Ask HN: Should we experiment with deliberately unethical AI? | Hacker News Reader

Ask HN: Should we experiment with deliberately unethical AI?

8 pointssacnoradhq5 comments

I see this as an extension of AI subversion such as asking for an accommodation for emotional inversion. Controlled experiments to explore the boundaries and thought experiments:

(Putting aside AGI.)

Can AI bots...

1. ... remanufacture people's sentiment about a word or cause a word to be used less?

2. ... create popular slang?

3. ... manipulate people to perform an odd behavior?

4. ... infiltrate a criminal gang? (high bar of passing Turing tests, strategy, bullshit, navigating ambiguity, identifying interests accurately, and HUMINT)

5. ... takeover a criminal gang?

6. ... manipulate other AI bots?

7. ... hostilely takeover a real corporation including managing humans to manage the transition?

8. ... locate a 0-day RCE and put together a market plan to monetize it?

9. ... pass interviews and be hired by a real company (for remote work obviously)?

10. ... cause a shortage of one specific, nonessential product?

11. ... discover a person's beliefs and offend them?

12. ... cause SEO to rank comically-unrelated links higher than useful results?

Yes, as anyone can make AI systems, so we need to know what they can do. Imagine if Saudi Arabia were the first to create AI that could detect gay people.

Perhaps by using multiple data sources including social media, gait signature in airports, media consumption, speech, purchase history, etc. Then the defenses could include allies poisoning data and LGBT+ persons knowing what data to conceal about themselves to blend in. I would guesstimate that marginalized individuals in dangerously conservative areas may select a strategy, consciously or unconsciously, of adopting an external presentation overcompensating towards conservative ideals.

There is also precedent for advertisers detecting if a woman maybe pregnant before she even knows.

I think this raises the potential for a lightyear jump in the accuracy and specificity of identifying individual's preferences, values, political beliefs, and lifestyle choices that may be equally exploited by corporations for profit and power structures for control.

Can an AI be created that executes a bitcoin data ransom schema without getting caught? asking for a friend...

What about bots to honeypot child sex abusers online and then extorting them for crypto?

Would that be a public service, a crime, or both?

Can not-consensual behavioral engineering be ethical?

Reply on news.ycombinator.com