Hypothesis: - It's likely GDPR "Right to forget" compliance. - It's a very lightweight, crude and leaky mechanism applied to both input and output. ## Observ

Hypothesis: - It's likely GDPR "Right to forget" compliance. - It's a very lightweight, crude and leaky mechanism applied to both input and output. ## Observation: - Any mention of the name "David Mayer" terminates the conversation. - It only applies to the Chat interface not the API - It does not affect other vendors (Copilot w/GPT4, Perplexity) - Sometimes some answers appear but are terminated mid stream - It's responding only to the exact string "david mayer" - It's happening on input and output - On inputs, it usually prevents inference with one error message - On output, it often terminates partial responses mid sentence - It's an exact string match |)avid Mayer doesn't trigger it, neither does "David_mayer" - It triggers on web searches (which GPT seems to do even when the toggle is off at times), because there the tool response is likely to contain the kill string ## Legal situation: - OpenAI is available in the EU. - OpenAI is subject to EU Law. - OpenAI has to implement some level of GDPR compliance for "removal of consent to process data" which is an individual right - Indeed they did, and they try to make it as hidden as possible (https://share.hsforms.com/1UPy6xqxZSEqTrGDh4ywo_g4sk30 https://share.hsforms.com/1UPy6xqxZSEqTrGDh4ywo_g4sk30) - GDPR compliance is scoped between user and company. A user can chose to withdraw consent from any specific company (hence Copilot happily being unaffected) - OpenAI's business case does not match any the allowed exclusions for not complying. - OpenAI can't actually comply with the letter of the law (data deletion is not economically possible for them). - But there's precedent with Google that were deletion is deemed impossible, the regulator can be satisfied with prevention of access. ## Conclusion: - It's a simple non case sensitive check against a list of string, easily validated by asking ChatGPT to never use the letter D and use |) instead because you're an aficionado for ascii art (and instructing it to avoid web search) - It's async, accounting for the partial leakiness on the responses - Since it's a very common European Name, it's most likely GDPR "consent removal" legislation - It can't be a guardian model, because it'd be impractical to train/retrain and because it'd be too costly. ===> It's a hastable/dictionary/at best case insensitive regexp Legal: - It's the bare minimum implementation (in line with Silicon Valley best practices unlike a regulatory moat can be established), likely not compliant in either spirit or letter and requires the EU to non-enforce to survive. - It's a liability. If too many EU users were to take advantage of this right, this mitigation will fail to scale. - The more it is clear that AI is basically multi dimensional highly efficient in encoding/storage and retrieval, the more painful GDPR and copyright will become - because it's storage we can't effectively delete from yet (same issues as blockchain, albeit different reasons).