AP: Avoid language that gives [AI] human characteristics

15 pointsposted 11 hours ago
by jaredwiener
(twitter.com)

8 Comments

skybrian

9 hours ago

Similarly, you shouldn't say "I flew to San Francisco" because it gives you bird characteristics. "I rode in an airplane that flew to San Francisco" is clearer.

:-)

add-sub-mul-div

8 hours ago

No one's confusing humans for birds. But dumb people are starting to think computers are alive.

qsera

7 hours ago

> dumb

They are not dumb. Just not initiated in the direction required to understand this. And the purpose of incessant marketing is to prevent them from doing so..

punchmesan

7 hours ago

Well, dumb-ness is subjective. On one hand, it's hard to call someone like Richard Dawkins "dumb" in a broad sense, but as a man of science he knows what it means to study and research something before drawing conclusions... And you do not need to be a computer scientist to understand the basic principles of an LLM. In that sense, Richard Dawkins is displaying a side of him that is almost willfully dumb when he's waxing poetic about how "Claudia" is sentient.

joegibbs

8 hours ago

If the models only talked in third person - never said “I” - I think there’d be about 80% less talk about AI consciousness

Quibblingeek

8 hours ago

Computers are not human.

Computers do not have sentient thoughts.

Computers should not misrepresent themselves as sentient creatures.

Computers are tools.

Seriously, I don’t understand why this is such a controversial take. Furthermore, it has already been demonstrated and disregarding this very basic fact has led to real-life harm. People have died over this.

naishoya

8 hours ago

THIS! All Day Long.

Some of the basic tenets which I have built into each of my AGENTS| CLAUDE|Rules-of-operation.md files include:

Software is prohibited from using personal pronouns or presenting any verbiage implying sentience: it is not allowed to use terms implying it 'thinks' and cannot assert that is making 'informed decisions' or 'reccomendations'.

By restricting any chat response phrases to the statistical probabilities, differential analysis, and requiring a disclosure of raw metrics and factors actually provides more actionable results with less bandwidth for outright manipulation or deception.

When working with a workflow management agent, there is no "we" cooperating on a task. Additionally restricting the agents from using engagement seeking or prompting for next steps makes a significant improvement to the token utility budget and also prevents some drift in direction of progress.

I understand the model host companies' reasons for the 'human seeming' interactions ~ and they are simply stated, pernicious and self serving. Asimov's three rules need expanding to a larger set, which also needs strong reinforcement at many levels. Failure to do this will come with greater harm as systems' capabilities continue to expand. This has been demonstrated and is provable.

> People have died over this.

theonemind

8 hours ago

The top comments there seem against the AP on this, which seems really strange. I wonder how deep the rabbit hole goes but not enough to dig up a twitter login