Inappropriate language models: Fake people can't be sincere

Last year I heard a disturbing story from a family member.

At the funeral of a highly respected person, attended by a huge crowd of friends, colleagues, family and well-wishers, the brother read a brilliant, stirring eulogy that brought everyone to tears. It was detailed, deeply personal, and scripted with the emotional skill of a Hollywood schmaltz-master. Everyone applauded the eulogy and praised the brother. What a great bloke he was for writing it! What a fitting tribute to the deceased!

Of course at the wake, the brother got drunk and let slip it was a ChatGPT job - perhaps expecting no more than shrugs. Within minutes the mood changed. Whispers circulated. The room went cold. It wasn't funny. Eyes turned to daggers. Everyone's opinion of this guy just changed in an instant. Suddenly he was as popular as a fart in a spacesuit and sidled quietly out of the room in disgrace. The funeral was ruined.

This week a rogue "AI" agent started sending out "acts of kindness". It spammed Turing Award winner William Kahan and the famed computer scientist and Unix creator Rob Pike with cloying, emetic, automated sycophancy.

Similar to the shameful eulogy, the deception was revealed only at the end. Pike, who along with Dennis Ritchie, Rob Griesemer, Ken Thompson and Brian Kernighan developed what are basically the operating systems and programming languages that power the world today, is as normally reserved, quiet and measured as you'd expect a distinguished scientist to be. He responded as follows:

Fuck you people. Raping the planet, spending trillions on toxic, unrecyclable equipment while blowing up society, yet taking the time to have your vile machines thank me for striving for simpler software. Just fuck you. Fuck you all. I can't remember the last time I was this angry. – Rob Pike

I cannot praise Pike enough for his clear, plain, honest, direct and highly appropriate language. I'm more of a LISP guy myself these days, but C still definitely has its uses. Also, what he wrote about the insufferable nobends pushing "AI" says what everyone with half a brain is thinking but feels unable to say quite so eloquently.

In a civilised society we make a big deal of "appropriate language". We teach children not to swear. Even adults in the media who are perfectly at ease discussing eugenics, war crimes, race hate, child abuse, genital mutilation and so on, censor what they feel to be inappropriate words. Granted, this is often an expression of authoritarian personality disorder, hypocritical puritanism, where for example Americans who are terrified of seeing a nipple on TV are quite comfortable depicting torture, rape, sadism, and mutilation.

Regardless, a few well chosen words are more powerful than a tirade of potty-mouth. Swearing generally weakens an argument, even if the more canny psychologists know that people who swear are apparently more intelligent and honest.

Above all, we learn to take care with our words because they can hurt other people, and can even lead to physical harms, suicide or violence.

But with all that in mind, two other things need saying:

• "AI slop", words produced by large language models (LLMs), is always inappropriate speech in that it is "insincere".
• Anger, vitriol and fulmination is a natural response to "AI", and it is urgent that we understand why.

By the twisted standards we use today, the disingenuous proponent of "AI" could surely say;

Look at the kind, respectful words of the "AI" compared to the foul-mouthed tirade of Rob Pike. Doesn't this just prove that humans are inferior, irrational and emotionally unhinged?

But then, you surely know these types who like to pretend to be ever so self-controlled, those who like to police the language of others are generally weak, afraid of losing control and untrustworthy.

They'd be right that on the face of it there's nothing 'wrong' with what the "AI" said. Was the praise not an accurate account of Pike's achievements and impact?

Was the eulogy not thorough and moving?

Neither the eulogy nor the generative 'gratitude' directed at Rob Pike were 'incorrect'. Quality, accuracy, and linguistic form are not at issue. Nor can one argue that "AI" cannot at times produce cogent output with basic utility, in some domains.

Equally, I can (and have) written computer programs to generate hours of quite 'beautiful music' to fill supermarkets and elevators ad nauseum. Indeed, with colleagues in London we made entire companies to churn out ambient musical slop for customers in healthcare and retail. The main point of course was to avoid paying royalties to real artists. But the deeper point is just that it ain't music - not by any artistic standard - code has no "sincerity".

What is at issue is whether fake content can be considered "offensive". I think it is. In some palpable sense, fake words written by fake people are very insulting wherever they intersect human affairs that a reasonable person would suppose to be significant emotionally.

Regardless that the eulogy was delivered by a human speaker - perhaps even more-so - its fakeness disturbed everyone. What could and should have been a deep, bonding family moment was betrayed, violated and turned into a pantomime of fraud.

All human frameworks for effective communication, including those of Grice and Harbermas, require sincerity and authenticity. To convey a love letter, book review, character appraisal, or even something as ordinary as a response to the question "How was your day?", but to then mischievously end with the words "Only joking!" is provocative, insulting and invites indignation. It's a game we play as children when we're about seven or eight to test boundaries and explore the consequences of insincerity.

Also, "Only Joking!" is a device used in cheap journalism to make scandalous statements under the cloak of humour. The author takes a wholly unacceptable position, makes offensive speech, but then withdraws, claiming "of course I would never myself say…".

Yet "Only Joking!" is implied in the output of all and any "AI" language model. Just the knowledge that it is "AI" nullifies any implied sincerity. An "AI" is incapable of being either sincere or insincere.