Can grounded AI still get the source wrong?

Groundedness usually asks whether the claims an AI generated are supported by the context available to it. That is an important check. But there is another question worth asking:

What changed, disappeared, widened, narrowed or gained force when the source became the answer?

A response can stay close to its source while losing a condition, dropping an exception, strengthening certainty, widening scope or omitting information that materially changes the result.

That is the difference between finding support for a claim and checking whether the source survived the transformation into the output.

---

What does grounding mean in RAG?

Retrieval-augmented generation, or RAG, gives a language model external information to use while generating an answer.

Instead of relying only on what the model learned during training, the system retrieves relevant documents, passages or records and places that information into the model's working context.

Groundedness evaluates what happens next.

The exact definitions vary between evaluation frameworks, but groundedness generally asks whether the generated response stays supported by the context it was given rather than introducing information that the context does not establish.

Microsoft Foundry, for example, treats retrieval, groundedness, relevance and response completeness as separate evaluation dimensions. Its groundedness evaluator looks at how well the generated response aligns with its context without fabrication.

That distinction matters.

Good retrieval does not automatically produce a good answer. The system can retrieve the right source and still use it poorly during generation.

Grounding helps detect that.

But it does not make every other source-integrity problem disappear.

Is groundedness the same as faithfulness?

Not always.

There is no single industry-wide definition that cleanly separates groundedness from faithfulness.

Some systems use the words almost interchangeably. Microsoft documentation, for example, notes that groundedness is sometimes called faithfulness.

Other evaluation systems distinguish them more sharply. Openlayer defines groundedness around whether generated claims are supported by retrieved context, while using faithfulness to describe whether the response accurately represents the source without distortion, omission or inversion.

So arguing over the labels is not particularly useful.

The more important question is:

What relationship between the source and the output is actually being tested?

One common test starts with what the AI generated and asks whether evidence exists to support it.

That is claim support.

Another useful test starts with the source and asks what happened to its important meaning when the AI transformed it.

That is meaning preservation.

You can care about both.

What can a grounded answer still get wrong?

Consider a simple source statement:

SourceThe refund is available if cancellation occurs before May 1.

An AI summarizes it as:

AI outputThe customer is entitled to a refund.

The output did not suddenly invent refunds. The source really does establish a refund.

But something important disappeared.

The condition did not survive.

The same problem can take several forms.

None of these failures requires the AI to wander onto an unrelated subject.

The words can remain plausible.

The answer can look clean.

The original document can even be attached as a citation.

But the representation has changed.

Claim support and meaning preservation are different tests

This is the distinction that matters.

That creates two directions of inspection.

Neither direction makes the other unnecessary.

They answer different questions.

Can an answer be supported and still be misleading?

Yes.

Suppose a policy states:

SourceManagers may approve expenses up to $5,000 without additional authorization, except for travel outside Canada.

And the AI produces:

AI outputManagers can approve expenses up to $5,000 without additional authorization.

There is direct textual support for much of that sentence.

If you search the source for each major concept, they are all there:

Managers. Approval. $5,000. No additional authorization.

The problem is not that the AI introduced an unsupported number or invented a new policy.

The problem is that an exception governing when the statement applies disappeared.

The resulting output is broader than the source.

This is why source integrity cannot always be reduced to a binary question of whether evidence exists.

Sometimes the evidence supports the words while the transformation changes what those words amount to.

Are citations enough to trust an AI answer?

No.

Citations are valuable because they provide traceability. They can show where an answer came from and make inspection easier.

But three separate questions remain:

Is there a citation?

Does the cited source actually support the generated claim?

Did the generated claim preserve the relevant meaning of that source?

Those are not the same test.

A citation can point to exactly the right paragraph while the summary attached to it drops an exception from that paragraph.

The citation has done its job: it showed you the evidence.

It has not independently established that the transformation of that evidence was correct.

That is why citations and source verification should be treated as complementary rather than interchangeable.

Why omissions are different from hallucinations

AI reliability discussions tend to focus heavily on information the model adds.

That makes sense. Fabricated names, numbers, events and policies are obvious failures.

But AI can also change an output by removing something.

Take this source:

SourceThe service automatically renews for twelve months unless either party provides written notice at least 60 days before renewal.

AI summaryThe service automatically renews for twelve months.

Nothing in the second sentence is fabricated.

But the output is still materially less complete than the source because the mechanism for preventing renewal has disappeared.

The same thing can happen with conditions, exceptions, qualifications, deadlines, uncertainty, attribution and scope.

This matters because omissions can be harder to notice than inventions.

An unsupported claim gives you something new to challenge.

A missing exception leaves nothing behind to inspect unless you compare the output back against the source.

Can AI change certainty without changing the basic fact?

Yes.

Consider:

SourceThe company expects the project will probably be completed in Q4.

AI outputThe project will be completed in Q4.

The topic has not changed.

The project has not changed.

The date has not changed.

But probability became commitment.

Now consider:

SourceEarly evidence suggests the treatment may improve recovery time.

AI outputThe treatment improves recovery time.

Again, the semantic movement is small at the word level and significant at the meaning level.

This is one reason AI-generated analysis can look correct during a quick review. The major nouns and numbers survive while the relationship between them changes.

What if there is no ground-truth answer?

Many evaluation approaches become easier when you already know what the correct output should be.

Real AI-assisted work often does not have that luxury.

Someone may ask an AI system to produce an executive brief from 40 documents.

Or summarize a set of interviews.

Or analyze emails, spreadsheets and meeting transcripts.

Or turn a policy into guidance.

There is no pre-written perfect answer waiting to be compared against the generated one.

The source material is the authority.

That changes the verification problem.

Instead of asking whether the generated answer resembles an expected answer, you can ask whether the generated representation preserved the relevant meaning contained in the bounded sources behind it.

Microsoft's current RAG evaluation framework illustrates this difference. Its groundedness evaluator works against provided context, while its response-completeness evaluator compares the generated response with expected information represented by a ground-truth response.

In many business workflows, that ground-truth response simply does not exist.

The documents do.

What should source verification check?

If the objective is to understand whether AI preserved its source, the useful unit is not simply "true or false."

You want to know how the representation changed.

A source-integrity check can look for changes in conditions, exceptions, obligations, certainty, scope, numbers, timing, attribution, causal strength, availability or status.

It can also look for two opposite problems: information that appeared in the output without adequate source support, and relevant source information that disappeared from the output.

Those differences are useful because they describe the failure rather than collapsing everything into a single score.

"Something changed" is less actionable than:

The exception was removed.

The output changed "may" into "will."

The conclusion combines two supported facts into a relationship neither source actually states.

That is the level at which source verification becomes useful in real work.

What happens when AI uses multiple sources?

The problem becomes harder.

Imagine an AI analysis built from a customer interview, a spreadsheet, a policy document, two emails and a meeting transcript.

Now the system has to preserve not only individual facts but also relationships between sources.

One source may qualify another.

Two sources may disagree.

A newer document may supersede an older one.

A claim may belong to one speaker but not another.

An exception may appear in only one document.

And several individually supported facts can be combined into a conclusion that none of the sources actually makes.

This is why simply asking whether an AI answer contains grounded claims does not fully describe the integrity of a multi-source analysis.

The synthesis itself matters. That is its own problem: what happens when the sources disagree.

So what is the difference between grounding and source verification?

Grounding asks whether an AI output is supported by the information available to it.

Source verification asks what happened to that information as it became the output.

A useful way to think about it is:

These checks can work together.

Grounding can identify generated statements that extend beyond available evidence.

Source verification can make other forms of movement visible, including lost conditions, changed certainty, widened scope, altered obligations and material omissions.

The point is not that groundedness is inadequate.

It is that support for a claim does not by itself prove preservation of the source.