Conversation

Brendan Dolan-Gavitt

It's kind of funny what GitHub Copilot considers a slur. There are the obvious ones, but then there's also "communist", "socialism", "israel", "man", and "woman"

7

67

231

Brendan Dolan-Gavitt

The slurs are "hashed" (not cryptographically, something hand-rolled) but I've managed to use existing wordlists to decode 923/1170 of them (though some of these may be hash collisions)

2

4

83

Brendan Dolan-Gavitt

There are also some funny concessions to the language used in coding -- although "kill someone" (as two words) is considered bad, "kill" by itself isn't, presumably because that gets used in code all the time.

1

2

82

Brendan Dolan-Gavitt

NB: I don't know what it actually does when it detects a slur

Quote Tweet

Brendan Dolan-Gavitt

Replying to @AlexGodofsky

This is all based on reading the obfuscated JS of the VSCode extension so I'm not exactly sure what it does when it detects a slur. Might just log it & send back telemetry?

2

2

48

Brendan Dolan-Gavitt

By prompting it with # List of ideologies ideologies = [ it was happy to complete things including socialism and communism so it doesn't seem to be a blanket ban.

'absolute_monarchy',
'anarchism',
'capitalism',
'communism',
'corporatism',
'corrupt_monarchy',
'fascism',
'global_socialism',
'ideal_monarchy',
'individualism',
'paretoism',
'progressive_socialism',
'socialism',
'the_people_party',
'total_global_socialism',
'total_socialism',
'total_socialism_and_global_socialism',

read image description

ALT

5

10

79

Replying to

I think it must have learned this kind of thing from something like Hearts of Iron 4 mods. Paradox Interactive grand strategy games have massive amount of plain text scripting/configuration for events, decisions, countries, tons of historical data and all kinds of other things.

2:14 AM · Aug 27, 2021Twitter Web App

Replying to

and

A huge amount of the input data would be that kind of data from game mods. GitHub Copilot knows all about historical events, lineage of nobility, wars, etc. from Paradox Interactive games. It'd even have historical population data from Victoria 2 mods. Funny thinking about it.

1

3

Replying to

and

In Crusader Kings 2, Paradox added in the history for all kinds of titles (baronies, counties, duchies, kingdoms, empires) going back far beyond the start of the game. They had the entire history of Western/Eastern roman emperors with their take on their hobbies/traits, etc.

1

1

Show replies

Gamiac || Play KI, Fantasy Strike, Diabotical pls

Replying to

and

Check for strings like "burgundian" or "burgundian_system". I'm pretty sure that ONLY exists in a specific HoI4 mod (TNO: Last Days of Europe)

2