231 — code and benchmark: https://github.com/xMiles/HaluCheck