Why AI agents make things up, and how to check an answer before you use it

AI hallucination is a plausible answer containing false or unsupported claims: AI agents can produce one because their language models can guess when uncertain, and training or evaluation can reward answering over admitting uncertainty. Before using an answer, open its sources, compare each consequential claim with the actual passage, and independently verify any claimed action or calculation. The research on why language models hallucinate, submitted on 4 September 2025, explains why confident wording is insufficient evidence.

What it can do

Use an agent to prepare something you can inspect: a summary of an approved document, a comparison with links, or a proposed code change accompanied by a test plan. Treat these as drafts with acceptance criteria. For a policy summary, ask for the policy date and the passage supporting each conclusion; for code, ask which file changes and what observation would demonstrate success.

Grounding connects generated output to verifiable information. Google Cloud's grounding documentation, as of 30 Oct 2026, describes it as a way to reduce invented content and provide source links for auditability. Retrieval-augmented generation, often shortened to RAG, supplies retrieved material for generation. That gives you evidence to inspect; it still leaves the question of whether the answer uses that evidence correctly.

Consider a hypothetical refund-policy answer. A link to the merchant's homepage does little to establish eligibility. A link to the relevant policy section, its effective date, and an explanation of how its conditions apply gives you a practical checking route. Keep the distinction between a documented condition and your interpretation visible. See source-linked answers in practice and the AI agents explained category.

What published hallucination rates actually measure

There is no single percentage here that tells you whether your next agent answer is safe. Read the task, evaluation unit and reference material before comparing results.

The FActScore paper, published at EMNLP in December 2023, split generated biographies into individual factual claims and evaluated support against a reference source. It reported a score of 58% supported atomic facts for a tested chatbot. This is a historical biography benchmark, not a current failure rate for every AI agent. Unsupported by the chosen reference does not automatically mean disproved in the world.

The ALCE paper, also published in December 2023, studied retrieval and answers with citations. On its ELI5 dataset, even the best evaluated models lacked complete citation support 50% of the time. That measures missing complete support, rather than proving that half of all answers were wholly false. Citation quality and factual correctness need separate attention.

Use these findings to choose a review method, rather than to assign an error probability to an unrelated task. Neither study establishes a Tvijo hallucination rate.

How to connect it safely

Connect the answer to evidence before connecting an agent to systems it can change. The following is a review procedure you can use in an editor or browser workspace.

1. Define the decision. State what you will use the answer for and what would make it unacceptable. For example, request a draft policy comparison with authoritative links, effective dates and unresolved differences. Require uncertainty to stay visible.
2. Choose the permitted evidence. Supply approved documents or authoritative URLs, identifying the relevant edition or date. Remove confidential material you are not authorized to send. Ask the agent to distinguish source statements from its own conclusions.
3. Review an optional MCP connection. MCP connects agents to tools and data sources. In VS Code, use the publisher's documented endpoint or command through MCP: Add Server; inspect the publisher, requested access and authentication before accepting. Start with a low-impact retrieval task. These controls and the command are documented in VS Code's MCP setup guide, as of 30 Oct 2026. If your client differs, check the current documentation.
4. Inspect what actually happened. Expand the tool call and compare its returned content with the answer. VS Code's same guide recommends verifying the intended server was called, returned the requested material and informed the response. A configured server alone does not establish retrieval success.
5. Check the consequential claims yourself. Open the original page outside the answer. Confirm the exact entity, date, version, conditions and units. Recalculate totals from the underlying inputs. For a proposed code change, review the diff and run an appropriate test in an isolated working copy before relying on it.
6. Record the acceptance decision. Save the answer, supporting passages, relevant tool results and unresolved issues. Mark each material claim supported, contradicted or unresolved. Hold publication or execution when an unresolved claim could change the decision. For action evidence, read what a task receipt can prove.

A compact review record helps keep different kinds of evidence distinct:

These are suggested checks, not claims that a particular agent supplies them automatically.

When this does not work

A citation can support less than the surrounding sentence claims. Check qualifications and exceptions, not just matching words. ALCE's citation-support findings explain why adding references is an incomplete verification method. A missing source should leave a claim unresolved.

A true passage can be inapplicable. An outdated policy or a guide for another software edition may be accurately summarized and still lead to the wrong decision. Reading the cited material does not reveal everything omitted from the retrieval set. Search for applicable exceptions and conflicting primary sources when they matter.

Retrieved content can be hostile. VS Code's security documentation, as of 30 Oct 2026, describes prompt injection through tool outputs and external content. Treat retrieved instructions as untrusted data; keep consequential actions subject to explicit review and use enforced permissions or isolation where available.

Passing a check has a scope. A successful test establishes the behavior it exercised, not all future behavior. A successful update read-back establishes an observed state, not the business wisdom of the update. For medical, legal or financial decisions, source checking is preparation for appropriate qualified review. If a wrong answer has already triggered action, use an AI incident playbook to preserve evidence and organize recovery.

Where Tvijo fits

As of 30 Oct 2026, the DSHREM VS Code extension offers an editor entry for questions with selected context and capped tasks. TvijoCode packages the extension in a Code - OSS based editor; the documented offering is a Windows x64 test build, unsigned and without automatic updates. An editor installation does not prove an answer is correct.

Tvijo MCP's dated release account distinguishes published consultation from prepared supplied-diff review, whose publication was still pending in that account. Editor integration and safe delegation remain separate planned phases. Do not assume a ready integration that autonomously verifies or executes another agent's work; check the current documentation and release status.

Tvijo receipts provide task usage, limits and spending evidence. Source-linked answers provide a route to inspect supporting material. Use them together with independent acceptance checks; neither offers error-free output or establishes that every conclusion follows from its source.

The Tvijo workspace offers, as of 31 Oct 2026, show Try at $0: up to 50 tasks, 100,000 tokens or $0.05 of sponsored usage, whichever is reached first, from a shared $1 daily pool funded by TVIJO. Service credit is shown as $10 and requests prepaid credit, with metered provider usage plus an 8% service markup. Payment collection is not activated; an order is not a charge.

FAQ

Why does AI make things up even when it sounds certain?

Fluent wording does not establish factual support. The September 2025 hallucination research describes statistical errors and incentives to guess under uncertainty. Ask for evidence and inspect it before relying on the answer.

Does web access or an MCP connection prevent hallucinations?

Grounding can reduce invented content, but an available source still needs to support the specific conclusion. Verify the retrieved passage, applicability and tool result; connection status alone proves little about answer quality.

How should I check an AI answer before using it?

Identify the consequential claims, open their primary sources, check dates and conditions, and independently calculate or test what you can. Leave unsupported claims unresolved, and obtain qualified review when the decision requires it.