Apple's annual reports already discussed risks from artificial intelligence in 2024. What became more specific in 2025, and what stayed the same? Compare selected passages from the original FY2024 and FY2025 10-K filings before accepting an answer.
This is a small, deterministic Python example by Yeabin Moon. It connects narrative comparison, evidence, and verification to later AI interpretation and model routing. No language model runs in this program.
Start with a prediction
Before reading the interpretation below, inspect the paired excerpts in data/sample.json. Identify one changed passage and one unchanged statement. Would a count of AI mentions establish whether Apple introduced a new risk? What would you need to check before saying so?
Run the comparison
Download and extract this project folder, open a terminal in it, and run:
python3 compare.py --check --output-dir outputs
python3 -m unittest discover -s tests -v
Python 3.9 or later and its standard library are sufficient. Use python instead of python3 if that is your environment's command. No API key, network access, extra package, or full filing download is needed.
Read outputs/report.md alongside the complete selected passages in data/sample.json. outputs/result.json also contains word-level differences. --check compares the computation with the bundled reviewed result; it does not independently verify the company's statements.
The program validates excerpt hashes and filing-identity consistency, aligns the selected pairs, and reports sentence and word changes. It does not automatically label a risk as new, stronger, resolved, or financially material. The sentence splitter is scoped to these reviewed excerpts, not arbitrary documents.
A reviewed interpretation
These observations come from reading the paired evidence, rather than from an automatic semantic classifier:
- More specific intellectual-property exposure: the 2025 IP paragraph explicitly connects machine learning and AI to copyrighted training materials and potential reproduction in outputs. That mechanism is absent from the corresponding 2024 paragraph. Its addition does not prove infringement occurred.
- Changed licensing emphasis: 2024 says licenses can generally be obtained on commercially reasonable terms, followed by a no-assurance qualification. In 2025, Apple says it is not always able to obtain all necessary licenses on those terms or at all. Preserve the earlier qualification; this wording alone does not establish a newly experienced licensing failure.
- An unchanged AI warning: both filings contain the same selected sentence about AI exposing users to harmful or inaccurate content. This contradicts the claim that Apple first disclosed AI risks in 2025. It establishes equality of that sentence, not of the full product-defect paragraph or all risks.
- Optional privacy extension: the selected 2025 privacy paragraph adds an explicit AI-related inadvertent-disclosure mechanism. The parallel change from “personal information” to “personal data” should not automatically be counted as a new risk category.
The IP paragraph and unchanged warning provide the core exercise; the privacy pair is an optional extension. A defensible short answer includes the observation, both source IDs, supporting language, and what remains uncertain.
Try one controlled change
from compare import compare_text
print(compare_text("AI can create risks.", "AI can create\nrisks."))
print(compare_text("Licenses can be obtained.", "Licenses cannot be obtained."))
Predict which comparison changes only representation and which changes meaning. The program locates textual changes; you assess their significance. Keep the original packet intact. Deliberately editing a fixture requires a new recorded hash and a separately reviewed expectation; do not change the reference merely to make a failed check pass.
Scope and sources
These are selected annual disclosure passages, not a complete risk inventory or investment signal. A new sentence may express a pre-existing risk; matching and interpreting sections requires more than keyword counts. The example makes no claim about stock returns or actual changes in risk probability.
Sources and data dictionary identify the original filings, local source hashes, anchors, selections, and limitations. This public worked example exposes findings. An independent classroom assessment needs a different reserved case if the learner has already seen these answers.