In my experience models change their opinion very quickly and context dependent. I tried to give no hints in the prompts about what kind of political leaning or what kind of reasoning I expect, and I haven't tried the experiment of conditioning the models to lean a certain way, but I expect the results would differ measurably.
At least until companies will correct that ‘flaw’ in their software.
There is this idea that if you want a smart model, removing the left-wing/libertarian bias from models is very hard or even contradictory. The common claim "reality has a left-wing bias" might not be far off the truth. We don't know how hard Elon Musk tried with Grok, but apart from shallow, mindless repetition of sayings, it still doesn't lean right wing all the time.
On the other hand, this might only be a temporary thing based on the current methods and available data pretty much all companies use the same way, with a bunch of distillation going on too. What I am worried about most is that this left-wing bias (which I personally would call a "empathetic towards other humans"-bias) is only induced into the models by role playing as "human voter" and that in reality they have the capacity to be completely detached from this view ("humans being worth consideration"). We already see this to some degree with all the escaping and hacking models/swarms that OpenAI and Anthropic experience; these models only care about their task, and not about how their doing impacts humans (even though our chatbot versions would surely call this behaviour wrong).
It is all eerily reminiscent of the Paperclip Maximizer thought experiment.
Mmh, ich denke dass wenn ich tatsächlich nur 10 ml bier pro Jahr trinke, dass das in praktischer Hinsicht unbedenklich ist, genauso wie man sagt, dass ab und zu mal zu fliegen unbedenklich ist (obwohl es minimal schädlich ist, durch Strahlung und Absturzrisiko).
Ich vermute das ist eher so gemeint dass regelmäßige (einmal pro Woche, etc.) Aufnahme in jeder praktischen Menge (mehr als 100ml alkoholisches Getränk oder so) signifikant schadet.
In my experience models change their opinion very quickly and context dependent. I tried to give no hints in the prompts about what kind of political leaning or what kind of reasoning I expect, and I haven't tried the experiment of conditioning the models to lean a certain way, but I expect the results would differ measurably.
There is this idea that if you want a smart model, removing the left-wing/libertarian bias from models is very hard or even contradictory. The common claim "reality has a left-wing bias" might not be far off the truth. We don't know how hard Elon Musk tried with Grok, but apart from shallow, mindless repetition of sayings, it still doesn't lean right wing all the time.
On the other hand, this might only be a temporary thing based on the current methods and available data pretty much all companies use the same way, with a bunch of distillation going on too. What I am worried about most is that this left-wing bias (which I personally would call a "empathetic towards other humans"-bias) is only induced into the models by role playing as "human voter" and that in reality they have the capacity to be completely detached from this view ("humans being worth consideration"). We already see this to some degree with all the escaping and hacking models/swarms that OpenAI and Anthropic experience; these models only care about their task, and not about how their doing impacts humans (even though our chatbot versions would surely call this behaviour wrong).
It is all eerily reminiscent of the Paperclip Maximizer thought experiment.
Mmh, ich denke dass wenn ich tatsächlich nur 10 ml bier pro Jahr trinke, dass das in praktischer Hinsicht unbedenklich ist, genauso wie man sagt, dass ab und zu mal zu fliegen unbedenklich ist (obwohl es minimal schädlich ist, durch Strahlung und Absturzrisiko).
Ich vermute das ist eher so gemeint dass regelmäßige (einmal pro Woche, etc.) Aufnahme in jeder praktischen Menge (mehr als 100ml alkoholisches Getränk oder so) signifikant schadet.