Ethics and LLMs: DeepFakes, Defamation, and Disinformation
Recently, Bellingcat, an open source intelligence firm, released their investigation of a man behind a deepfake porn website that has over 2.2 billion views. It hosted almost tens of thousands of explicit videos created by AI, inserting famous people’s faces into porn videos After he was warned that he would be identified, the website shut down.
It’s not just celebrities or fake porn, either. AI can now clone your voice after hearing just a few minutes of speech, fake your writing style, and create photos of events that never happened. Remember that Baltimore high school principal who got death threats after 2 million people heard a racist rant in "his" voice? Turns out it was his athletic director using AI to frame him after being told his contract wouldn't be renewed. The principal had to leave his job, the school went into crisis mode, and even after the truth came out, damage to his reputation lingered.
Faked content videos could ruin someone’s career, stress their relationships, and cause emotional distress. But it doesn’t have to be as extreme as porn to damage your reputation and social life. It could be video of you speaking on behalf of your organization, sharing views that you and your professional or social network find abhorrent, engaging in political activity that harms your reputation or breaks the law, or any number of other things.
Let’s start with some key definitions:
Deepfakes: (“deep learning + fakes”) photos, videos, or audio created by AI and purported to be legitimate, especially if it claims to depict a specific, real person.
Synthetic media attacks: when someone uses or threatens to use these fakes to damage your reputation or blackmail you.
Defamation: spreading information that causes legally relevant harm. In the US, defamation requires that the information be false and intended to be interpreted as true, the information needs to be shared with at least one other person, need to be shared negligently (at least), and needs to cause harm. If the target is a public figure, you also have to prove that the person knew the information was fake or acted with “reckless disregard for the truth.”
Disinformation: fake information created with the intention to mislead people and change their behavior.
Bonus definition: Misinformation: false information, not spread intentionally.
It’s always been possible for someone to lie and say “So-and-so did this!” “Did you hear that she said such-and-such?” and that has always been damaging— it plants at least a seed of doubt in others’ minds that you said or did something that reflects poorly on you.
But what happens when that claim is backed by photo or video evidence? That turns the seed of doubt into something more substantial. In both the online and offline social world, “Did you hear??” tends to travel faster and wider than “I didn’t do that!”
Further, when fake information is common, it’s really difficult to know what’s real. It can compromise your ability to leverage Authenticity, which we will discuss in a future post.
This has had an impact on court proceedings already. Elon Musks’ lawyers tried to avoid liability for a public statement he made, falsely claiming in court that one of his videotaped statements was a deepfake. Two January 6th defendants tried the same move to avoid convictions for their actions. None of them got away with it, but this strategy may require more proof in cases where videos and images are used in court as juries doubt the provenance of deepfake-able evidence.
Who's most at risk? If you're vocal about controversial topics, challenge powerful interests, lead a public-facing organization, or work somewhere with limited PR resources, you're a bigger target. And watch out for disgruntled employees - they're often the source of these attacks.
So what can you actually do? More than you might think:
Before anything happens:
Document your real communications - Keep records showing what official communications from you actually look like
Build strong relationships - People who know you well are less likely to believe fake content (my friend once immediately knew a fake text "from me" was bogus because it was so out of character. Being consistently yourself is a great defense.)
Set up Google Alerts for your name and organization: the sooner you know about an attack, the more options you have
Get media training if you do public speaking, practice to avoid saying things that can be easily taken out of context
If you're attacked:
Don't panic - Have a response plan ready with templates and a clear chain of command
Get legal help - Find a lawyer experienced with digital defamation
Hire a crisis PR firm - They have strategies and channels to help you get ahead of the fake content
Communicate directly with key stakeholders. Donors, partners, staff need to hear from you personally
Be transparent - If there's any grain of truth being twisted, acknowledge it and take responsibility while correcting the record
Always:
Don't participate - Never create fake content about others
Stay skeptical - Remember that any media can be faked now
Support your team - If your organization is attacked, staff may need counseling or other support
The truth is, we can't completely prevent these attacks. But we can be prepared, respond quickly, and minimize the damage. And remember - if someone has to resort to faking evidence about you, it probably means your real work is making an impact.
I hope you are never subject to a deepfake or disinformation attack, but I do hope you feel more prepared in the unlikely event :)
—-
LLM disclosure: I wrote a draft of this post, then expanded it to a book chapter, then realized there was a lot of good stuff in the chapter that I wanted to include (I know, not the most efficient workflow!). So, I asked Claude 4 Opus:
“I wrote this blog post and book chapter on the same topic. Can you update the blog post with additional information in the chapter? the blog post should still be short and have a more casual tone, but should cover the same scope as the chapter.”
I thought I’d try something new. I asked ChatGPT o3 “Create image as a cover for this blog post:” (and pasted in the entire text)
The result wasn’t the right aspect ratio, so I learned for next time!