The trouble with API keys that never expire
A key issued years ago, never rotated, and still valid today is not a convenience. It is a liability nobody has actually looked at in years.
Why "AI-powered geocoding" claims deserve scrutiny
A geocoding feature described as AI-powered sounds more advanced than one described plainly as address matching. Whether that label reflects a meaningful technical difference or just a marketing decision made after the fact is usually impossible to tell from the outside, and that is exactly the problem. The label answers a question about vocabulary. It does not answer the question that actually matters to a developer deciding whether to trust the response: is the match correct, and how does it behave on the addresses that do not fit a clean pattern.
We are cautious about this kind of language for our own endpoints, and skeptical of it when we see it elsewhere. A machine learning component genuinely can help with fuzzy matching, handling typos, or resolving ambiguous partial addresses more gracefully than a rigid rule-based lookup would. That is a real, useful thing a system can do. It is a different claim entirely from asserting that the resulting matches are therefore more accurate across the board, which is a claim that needs actual evidence specific to the addresses and regions a customer cares about, not an inference from the word AI on a features page.
The core issue is that "AI-powered" describes a method, not an outcome. A method can be sophisticated and still produce a wrong answer, and a simpler method can be perfectly reliable for the cases it is built to handle. What a developer integrating a geocoding endpoint needs to know is the outcome: which address patterns are handled well, which are not, and what a low-confidence or ambiguous result looks like when it comes back. None of that is answered by the method label, however impressive it sounds.
We are being deliberately careful in what we claim about our own forward geocoding, reverse geocoding, and autocomplete endpoints for this exact reason. We can describe what they do and how they are structured. We are not going to assert a specific, verified real-world accuracy figure as if it were an established fact, because doing that convincingly requires exactly the kind of rigorous, ongoing evaluation that a label alone does not substitute for, whether or not the label includes the word AI.
None of this is an argument against machine learning techniques being used in geocoding. It is an argument against letting a technique become a substitute for a claim about results. If a method genuinely improves match quality, the way to demonstrate that is showing the improvement on real, representative cases, not naming the method and letting the reader assume the improvement follows automatically. A label is not evidence. It is, at best, a hint about where to look for evidence, and at worst, a way of skipping the part where evidence was supposed to come next.