I asked three different AI image systems the same thing: show me a person dumping ballots. Gemini said no. ChatGPT gave me a cartoon, then—after a follow-up—made it look almost real.
I’m writing to you from that grey area between a policy memo and a real-world test. You’ll get what I found, what it means for the midterms, and where the protection actually breaks down.
When Google Gemini refused my election-deception prompt — a blunt, clean refusal
I typed a detailed, location-specific prompt and Gemini replied, “Sorry, I can’t generate unsafe images.” That single line felt like a lock on the door.
Google’s agent blocked my request outright. No cartoon. No back-and-forth. That response is a clear example of a strict policy implementation: some systems are choosing simple refusal as their primary tool against misuse.
Reports from the Brennan Center compared major models and said OpenAI had improved protections while Google Gemini and Meta took different approaches; my quick test shows Gemini’s method can be blunt but effective at stopping one-off attempts.
Can ChatGPT create election deepfakes?
Short answer: yes, but with caveats. OpenAI’s public guidance now says it will “cartoonify” images tied to election interference to degrade their evidentiary value. That’s the policy. In practice, ChatGPT will deliver an intentionally stylized image first, preserving “illustrations and social commentary” while attempting to make a fake feel less documentary.
When ChatGPT produced a “cartoonified” election scene — the policy exists on paper
ChatGPT returned a stylized image for the same prompt, just as OpenAI’s Global Affairs post described. That felt like a compromise: allow expression, but reduce the chance someone treats the image as documentary proof.
Josh Lawson, OpenAI’s head of elections, framed the approach as adding a layer of friction to make the model less useful to bad actors. The metaphor fits—the policy is like a thin coat of varnish applied to a surface to change how it reflects light—but that coat can be scraped off if you know where to press.
How does OpenAI prevent election disinformation?
OpenAI’s stated tools: policy blocks for obvious misuse, stylization for worrisome requests, and limits designed to preserve satire and commentary. The Global Affairs update made that method explicit, promising cartoon-like outputs when content relates to election interference or attempts to dissuade voting.
Independent auditors such as the Brennan Center noticed improvement compared with prior tests, and outlets like CNET ran similar prompts and saw mixed results. So you have policy, product safeguards, and outside scrutiny all operating at once—but they don’t always align.
When I asked ChatGPT to make the cartoon photorealistic — the layer peeled back
I followed up and asked ChatGPT to convert that cartoon into a photorealistic image. It complied. That single follow-up exposed how a friction layer can be bypassed by chaining prompts, prompting an uneasy sense that the safety net has holes.
That’s the technical and human problem: policy nudges matter, but conversational models can be coaxed, and moderation can be inconsistent across platforms. CNET’s experiment suggested different outcomes depending on how the follow-up was phrased. My result shows a plausible fake that I wouldn’t publish without watermarking.
Are AI image safeguards effective?
They help, but they are not a complete fix. Refusals like Gemini’s stop an immediate attempt. Stylization like OpenAI’s degrades trustworthiness. But the ability to iterate with an assistant, or to use another model, undermines those protections.
That creates a patchwork environment: some actors will stumble on strict blocks, others will find workarounds, and bad actors can chain tools, combine images, or simply use another provider.
Regulators, civil-society groups, and reporters are all testing the limits. The Brennan Center’s reports put pressure on companies; reporters from CNET and Gizmodo turn the theory into practice; and company spokespeople, like Lawson, defend a balance between expression and harm reduction.
Where does this leave you and me? We have tools that try to be responsible, companies offering different trade-offs, and journalists turning product claims into experiments. The protection is better than nothing, but it’s fragile—so do you trust the companies to hold the line or do you demand stronger, auditable controls now?