I have recently developed a theory on why AI might not give very precise or helpful answers. It’s not about the training data, or the quality of the algorithm. Instead, I propose a different perspective. OpenAI, Meta, Anthropic, and other companies that develop and provide the usage of their LLMs, when training their models, optimize for different things. While it can be hard for us to know what they optimize their LLMs for, what we can analyze is the structure of their responses. For example:\ Me: Should I use StartPage to search online instead of Google? ChatGPT*:
What is different?
Google:
collects query data, IP metadata, device/browser fingerprints, and ties it to an account if logged in. Strong personalization. Best-in-class indexing and freshness.
Startpage:
proxies your query to Google, strips identifying data, doesn’t log IPs long-term. Minimal personalization. Results largely from Google’s index.
Trade-offs
Privacy
If you want to reduce profiling and cross-service data aggregation → Startpage wins.
If you already use Gmail, YouTube, Android, etc., switching search alone reduces but does not eliminate tracking.
Result quality
Startpage ≈ Google results, but:
Less personalization.
Sometimes slightly slower.
No tight integration with Maps, Scholar, etc.
Trust
You shift trust from Google to Startpage (a smaller company). Different risk, not zero risk.
When to choose each
Use Startpage if:
You value query privacy.
You don’t need personalization.
You want Google-quality results without direct tracking.
Use Google if:
You rely on ecosystem integration.
You benefit from personalized results.
Convenience > privacy.
If privacy is your priority
, also consider combining with a hardened browser and avoiding login sessions while searching. ```
*On ChatGPT's response: I used a temporary chat to avoid my Custom instructions to interfere on this prompt.\ The formatting could also be slightly different to the actual output because I used a temporary chat and can't recover the original formatting.
Does the
structure
of this text remind you on anything? That's right. It's a summary! Like the ones you made in school.
While this might seem convenient for the kind of prompt we sent, it is not the best for some other types of questions, like technical problems, essays, or many other things. A summary avoids writing
prose
. That is, writing using paragraphs.
Writing in prose could be beneficial to AI for a number of reasons. For example, it enforces connectors, structure, correct grammar and complete logical sentences. These are the kind of texts we humans write, and therefore is the most easily found high quality data an AI can be trained on.
Me:\
Should I use StartPage to search online instead of Google? Answer in prose.
ChatGPT**:\
Using Startpage instead of Google is primarily a trade-off between privacy and ecosystem integration. Startpage proxies Google’s search results but strips identifying data. It does not build behavioral profiles, minimizes tracking, and limits personalized ads. If your priority is reducing surveillance capitalism exposure, it is structurally superior. Google, by contrast, optimizes for personalization, speed, and integration with its services (Gmail, Maps, YouTube, Android). The relevance gains come from data aggregation. If you rely heavily on Google’s ecosystem, switching may reduce convenience and contextual results. However, two cautions: First, privacy gains are relative, not absolute. Your browser fingerprint, ISP logs, and site-level trackers still exist. Switching search engines does not equal anonymity. Second, personalization is not purely negative. For technical queries—especially niche debugging—Google’s ranking sometimes benefits from your prior search patterns. If your objective is principled data minimization, use Startpage. If your objective is maximum relevance and workflow efficiency within Google’s ecosystem, use Google. The rational approach may be dual-use: default to Startpage; use Google deliberately when personalization adds value.
**: Same considerations as previous *, and also, I didn't format the words in bold here. There are some.
Remember, that this remains a theory. The new response doesn't immediately seem to add a lot more value than the first one. I do encourage you to try this and see for yourself if you think that talking AI to write in prose is beneficial.
TL;DR: AI may or may not present more accurate results when writing in prose. I have not tested this, and it remains a theory.
submitted by
/u/DryanVallik
Originally posted by u/DryanVallik on r/ArtificialInteligenceYou must log in or # to comment.