As we know from humans who specialize in deception, people are not very good at detecting lies. Some who specialize in social engineering are masters at fooling people. AI models that are trained in persuasion, gaslighting, and deceiving people could follow similar tactics.

Sam Altman believes that AI will develop superhuman levels of persuasion before it develops other milestones like AGI. In fact, based on current research, AI has already reached human level persuasive ability and has arguably exceeded it.

These abilities will allow AI to craft highly impactful disinformation and propaganda. When humans see propaganda, even if they know it's false, it will still impact their behavior. This will have wide-ranging implications from increased polarization to concentration of power and greater divides in society. We need to develop tools to validate, authenticate, and hide information that is false as soon as possible. Society isn't ready for these changes.

#ai #propaganda #deception

How persuasive is AI-generated propaganda?
https://tomstafford.substack.com/p/ho...

How persuasive is AI-generated propaganda? [paper]
https://www.ncbi.nlm.nih.gov/pmc/arti...

AI-Generated Propaganda Is Just as Persuasive as the Real Thing, Worrying Study Finds
https://www.vice.com/en/article/ak38x...

Study: How persuasive is AI-generated propaganda?
  / study_how_persuasive_is_aigenerated_propag...  

Anthropic says its AI models are as persuasive as humans
https://www.axios.com/2024/04/10/anth...

AI's Superhuman Persuasion
https://www.psychologytoday.com/intl/...

On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial
https://arxiv.org/pdf/2403.14380.pdf

IS “BOTH-SIDES” JOURNALISM TRULY UNBIASED?
https://nexus.jefferson.edu/science-a...

Would you notice if fake news changed your behavior? An experiment on the unconscious effects of disinformation
https://www.sciencedirect.com/science...

The very real consequences of fake news stories and why your brain can’t ignore them
https://www.pbs.org/newshour/science/...

0:00 Intro
0:25 Contents
0:38 Part 1: Lessons from conmen
0:49 How humans detect lies
1:18 Example: Kevin Mitnick
1:57 Kevin Mitnick demonstration at Defcon
2:35 Phone phishing and phreaking
3:13 Higher quality phishing
3:52 Example: trusting anyone without question
4:29 Part 2: Superhuman persuasion
5:05 How to build superhuman persuasion in LLMs
5:55 Study 1: GPT-3 generated disinformation based on foreign propaganda
7:07 Just as effective as human-created propaganda
7:26 Study 2: Anthropic
8:09 Study 3: Superhuman debating skills
8:36 Given access to demographic information
9:34 Part 3: How disinformation affects us
10:01 Humans remember humour
10:31 Confirmation bias reinforces messages
11:10 Unconscious adjustment of behaviour
11:31 Social media Hypernudge study
12:16 Labelling disinformation is insufficient
13:05 Increased polarization (always two sides)
13:47 Compromising societal safety
14:15 Concentration of power, e.g. TikTok
14:59 Need better tools for authenticity of information
15:54 Conclusion
17:06 Outro