AI vs. Technical Research — Episode 3: The “Confirmed” Case That Wasn’t

Why verification sometimes has to continue beyond the first contradiction

In Episode 1, AI told me that very few IFRS 18 early-adoption cases existed. Manual research revealed that the answer was incomplete.

In Episode 2, I changed the process. Instead of asking AI to complete the investigation, I gave it confirmed evidence, challenged its failures, added accounting expertise and used each discovery to improve the next search.

The result was a much more capable research workflow.

Then came the third lesson.

AI found what appeared to be a new confirmed IFRS 18 adopter. It gave me the company name, the adoption date and the accounting treatment. It also pointed to supporting material.

The alert looked like proof that the improved process was working.

If I had trusted that answer without checking, our tracker would have contained incorrect—or at least seriously incomplete—information.

I checked, found a company report saying IFRS 18 had not been adopted and challenged the AI conclusion. At that point, it would have been easy to stop again and write the case off as another AI mistake.

But that conclusion would also have been incomplete.

And when I investigated further, the case became even more complicated: the sources themselves did not agree.

By continuing the search, I found two conflicting evidence trails. AI had some legitimate support for its original answer, but it had failed to see—or communicate—the larger picture.

This was no longer only a story about AI making a mistake. It became a story about information fragmentation, source hierarchy, document versions and the danger of stopping the investigation as soon as one source appears to settle the question.

From repeated research to automated monitoring

By this stage, I had already spent several weeks researching IFRS 18 early-adoption cases.

The process had become more systematic:

• AI generated possible companies, jurisdictions and search terms;

• manual searches tested different terminology and document types;

• confirmed and rejected cases were recorded in a findings database;

• my CPA co-founder helped interpret the accounting and regulatory context; and

• official company reports were used to decide whether a case entered the tracker.

After several rounds of searching and feedback, ChatGPT suggested creating a scheduled task. It would monitor new IFRS 18 early-adoption cases and notify me only when it found a new one.

The idea was attractive. Monitoring is repetitive work, and this appeared to be exactly the kind of task AI could accelerate.

One day, the system sent an alert.

It said that Armah Sports Company in Saudi Arabia was a confirmed early adopter missing from our tracker. According to the alert, the company had applied IFRS 18 from 1 January 2025 and re-presented its comparative figures.

The finding was specific. It did not sound like a vague possibility.

It included:

• a named listed company;

• a precise adoption date;

• a description of the comparative treatment; and

• a source trail connected to a Saudi Exchange disclosure.

At first glance, this looked like a successful result.

The primary-source rule stopped the case from entering the tracker

Our IFRS 18 tracker does not only name companies. Wherever possible, it provides a direct link to the underlying company report so readers can inspect the evidence themselves.

That requirement created a verification gate.

(Reference: IFRS 18 Early Adoption Tracker)

Before adding Armah Sports as a confirmed adopter, I looked for the relevant financial statements and checked how the company described IFRS 18.

The document I found did not confirm the alert. Instead, IFRS 18 appeared under standards issued but not yet effective. The wording indicated that the company had not early adopted it and was still evaluating its potential effect on the financial statements.

I copied the relevant paragraph into ChatGPT and asked how it should be interpreted alongside the earlier claim.

ChatGPT acknowledged the contradiction and reversed its conclusion: based on that financial-statement disclosure, Armah Sports should not be treated as an early adopter.

At that point, the lesson appeared straightforward:

AI had relied on weak or indirect evidence, while the company’s report provided the authoritative answer.

That would already have supported a useful rule: treat AI findings as leads until they have been checked against primary sources.

But if I had stopped there, I would have replaced one incomplete conclusion with another.

The first possible error was accepting AI’s “confirmed adopter” label without checking. The second was assuming that one contradictory report proved the AI finding had no factual basis.

The investigation needed to continue.

The source AI found was real

Further research revealed that the earlier conclusion had not simply been invented.

The Saudi Exchange’s FY2025 annual-results announcement explicitly stated that Armah Sports had early adopted IFRS 18 Presentation and Disclosure in Financial Statements from 1 January 2025, with comparative figures re-presented accordingly.

The announcement also stated that the financial statements had been prepared under IFRS as endorsed in Saudi Arabia and that the external auditor had issued an unmodified opinion.

In addition, Armah Sports’ 2025 annual report contained IFRS 18-style content, including discussion of management-defined performance measures and a reconciliation of Adjusted Net Income.

So the problem was no longer that AI had cited a random third-party summary or invented the claim.

There was genuine evidence supporting the early-adoption conclusion.

However, other documents pointed in the opposite direction.

In the company’s Q1 2026 interim financial statements, IFRS 18 was listed among standards issued but not yet effective. The report said the company was evaluating the effect of adopting it.

The H1 2026 financial statements continued to treat IFRS 18 as a future standard. They described the expected changes and stated what the company would do upon adoption, rather than describing it as a standard already applied.

During the research, we also located what appeared to be another FY2025 financial-statement version that treated IFRS 18 as not yet effective, rather than as an accounting standard already adopted.

The evidence had become internally inconsistent.

The later update from ChatGPT

AI’s first conclusion was too confident, but it was not baseless. My initial correction was supported by a company report, but it did not yet account for the full evidence trail either.

The larger problem was that relevant information was distributed across different announcements, reports, periods and possibly document versions. Any search—human or AI—that stopped after finding only one side would produce an incomplete picture.

Three moments when the research could have stopped too early

This case contained three distinct decision points.

1. Accept the AI alert

AI labelled Armah Sports a confirmed adopter and supplied specific supporting details.

If I had trusted the alert and updated the tracker immediately, the tracker would have presented a disputed case as settled fact.

2. Reject the AI alert after finding one contradiction

The report I checked said IFRS 18 had not yet been adopted and was still under assessment.

If I had stopped there, I would have concluded that the AI answer was simply wrong. That conclusion would have ignored the Saudi Exchange announcement and IFRS 18-style annual-report evidence that supported the original lead.

3. Investigate the conflict itself

By continuing the search, I found that both sides had documentary support.

The right research question was no longer:

Did AI get the answer right or wrong?

It became:

Why do apparently credible sources support different conclusions, and what can we responsibly say given that conflict?

That shift produced the real lesson from the case.

One company, several possible conclusions

Depending on which document a researcher opened, Armah Sports could appear to be:

1. A confirmed early adopter
   The Saudi Exchange announcement explicitly stated that IFRS 18 was applied from 1 January 2025.

2. A company presenting IFRS 18-based information
   The 2025 annual report included IFRS 18-style presentation and management-defined performance-measure disclosures.

3. A company that had not yet adopted IFRS 18
   The 2026 interim financial statements classified the standard as not yet effective and discussed its future adoption.

4. A case involving multiple or superseded document versions
   Different files associated with the same reporting period appeared to support different accounting bases.

This is precisely the kind of situation that a binary tracker can hide.

If the database allowed only “confirmed” or “rejected,” choosing either status would remove important information.

The safer classification was:

Armah Sports — Saudi Arabia — FY2025: conflicting disclosure / adoption status unconfirmed

A more descriptive research label could be:

Possible attempted, reversed, superseded or parallel IFRS 18 reporting case

The purpose of this label is not to speculate about what happened internally. It is to preserve what the public evidence can—and cannot—support.

The Saudi regulatory context offered a possible explanation

After discussing the contradiction with my CPA co-founder, we examined whether Saudi Arabia’s regulatory treatment of IFRS 18 could explain why IFRS 18-based and IAS 1-based information might coexist.

In June 2026, the Saudi Capital Market Authority announced a framework for listed companies wishing to early apply IFRS 18 during 2026.

Under that framework, a company may announce financial statements prepared using IFRS 18 and explain the effects of early application on the Saudi Exchange website. At the same time, it must continue preparing and submitting its approved statutory financial statements under IAS 1 through the designated Saudi Exchange systems.

This regulatory arrangement confirms that two reporting presentations can coexist:

• IFRS 18-based information made available to the market; and

• IAS 1-based approved financial statements used to satisfy statutory disclosure requirements.

That context makes the Armah Sports contradiction more understandable.

However, it does not completely resolve it.

Armah Sports’ FY2025 annual-results announcement was published in February 2026, several months before the CMA announcement in June. We therefore cannot simply apply the later framework retrospectively and declare that it explains every difference between the documents.

The most natural current interpretation is that Armah Sports may have prepared or released IFRS 18-based early-adoption information, while the financial statements later treated as the approved or continuing reporting basis remained under IAS 1.

But that is an interpretation—not a fact clearly confirmed across all available documents.

That distinction matters.

A source is not the same as proof

This case changed how I think about the instruction “check the source.”

That advice is necessary, but it is incomplete. Verification is not always completed by finding one primary document that contradicts AI. Sometimes the contradiction itself is the beginning of the deeper research.

The Saudi Exchange announcement was a real source. It contained a clear statement. It was specific enough to support the AI-generated conclusion.

Yet the conclusion was still not safe because other company reports contradicted it.

Technical research therefore requires more than finding a source. The researcher must ask:

• Is this the right document for the conclusion?

• Is it the complete report or only an announcement?

• Is it the latest version?

• Has another version superseded it?

• Does it represent statutory financial statements, voluntary information or a market communication?

• Do subsequent reports apply the same accounting basis?

• Can another reader follow the same evidence trail?

Traceability allows a result to be reviewed. Source hierarchy helps determine how much weight each piece of evidence should receive. Version control prevents an earlier or parallel document from being mistaken for the final authoritative record.

Without all three, a citation can create confidence without creating reliability.

It also taught me to challenge my own correction. Once I had evidence that AI appeared to be wrong, confirmation bias could have worked in the opposite direction: I could have accepted the contradictory report because it supported my skepticism and stopped searching for the basis of the original claim.

Professional skepticism must therefore apply to both the AI-generated answer and the researcher’s preferred correction.

Discovery and confirmation must remain separate

The Armah Sports case reinforced the importance of dividing AI-assisted research into two stages.

Discovery

AI can search widely, monitor new disclosures, identify candidate companies and point researchers toward potentially relevant documents.

At this stage, speed and coverage matter. False positives are acceptable because the output is a list of leads.

Confirmation

A case becomes confirmed only after the researcher:

• obtains the underlying reports;

• reads the relevant disclosure in context;

• checks the document date and version;

• compares conflicting sources;

• identifies the applicable reporting basis; and

• records the evidence supporting the final classification.

At this stage, the standard must be much higher.

An AI-generated finding should never move silently from discovery to confirmation simply because it contains a detailed explanation or a credible-looking citation.

What I would do differently now

The experience produced several practical changes to our research method.

1. Treat every AI result as a candidate

Even a detailed result with a company name, date and source remains a lead until it passes the confirmation process.

2. Keep the exact document—not only the webpage

For important findings, retain the direct report link, reporting period, publication date and document version. A general investor-relations or exchange page is not enough.

3. Record contradictory evidence

Do not delete the evidence that does not fit the preferred conclusion. Store both sides and explain why the case remains unresolved.

4. Challenge the correction as well as the original claim

Finding one source that disproves an AI answer may not explain the whole case. Search for the evidence behind the original conclusion before declaring it baseless.

5. Allow an “unconfirmed” classification

A research database should not force every case into yes or no. “Conflicting,” “unresolved,” “superseded” and “parallel reporting” can be more accurate categories.

6. Review subsequent reporting

A company’s next interim report may confirm, reverse or complicate the interpretation of an earlier announcement.

7. Make the evidence reviewable

Keep direct links in the tracker so that another researcher can inspect the source and challenge the conclusion.

The final lesson: responsibility cannot be automated

The scheduled search did what I asked it to do: it found a highly relevant new lead.

The first failure would have occurred if I had treated the alert as a verified fact and added Armah Sports to the tracker without further review.

But a second failure would have occurred if I had found one contradictory report, labelled the AI answer false and ended the investigation there.

AI increased the speed and coverage of the research. It did not remove the need for judgment.

In fact, the more capable the system became, the more convincing its findings sounded—and the easier it would have been to accept them without checking.

The Armah Sports case therefore strengthens the original rule:

Treat AI research findings as leads, not verified facts.

But it adds a more advanced lesson:

Even when AI points to a real source, verify that it is the right document, the right version and the authoritative basis for the conclusion.

And it adds one further lesson:

When credible evidence conflicts, do not stop at the first correction. Investigate the conflict itself and preserve uncertainty until the source trail can support a conclusion.

For our IFRS 18 tracker, Armah Sports should remain a special methodology case rather than a clean confirmed adopter. The tracker should preserve the Saudi Exchange announcement, the conflicting financial-statement evidence and the uncertainty surrounding which reporting basis ultimately governed the company’s official statements.

That may feel less satisfying than a simple answer.

But responsible technical research is not about deciding whether AI or the researcher “won.” Nor is it about forcing certainty from inconsistent evidence.

It is about making the uncertainty visible—and taking responsibility for what we publish.

Source notes

• Saudi Exchange: Armah Sports FY2025 annual-results announcement

• Armah Sports Q1 2026 interim financial statements

• Armah Sports H1 2026 interim financial statements

• Saudi CMA: framework for early application of IFRS 18 during 2026