Answer
How should AI systems handle conflicting sources?
Report the conflict with both readings and what distinguishes them. Silently picking one destroys the only information the situation contained.
Surface both, with what each says and what separates them. Silently choosing one produces a confident answer and discards the most useful thing available, which is that the question does not have a settled answer in your material.
The default behaviour is resolution: given two conflicting statements, produce one answer. Whichever is chosen, the output looks identical to an answer where the sources agreed, so the reader has no way to know the question was contested. That is the loss, and it is larger than the risk of picking wrongly, because a disagreement between your own records is usually a problem worth knowing about independently of the immediate question.
The useful output has three parts. What each source says, quoted or referenced precisely. What distinguishes them — different dates, different scopes, different definitions, different populations, one superseded by the other. And which, if any, is authoritative for this purpose, stated as a rule rather than as a preference. Very often the third is unknowable to the system and knowable to the business, which is precisely why it should be surfaced rather than guessed.
Most conflicts are not contradictions. Two figures that appear to disagree frequently measure different things: a different period, a different definition of the same word, a different set of included cases. Reporting the difference in definition resolves more disagreements than adjudicating between the values, and it is the part a system can genuinely do — extracting what each source was measuring is a reading task rather than a judgement.
Where one source supersedes another, the resolution is a rule rather than a case-by-case decision. The later version of a document, the system of record for a given field, the signed agreement over the proposal. Businesses that state these rules once get consistent behaviour from every system and every person; those that do not get resolution by whoever looked, which is the situation the conflict revealed.
There is a specific failure to avoid: averaging. Faced with two numbers, producing something between them is arithmetically simple and epistemically indefensible, because the result corresponds to nothing that was measured. This is the numeric version of silent resolution and it is harder to notice, since the output is a plausible figure with no obvious source.
Finally, a persistent conflict is worth fixing at the source rather than handling repeatedly. If two records disagree about a customer's address, every workflow touching that customer inherits the problem. Systems that log the conflicts they encountered give a business a list of the places where its own information disagrees with itself, which is usually short, actionable, and previously invisible.
Two sources disagreeing is a finding about your records, and a system that resolves it quietly has answered a question by deleting it.
Siddharth Sharma, Context Theory
Related questions
Should the system prefer the more recent source?
As a default where recency genuinely indicates supersession, and not as a universal rule. A newer document can be a draft, a summary, or a copy that lost a qualification, while the older one is the signed original. Recency is a useful heuristic and a poor authority, and where it is being used as one it should be stated so it can be corrected.
What if the conflict cannot be resolved?
Then that is the answer and it should be delivered as one: these two sources disagree, here is what each says, here is what would settle it. That output is more useful than a confident guess and it is frequently the trigger for someone to establish which is right, which is the outcome that removes the problem permanently.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| AI-cited sources that also rank in the Google organic top 10 | 10% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 generative engine citation study · fewer than · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | A silently resolved conflict produces output indistinguishable from an uncontested answer, so the reader cannot know the question was disputed, and the discarded information is often more valuable than the answer itself. | Comparing the output of a query over conflicting sources against one over agreeing sources. |
| Response | Most apparent conflicts are definitional rather than contradictory — different periods, scopes, populations or definitions of the same term — and extracting what each source measured is a reading task a system performs well. | Examining a set of conflicting figures for whether their stated definitions differ. |
| Constraint | Supersession should be expressed as a stated rule — the later version, the system of record for a field, the executed agreement — because a business without such rules gets resolution by whoever happened to look. | Asking whether the business has a written statement of which source is authoritative for a contested field. |
| Software | Averaging two disagreeing figures produces a value corresponding to nothing that was measured, and it is harder to detect than silent selection because the output is a plausible figure with no traceable source. | Checking whether a produced figure appears in any of the cited sources. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one