A language model doesn’t know your business. It knows what it was trained on, plus whatever it can look up at the moment you ask. When the answer it needs isn’t in either place, it does what it was built to do: it produces the most plausible sentence it can. Sometimes that sentence is wrong, and it reads exactly as confident as one that’s right.
That isn’t a reason to avoid AI. It’s a reason to be deliberate about what your AI is allowed to read.
The problem has a name
NIST’s Generative AI Profile (AI 600-1) lists it as one of twelve risks specific to generative AI. NIST calls it confabulation: “the production of confidently stated but erroneous or false content (known colloquially as ‘hallucinations’ or ‘fabrications’) by which users may be misled or deceived.”
The word that matters there is confidently. A University of Oxford team publishing in Nature in 2024 built a method for flagging when a model is likely making an answer up, precisely because the wording alone doesn’t give it away. And when Columbia’s Tow Center for Digital Journalism tested eight AI search tools on 1,600 citation lookups in early 2025, the tools misidentified the source more than 60% of the time, and rarely hedged when they did.
None of that means AI is unreliable by nature. It means an AI answer is only as good as the evidence underneath it.
Grounding works, and it has limits
The most studied fix is retrieval: before the model answers, the system fetches relevant documents and hands them over as context. Researchers call this retrieval-augmented generation, or RAG. A peer-reviewed study from Facebook AI Research (Shuster et al., EMNLP 2021) found that retrieval significantly reduces knowledge hallucination in conversational AI.
Stanford’s RegLab tested how far that goes in the real world. In a preregistered study published in the Journal of Empirical Legal Studies in 2025, researchers ran over 200 legal questions through commercial legal research tools built on RAG. The tools hallucinated less than general-purpose GPT-4, which was wrong on 43% of queries. They still hallucinated between 17% and 33% of the time, which is why the authors concluded that vendor claims of “hallucination-free” output were overstated.
The most useful finding is why they failed. In the authors’ words: “Many failures in the three systems stem from poor retrieval.” The tools often failed to find the most relevant source available and answered from a near-miss instead.
So the lesson isn’t “RAG doesn’t work.” It’s that grounding is only as good as what the system can retrieve. Feed a model the right source and it does well. Feed it a near-miss and it will write you a fluent answer about the wrong thing.
The best data isn’t in search results
Most general-purpose assistants ground their answers in the open web. That’s useful for a lot of questions. But the sources you’d actually stake a decision on usually aren’t sitting on a public web page:
- APIs. Live pricing, inventory, shipping status, weather, market data. It changes by the minute, and a search snippet from last week is the wrong answer.
- Registries. Business entity filings, license lookups, vulnerability catalogs, regulatory dockets. These are the authoritative record, and they’re usually queried, not crawled.
- Publishers. Licensed research, trade journals, legal and medical databases. The open web is also getting less open: MIT’s Data Provenance Initiative (NeurIPS 2024) found that in a single year, roughly a quarter of the content from the web’s most critical sources for AI training became restricted to AI crawlers through robots.txt.
- Your own data. Contracts, tickets, SOPs, the CRM, the file share. No public model has ever seen it, and none should.
Each of those is a high-fidelity source. Each of them, until recently, was also a separate integration project with its own authentication, its own format, and its own maintenance burden. Three AI tools and five data sources meant fifteen custom connectors. Anthropic put it plainly when it introduced the Model Context Protocol in November 2024: “Every new data source requires its own custom implementation.”
One standard instead of fifteen connectors
The Model Context Protocol (MCP) is an open standard for connecting AI applications to data and tools. Each source is exposed once, as an MCP server. Any AI application that speaks MCP can then use it. The math changes from every-tool-times-every-source to every-tool-plus-every-source.
It’s also no longer one company’s project. In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI and backed by Google, Microsoft, AWS, Cloudflare, and Bloomberg. ChatGPT, Gemini, Microsoft Copilot, and Visual Studio Code all support it.
In practice, that means a single connection layer can give your AI access to the sources that actually deserve trust: a publisher’s licensed archive, a government registry, your accounting system, and your internal documents, each with its own permissions, through one consistent interface. The answer comes back grounded in a named source you can check, instead of in whatever the model happened to remember.
What a well-connected AI looks like
Connecting more data is only half the job. The other half is connecting it carefully:
- Start with the question, not the data. Pick the decisions your people make every week and the sources they already trust to make them. Connect those first.
- Prefer the authoritative source. Query the registry instead of a page that quotes the registry. Pull from the live API instead of a cached export.
- Require citations. Configure your assistant to show where every factual claim came from. An answer you can trace is an answer you can verify in seconds.
- Give the AI the least access that does the job. Read-only where possible, scoped per user, logged. An AI connected to your data is a privileged user, and it deserves the same controls as one. Mixing private data with untrusted content is where the lethal trifecta lives, so design that boundary on purpose.
- Keep a human on the high-stakes calls. The Stanford authors’ conclusion applies outside law too: better retrieval reduces errors, and review catches the rest.
The bottom line
Hallucination is a known, measurable, and manageable property of today’s AI. The single biggest lever you control is what your AI can read. Connect it to authoritative APIs, registries, licensed publishers, and your own records, and it stops guessing and starts citing.
That’s the integration work we do. If you want your AI grounded in sources you’d actually stand behind, our LLM integration service connects them securely, with the permissions and audit trail to match.