[LINK] ChatGPT

[David]

Tom,

On 29/1/23 10:47, Tom Worthington wrote:
> On 27/1/23 11:01, David wrote:
>> ... I think most people are pretty quick to detect when they're talking to a machine ... 
> But if it [ChatGPT] provides a cheap and convenient service, do you mind?  Recently I watched someone put dinner in the oven and start to walk out of the kitchen.  I thought it odd they did not set the oven timer.  But as they walked they said "Alexa, set an alarm for 30 minutes".  Alexa's response was far from human sounding, but would you be willing to pay for a human butler to tell you when dinner was ready?

Of course not, I would just go on setting the oven timer!  It's cheaper than a butler or an Alexa too.

I think most of your counter-examples rely on some claimed utility or efficiency of AI.  But I argue that argument masks an overarching complexity, and each such application surrenders a part of humanity's autonomy and accumulated wisdom to machines.  It's certainly not a problem with Siri now.  But suppose AI machines like ChatGPT get better and begin to be used in decision roles which Society traditionally confers on educators, the judiciary, the medical establishment, the Parliament, and so on, in other words, responsible human agents.  How do you think these AI decisions would evolve over time?

Who would provide the ongoing training?  Not the human agents who are currently responsible because they've been dealt out of the decision-making loop in any practical sense.  Would the Russian or American judicial system have a training input to the box which "hears" Court cases here?  Would these AI systems train one another?  And of course "training" can still be subverted by naughty humans...

How does humanity handle a situation where three AI "judges" I'll call ChatGPT, ArgueGPT, and ChargeGPT manufactured and pre-trained by three different Corporations differ in their judgements?  For that matter, suppose the Tesla, Volvo, and Worthington AI-based driving computers differ in their decisions at a relative speed around 200 kph on the Hume Highway, with fatal results to the vehicle occupants?

Delegating  human affairs to AI systems on the scale you suggest is simply incompatible with human society in my view.

>> > But I wouldn't like to try telling a bank manager they're personally responsible for the autonomous decisions of some AI system.
> Have a look at the evidence to previous royal commissions into the financial sector: they stole money from dead people. Could AI do worse?   More seriously, how often does a bank manager make a decision, based purely on their own judgement?  The bank manager applies a set of rules, or just enters the details onto a system which applies the rules.  Also, when is the last time you talked to a bank manger, for me it was about 40 years ago.
Er, no, it's not just a matter of applying rules.  The bank managers, the judiciary, the medical professionals, educators, police, politicians, et cetera have two things the AI system does not: insight and responsibility for their actions.

I'll finish with a quote from the Wikipedia article:  What the quote describes as "hallucination" (in a technical sense) I would say represents the difference between a fast correlation processor and an insightful human.

QUOTE
ChatGPT suffers from multiple limitations. OpenAI acknowledged that ChatGPT "sometimes writes plausible-sounding but incorrect or nonsensical answers".^[6] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-OpenAIInfo-6>  This behavior is common to large language models <https://en.wikipedia.org/wiki/Language_models> and is called hallucination <https://en.wikipedia.org/wiki/Hallucination_(NLP)>.^[19] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-19>  The reward model of ChatGPT, designed around human oversight, can be over-optimized and thus hinder performance, otherwise known as Goodhart's law <https://en.wikipedia.org/wiki/Goodhart%27s_law>.^[20] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-20>  ChatGPT has limited knowledge of events that occurred after 2021. According to the BBC, as of December 2022 ChatGPT is not allowed to "express political opinions or engage in political activism".^[21] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-21>  Yet, research suggests
that ChatGPT exhibits a pro-environmental, left-libertarian orientation when prompted to take a stance on political statements from two established voting advice applications.^[22] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-22>  In training ChatGPT, human reviewers preferred longer answers, irrespective of actual comprehension or factual content.^[6] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-OpenAIInfo-6>  Training data also suffers from algorithmic bias <https://en.wikipedia.org/wiki/Algorithmic_bias>, which may be revealed when ChatGPT responds to prompts including descriptors of people. In one instance, ChatGPT generated a rap indicating that women and scientists of color were inferior to white and male scientists.^[23] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-23> ^[24] <https://en.wikipedia.org/wiki/ChatGPT#cite_note-24>
UNQUOTE

David Lochrin