A financial table can show how sales changed. A risk disclosure can describe a mechanism that those numbers do not capture: a licensing constraint, a potential privacy failure, or a new use of technology. How do you distinguish a meaningful addition from a routine rewrite?
Start with two original Apple annual reports: fiscal 2024 and fiscal 2025. We selected three pairs of passages from Item 1A, Risk Factors. Your task is to write a short evidence brief explaining one addition, one change in emphasis, and one warning that stayed the same.
First, challenge a claim
Consider this deliberately flawed sample analyst brief:
Apple first disclosed AI risks in 2025. The new language proves that AI caused licensing problems.
What would you need to check before accepting either sentence? Keep that question in mind as you read the selected passages. This is a constructed teaching claim, not a statement from Apple or the output of a model run.
Read the same discussion in both years
Open the 2024 Risk Factors source and 2025 Risk Factors source. Locate the heading about reliance on third-party intellectual property, on printed page 9. The download also includes the selected paragraphs for offline reading.
| Look for | Question to answer |
|---|---|
| The intellectual-property paragraph | What specific mechanism appears in 2025? How does the licensing language change? |
| The AI safety warning in the product-defect discussion | Is this warning new, or does the same sentence appear in both years? |
| The privacy paragraph, as an extension | Which change adds a risk mechanism, and which could simply be a terminology change? |
The anchors locate nearby source sections; use the supplied headings and paragraph prefixes to find the exact passages. Compare matching topics rather than assuming paragraph positions stay fixed.
Produce an evidence brief
For each finding, record the before passage, after passage, type of change, possible business implication, and remaining uncertainty. Include a link or locator for both years. A possible implication is a question for investigation, not proof that a loss occurred.
Compare your reasoning with the worked interpretation
New detail: the 2025 intellectual-property paragraph explicitly discusses AI training with copyrighted materials and potential reproduction of those materials in outputs. This identifies a specific mechanism worth investigating.
Changed emphasis: the 2024 paragraph says licenses can generally be obtained on commercially reasonable terms, but immediately qualifies that with uncertainty. The 2025 paragraph says the company is not always able to obtain all necessary licenses on such terms or at all. Preserve the earlier qualification: this comparison alone does not establish a newly occurring licensing failure or that AI caused one.
Unchanged control: the selected sentence warning that AI features may expose users to harmful or inaccurate content appears in both years. The claim that Apple first disclosed AI risks in 2025 is therefore unsupported. This is an unchanged sentence, not a claim that the whole paragraph or the company’s risk stayed unchanged.
Privacy extension: the 2025 paragraph adds a connection between AI features and inadvertent personal-data disclosure. Replacing the term “personal information” with “personal data” elsewhere in the paragraph does not by itself establish a new risk category.
Run a transparent comparison
The companion Python example contains six selected excerpts, source metadata, a comparison script, tests, and reviewed reference outputs. It runs offline with Python 3.9 or later and the standard library:
python3 compare.py --check --output-dir outputs
python3 -m unittest discover -s tests -v
Read the passages in data/sample.json alongside outputs/report.md. The program reports textual differences and equality. It does not decide whether a change is economically important; the worked interpretation above is a human-reviewed explanation.
Try adding whitespace to a scratch copy of a sentence. Then substitute a materially different word. Which changes should survive normalization? Keep the bundled source excerpts intact so their integrity checks remain useful.
Turn the task into an AI workflow
The current example supplies a baseline and a small evaluation case. A later model-based extension could:
- Retrieve matching passages and retain the filing identifiers and source locators.
- Propose structured findings: addition, changed emphasis, or unchanged language.
- Attach before-and-after evidence to every finding.
- Check whether the supposedly new idea already appeared in another selected passage or elsewhere in the earlier filing.
- Revise unsupported claims and stop with a qualified answer when the available evidence cannot settle them.
You can try those steps with an AI assistant and audit its answer against the sources. The downloadable program itself makes no model calls. Source selection and this worked answer are already visible, so this is a learning example rather than a hidden evaluation set.
The question is inspired by Lazy Prices: can changes in business language contain useful analytical information? This exercise identifies and interprets selected disclosure changes. Establishing that investors missed the information or that it predicts returns would require separate evidence and a carefully timed study.
What counts as a finished answer?
Finish when your brief identifies a supported addition, preserves the licensing qualification, recognizes the unchanged warning, and states what the filings cannot establish. More text or a more confident model is not a substitute for those checks.