Can AI show where its answer came from? Yes, but a source link is a pointer to check, not proof.
Short answer. Yes, when the system retrieves passages from your documents and is told to answer only from them. It can then show the passage, the file and the page beside each claim. A shown source is a pointer to check, not proof. In a human audit of four generative search engines, only 51.5 per cent of generated sentences were fully supported by their citations, and 74.5 per cent of citations supported their sentence. On the ALCE benchmark even the best models lacked complete citation support about half the time. The remedies are design choices: quote the exact passage, let the reader open the page, make not in the documents an allowed answer, and sample citations for checking.
01
Three kinds of citation, from easiest to hardest to check
Kind
What it is
How checkable it is
Retrieved passage shown
The system fetches passages from your files and shows exactly which ones the model was given.
Easiest: the reader can read the passage and judge whether it backs the claim.
Citation written by the model
The model names a source as part of its answer, from the text it was given or from memory.
Harder: the source can be wrongly attached, or invented if it came from memory.
Source recalled from memory with no retrieval
A plain chatbot lists references it believes support the answer.
Hardest: a 2023 study found 55% of GPT-3.5 and 18% of GPT-4 citations in generated literature reviews were fabricated.
02
Two ways a citation can fail
Liu, Zhang and Liang audited four generative search engines in 2023 and separated two measures. Citation recall asks whether every statement is fully supported by citations. Citation precision asks whether every citation really supports its statement. On average only 51.5% of generated sentences were fully supported by citations, so claims were left unbacked, and only 74.5% of citations supported their associated sentence, so a quarter of citations pointed at something that did not back the claim.
The ALCE benchmark of Gao and colleagues reached the same conclusion from another direction. Even the best systems it tested lacked complete citation support 50% of the time on the ELI5 questions. Retrieval, first described by Lewis and colleagues in 2020 for knowledge-intensive tasks, makes answers more traceable; it does not make every sentence faithful to its source.
03
What a person must still check
Open the cited page and read the sentence. Does it say what the answer says, or only something near it?
Numbers, dates and conditions. A passage can mention a figure that an answer misstates.
Negations and exceptions: the claim may leave out an unless that sits two lines below.
That the passage comes from the current version of the document.
Inference. If the answer says more than the passage does, treat the extra as the model’s own.
04
How an office would set this up
Keep the file name, page number, section heading and date with every passage when you index documents.
Show the retrieved passage text beside the answer, not only a link or a number.
Tell the model to answer only from the passages and to say not in the documents when they do not cover the question.
Log which passages were given to the model for each answer, so a doubtful answer can be traced.
Sample 20 answers a week, check each against its passage and record the errors you find.
Raise the sampling when you change the model, the prompt or the documents.
05
What ANVE does today
ANVE’s stated aim is every answer cited, every action owned and every model call logged. In its documents function, question answering over the office’s own documents shows the source passage. In the meeting minutes maker, every action item must quote the transcript line it came from. In the agent loop, a claim made without a successful tool call is challenged. The research maker looks sources up online only when asked.
These run today in ANVE’s development build. No sealed unit exists yet, and ANVE publishes no citation accuracy figure.
06
Questions people ask
Does a citation mean the answer is correct?
No. A citation shows where the system says it found the claim. Studies of citation quality find that many citations do not fully support the sentence they are attached to, so a reader should check anything that matters.
Can AI invent a source?
Yes, especially when it answers from memory with no retrieval. A 2023 study found 55% of GPT-3.5 and 18% of GPT-4 citations in generated literature reviews were fabricated. Retrieval from your own files reduces the risk, because the system can only cite passages it has.
What should a good citation contain?
The file name, the page or section, the exact passage, and the version or date of the document. With those a reader can verify the claim in seconds.
How many answers should an office check?
At first, a sample every week, say 20 answers, and more after any change of model, prompt or documents. If the errors found are rare and minor for a few months, you can sample less.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
