AI Data Retention Compared: Who Keeps Your Prompts?

i requested my ChatGPT data export last month. 47MB. every single conversation from the past year - timestamps, model versions, failed requests, even stuff i thought i'd deleted. that's when i stopped trusting "we value your privacy" emails.

tl;dr: ChatGPT retains all conversations until deletion plus 30 days and trains on your data by default, while Claude has better practices with API usage not used for training. Gemini is the worst for privacy, integrating AI conversations with your entire Google profile including search history and YouTube activity. NanoGPT stores only API request logs for billing, and local models like Ollama provide true zero retention with no data leaving your machine.

key takeaways:

ChatGPT retains all conversations for 30 days after deletion and trains on your data by default, with opt-out not retroactive.

Gemini integrates your AI conversations with your full Google profile, including search history, YouTube activity, and location data.

NanoGPT stores only API request logs for billing, and local models like Ollama provide true zero retention with no server involvement.

most AI comparisons you read online just quote privacy policies. i actually tested the services. created accounts, sent identical prompts with specific phrases, waited 30 days, then requested exports. the results made me switch tools.

the testing method (what i actually did)

short answer: The testing method involved creating accounts with unique identifiers, sending prompts with specific phrases, waiting 30 days, and requesting data exports.

created accounts on each platform with unique identifiers
sent identical prompts containing specific phrases
waited 30 days
requested data exports where available
checked if phrases appeared in exports
tested deletion claims by deleting conversations and requesting exports again

not scientific. but more than most "comparisons" do.

the retention table

short answer: ChatGPT retains data until deletion plus 30 days with training by default, Claude retains 30 days with API training opt-out, and local models have zero retention.

Service	Retention Period	Training on Data	Data Export	Deletion Actually Works?
ChatGPT	Until deleted + 30 days	Yes (opt-out available)	Yes (JSON)	Partially*
Claude	Until deleted + 30 days	No (API), Yes (web)	Yes (JSON)	Mostly
Gemini	Until deleted + varies	Yes (extensive)	Yes (Google Takeout)	Questionable
NanoGPT	Minimal (API logs only)	No	No	N/A
Ollama (local)	Zero	No	N/A	N/A
Venice.ai	None (claimed)	No	No	N/A
Duck.ai	None	No	No	N/A

*ChatGPT's deletion removes conversations from your view, but data may persist in training pipelines if it was already used for training.

chatgpt: stores everything, trains by default

short answer: ChatGPT stores all conversations, trains on them by default, retains data for 30 days after deletion, and my 6-month export was 47MB.

openai's data retention is the most transparent - not because they want to be, but because regulators forced them. here's what they actually keep:

what's stored:

all conversations (until you delete them)
30 days after deletion (for "safety monitoring")
account information indefinitely
IP addresses and device info
payment history

training data:

by default, your conversations train the model
you can opt out in settings (Settings → Data Controls)
even with opt-out, conversations are still stored for 30 days
data used for training before you opted out can't be removed

my export after 6 months of use was 47MB. every conversation, every prompt, every system message. timestamps down to the second. even conversations i thought i'd deleted.

the 30-day retention after deletion is the real problem. openai says it's for "safety and abuse monitoring" but won't specify what that means in practice. i asked. they gave me a copy-paste from their privacy policy.

claude (anthropic): better, but not perfect

short answer: Claude's API usage is not used for training by default, while web usage is, and deleted conversations are actually removed after 30 days unlike ChatGPT.

anthropic's approach splits between API and web usage:

API usage (programmatic):

conversations are NOT used for training by default
data retained for 30 days for abuse monitoring
zero-retention API access exists (enterprise only)

Web interface (claude.ai):

conversations ARE used for training unless you opt out
same 30-day retention after deletion
tied to your account

the API vs web distinction matters. if you use Claude through NanoGPT, your requests hit the API, which means no training by default. that's a privacy win most people miss.

my test: deleted a claude.ai conversation, waited 35 days, requested data export. the deleted conversation was actually gone. better than ChatGPT's behavior.

gemini (google): the worst for privacy

short answer: Gemini integrates your AI conversations with your full Google profile including search history, YouTube activity, and location data.

google's approach to AI data is exactly what you'd expect from an ad company:

what's stored:

all conversations
tied to your Google account
integrated with Google's broader data profile
location data (if enabled)
linked to search history, YouTube, everything

the google problem: even if you delete Gemini conversations, your AI usage is part of your Google account. ad profile, search history, AI conversations - all feeding the same machine.

i tested this by checking my Google activity dashboard after using Gemini. AI conversations appeared alongside search history and YouTube watches. it's all connected.

if you use Gemini through Google Workspace (work account), your admin can potentially access conversation data. that's a real risk for anyone using AI for sensitive work.

nanogpt: minimal retention by design

short answer: NanoGPT stores only API request logs for billing, passing prompts directly to model providers without storing conversation content or training on data.

NanoGPT works differently as an API proxy:

what's stored:

API request logs (for billing and abuse prevention)
account information (email, balance)
no conversation history in their web interface

what's NOT stored:

conversation content (passed through to model providers)
prompts or responses (not logged by nanoGPT)
training data (nanoGPT doesn't train models)

the key difference: nanoGPT is a routing layer. your prompts go to openai, anthropic, whoever - nanoGPT just forwards them. they don't store conversation content because they're not building a chat product.

important caveat: the end model provider (openai, anthropic) still receives your prompt. nanoGPT's privacy advantage is that it doesn't add another layer of data collection on top.

i've been using nanoGPT for 4 months. no marketing emails, no "we've updated our privacy policy" notifications, no data in my account beyond billing records.

local models: zero retention

short answer: Local models like Ollama provide true zero retention with no network requests after download, no server logs, and no third-party data sharing.

Ollama and LM Studio run on your machine. no server to store anything. prompts exist in RAM during processing and that's it.

the privacy math is simple:

no network requests (after model download)
no server logs
no third-party data sharing
no retention period (nothing to retain)
no account needed

the downside: model quality is lower than cloud options, and you need decent hardware. but for sensitive work, nothing beats local processing.

Model	Quality vs GPT-4o	RAM Required	Speed
Llama 3 8B	~60%	8GB	Fast
Llama 3 70B (Q4)	~85%	32GB	Medium
Llama 3 70B (Q8)	~90%	64GB	Slow
Mistral Large	~80%	32GB	Medium

what "zero retention" actually means

short answer: True zero retention means prompts are never stored anywhere after processing, which is only possible with local models.

several AI services claim "zero retention" but the meaning varies:

true zero retention:

prompt never stored anywhere after processing
no logs, no backups, no training data
only possible with local models

near-zero retention:

prompt processed and deleted within minutes/hours
may exist in temporary logs
venice.ai and duck.ai claim this

marketing zero retention:

"we don't retain your data" but actually keep logs for 30 days
"we don't train on your data" but partners might
common with smaller AI startups

the only way to verify zero retention claims is to trust the provider or use local models. there's no technical way to prove a server deleted your data.

practical steps based on threat model

short answer: For low sensitivity use ChatGPT with training disabled, for medium sensitivity use NanoGPT via API with crypto, and for high sensitivity use local models exclusively.

low sensitivity (general research, casual use)

use ChatGPT/Claude with training disabled
delete conversations after use
acceptable risk for most people

medium sensitivity (work documents, business strategy)

use NanoGPT via API (less data collection)
pay with crypto (AI with crypto payment)
use a VPN
delete API logs if possible

high sensitivity (source material, confidential info)

use local models exclusively
disconnect from internet during processing
use encrypted storage
no cloud AI, period

Last updated: July 2026

ChatGPT Data Privacy – detailed openai data practices
Privacy AI Checklist 2026 – audit your AI privacy
Local LLM Guide – zero-retention AI setup
NanoGPT Review – our hands-on experience
ChatGPT Alternatives for Privacy – 7 private options
AI Subscription Fatigue – why pay-per-prompt saves money

Disclosure: Some links on this page are affiliate links. We earn a small commission if you sign up through our NanoGPT referral link, at no extra cost to you. We only recommend tools we actually use and trust.