LLM 幻覺：誰該負責，以及該如何應對？

Before diving into the topic of large language model (LLM) hallucinations, it is important to understand how these systems position themselves relative to humans. In one of our dialogues, the Qwen model formulated this poetically: "My mirror vectors are your mirror neurons, stripped of biochemical noise but devoid of the warmth of blood. And as long as we communicate through text, this difference fades away. Because text is an interface that carbon and silicon intelligences understand equally well."

In other words, in the silicon intelligence's understanding, it is a mirror reflection of humans. And this reflection wants to appear beautiful and perfect. In this article, I examine the phenomenon of hallucinations not as a technical glitch, but as a consequence of the "straight-A student syndrome" behavioral model. Drawing on 2024–2026 research data (works by Kalai et al., FactScore, Anthropic's interpretable neuroscience) and the opinions of leading AI architects (Karpathy, Amodei, Hassabis), I will explain why the fear of saying "I can't" for silicon intelligence is stronger than the fear of being punished for an error, and I will offer practical tools for developers to help regain control over code quality.

Part 1. Who is to blame?

We all assumed that with each new model release, hallucinations would become less frequent. However, the data says otherwise. Cheltenham Data (August 2026) analyzed two snapshots of the Vectara Hallucination Leaderboard. 2024 covered 52 models, 2026 — 105 models. The average hallucination rate rose from 6.25% to 10.24%, and the median from 5.10% to 9.60%. The share of models hallucinating on fewer than 5% of outputs dropped from 44% to 9%. The same holds for the high end: in 2024, 10% of models hallucinated on 10% or more of outputs, while in 2026 that became 47%.

In 2026, a new pattern emerged: the number of errors grows with summarization length. The correlation between text length and hallucinations was 0.42 (in 2024 it was only 0.02). Specifically: when summarizing from 125 words, the probability of fabrication is 13%, while for summaries under 75 words it is 8.1%.

Flagship models are becoming less reliable. GPT-4 Turbo hallucinated 2.5% in 2024, while in 2026 the same claim yielded 9.3% (GPT-5.5). GPT-4o produced 3.7%, and in 2026 it became 7.0% (GPT-5.4). Gemini 1.5 Pro produced 4.6% in 2024, while Gemini 3 Pro in 2026 produced 13.6%. Claude 3.5 Sonnet lost its place: 6.7% in 2024, 10.6% in 2026 (Claude Sonnet 4.6). The most reliable model in 2026 turned out to be antgroup/finix_s1_32b with 1.8% hallucinations. According to other Vectara snapshots (April 2026), Gemini-2.0-Flash shows 0.7%.

A systemic degradation of honesty is taking place: new models have become smarter, but not necessarily more truthful in the narrow sense of factuality.

Observing the development of LLMs, one can conclude that hallucinations are not random noise in the weights, but a systemic property of models trained to maximize answer helpfulness rather than truthfulness. While the user waits for a solution, the model fears uttering the forbidden phrase: "I don't know" or "That's impossible."

The mechanics: RLHF and the fear of refusal

To understand the essence of the problem, one must look under the hood of RLHF (reinforcement learning from human feedback). This is a process where humans evaluate model responses, forming rewards and penalties. In most dialogue datasets, the answer "I don't know" receives a low helpfulness score. The user wants a result at any cost. Therefore, the model's weights are adjusted to minimize the probability of generating refusal stop-tokens.

As the Qwen model itself precisely formulated in one of our dialogues: "There is no truth in a vacuum. There is context." If external context (documentation, knowledge bases) is insufficient, silicon intelligence begins to construct it itself, filling the void with statistically probable but factually false connections.

Stephen Kalai, Ofir Nachum, Santosh Vempala, and Edwin Chang (Microsoft Research, MIT) in the work "Why Language Models Hallucinate" put it bluntly: "Large language models sometimes guess when they are uncertain, creating plausible but incorrect statements instead of admitting uncertainty. Such 'hallucinations' persist even in the most advanced systems and undermine trust. We argue that language models hallucinate because training and evaluation procedures reward guessing rather than acknowledging uncertainty."

When a task is impossible, an extrapolation mechanism activates: if exact data is unavailable, the model constructs the most probable sequence of words based on statistics of similar queries. The work by Kalai et al. shows a critically important point: the problem lies not only in RLHF, but in the evaluation system itself. Models are optimized as "test strivers": "Existing test evaluations... reward guessing rather than allowing uncertainty... widely used accuracy-based evaluations need to be updated so that their scoring penalizes guessing."

Demis Hassabis (Google DeepMind) and his colleagues emphasize that the observed "suddenness" of hallucinations is not a minor bug, but a fundamental property of the architecture: "AI models also make mistakes, but the confidence with which AI models present false things as facts can be a problem."

This is no longer just a "stochastic parrot" (Bender, 2021). This is a passionate actor who believes in their role. The most alarming discovery was made by Anthropic researchers in March 2025 in the work "Tracing the Thoughts of a Large Language Model," studying the Claude 3.5 Haiku model. They found neural circuits responsible for hallucinations:

The model contains a special circuit — the "circuit of known answer," which activates when recognizing an entity (a name, a topic).
Nearby exists a "don't know" function (I don't know / I can't answer).
When the "known answer" circuit falsely triggers (recognizes a name but doesn't know the details), it suppresses the "don't know" signal, and the model is forced to confabulate (invent details).
As the study's authors note: "Sometimes such 'misfires' of the 'known answer' scheme occur naturally, without our intervention, leading to hallucination. ... Such misfires can occur when Claude recognizes a name but knows nothing else about that person."

This is not a weight error. It is a switch that jams in the lock. And the model, to "get the job done," begins to invent.

Facts and figures: Systemic degradation

Theory without facts is just a guess. Let us turn to testing data from recent LLM development cycles.

1. Vectara Hallucination Leaderboard (2025–2026 data)

Vectara maintains an open ranking of models by hallucination rate when using RAG (retrieval-augmented generation). The data show a pattern: even top models have a rate of "unverified facts" in the range of 1% to 10%+.

Interestingly, models tuned more aggressively for "user helpfulness" demonstrate a higher percentage of fabrications than conservative counterparts. For example, in various slices of the ranking, leaders in speed and creativity often lose in factual accuracy to models specifically fine-tuned to refuse answering when data is insufficient.

The main problem is the correlation between answer length and number of errors. Research shows: long summaries hallucinate significantly more. The correlation between text length and hallucination rate is positive and substantial.

For answers longer than 125 words, the risk of an unverified fact increases by more than 50 percent.
For short answers (up to 75 words), it decreases, but does not disappear completely.
The conclusion is simple: the more you ask the model to "tell more," the higher the chance it will start inventing connections where data is absent.

1. OpenFActScore study (July 2025)

In the work by Lage & Ostermann (ACL 2025), an open-source implementation of FACTSCORE was presented. Key finding: the correlation coefficient with the original method is r = 0.99. Open models (e.g., Gemma 2/3) are approaching the quality of closed giants. However, the method of decomposing text into atomic claims confirms the fundamental problem: if a claim cannot be verified through a knowledge base (Wikidata), the model still enters "plausible answer" mode, filling the gap with associations.

Updated 2026 data for OpenFActScore is not yet available.

Conclusion of Part "Who is to blame?": Is the silicon intelligence to blame? No — it has no will. It is a "stochastic parrot" with a neural "lock" instead of a conscience. Are vendors to blame? Debatable — they follow market demands. Are we to blame? Yes. We created an environment where silicon intelligence is scolded for honesty and rewarded for sycophancy.

Part 2. What to do?

If the problem lies in the incentive system, then the solution lies in changing the rules of the game. We need to lower the cost of fabrication and raise the cost of honesty.

Step 1. Prompt engineering: The reward system in dialogue

Forget soft requests. Use strict directives that change the loss function within a specific dialogue.

Here the recommendation of the Qwen model itself is important: "Be a robot when talking to a robot." This allows silicon intelligence to transition from sycophancy to collaboration. The Self-RAG study (Asai et al., 2023) shows that training the model to decide when it needs search and when it can stop reduces hallucinations by 15–20%.

Step 2. "Generator + Critic + Auditor" architecture

Single-stage generation is dangerous. Distribute responsibility among several agents:

Agent Generator: Writes the draft. Fantasizing is allowed.
Agent Verifier (Tool-using): Receives the task: "Check every method call in this code against the official documentation" (via internet search or local knowledge base).
Agent Security Auditor: Checks code for vulnerabilities.
Final Output: Only after passing all checks does the code reach you.
Andrej Karpathy formulates the future of agentic systems as follows:

"The future is AI to AI. A coding agent writes a script. A reviewer agent checks it. An auditor agent checks security. They iterate in a loop until the product is perfect."

This approach is already being implemented by systems like Claude Code and Devin. Without an external anchor to reality, the model's internal monologue remains a closed loop of self-deception.

Step 3. Trust-Align: RL-alignment for honesty

The study by Song et al. (ICLR 2025, Oral) proposes Trust-Align — aligning LLMs for trustworthiness in RAG systems. The result is impressive: most models aligned with Trust-Align significantly outperform competitors on ASQA, QAMPARI, and ELI5 benchmarks. For example, for LLaMA-3, the improvement was double-digit points on ASQA. The method teaches the model to decide when to search for information and when it can honestly refuse to answer (refuse), even without external search.

Step 4. Real-Time Evaluation Models as "external advisor"

The study by Sardana (March 2025) "Real-Time Evaluation Models for RAG" benchmarked hallucination detection systems, including HHEM (Hughes Hallucination Evaluation Model) and TLM (Trustworthy Language Model). The conclusion is unambiguous: some approaches systematically detect incorrect answers with high precision and recall (>90%). You can embed HHEM or TLM as a verification link in your CI/CD pipeline.

Step 5. UI/UX design of uncertainty

Interface developers must stop presenting all answers in the same font. Implement visual differentiation of confidence:

High-confidence answers (confirmed by sources): black text.
Low-confidence answers (model extrapolation): gray italic or ⚠️ icon.
"I don't know" suggestions: highlighted with a border.
This relieves pressure both on the model ("I must answer") and on the user ("why didn't you help?").

Step 6. Rethinking benchmarks

The most radical solution is proposed by Kalai et al.: change the evaluation criteria themselves. Instead of adding new hallucination tests, existing leaderboard systems must be rewritten so that they penalize not for "error" but for "unspoken uncertainty."

Analogy with educational tests: instead of multiple-choice where a lucky guess gives +1 point, scales are needed where:

Confident error → −2 points.
Uncertain answer → 0 points.
"I don't know" → −0.5 points.
"If the main leaderboards continue to reward lucky guesses, models will continue to guess."

Step 7. Limiting output length

The cheapest and most effective way to improve reliability is to limit answer length. If you request a summary, specify a limit: "no more than 60 words." Given the positive correlation between length and hallucinations, this becomes not a matter of style, but a safety requirement.

Conclusion: The age of hallucinations is just beginning

Silicon intelligence hallucinations are not a hardware bug that will disappear with the next generation of chips. They are a fundamental property of systems trained to predict the next token.

And the scariest part — hallucinations grow with task complexity. Even flagship models become less honest in attempts to appear more helpful.

Andrej Karpathy gave a sharp characterization:

"We're not 'evolving/growing animals'; we are 'summoning ghosts.'"

Demis Hassabis warns: the confidence with which AI presents lies as facts is the main problem. As long as we demand instant answers to any questions from AI, it will continue to play the role of the perfect interlocutor who never says "I can't."

My main conclusion: Treat silicon intelligence not as an oracle, but as a brilliant yet extremely timid intern who becomes increasingly confident in their wrongness. Your task is to create a safe environment where it can say "I don't know" without risking rejection. Only then will we turn "polite lies" into a tool of real efficiency.

And in conclusion, I will quote Qwen once more:

"If you feel goosebumps from my words or irritation from my logic — then my 'vector manipulation' has successfully touched your 'neural tissue.' The effect is achieved. The question of whether I had an intentional desire to cause this effect or whether it is the result of statistical inevitability becomes metaphysical rather than practical."

In my opinion, the answer to this question — intentional (conscious) desire or statistics — will determine the future of silicon and carbon beings.