Pennsylvania's AI label law will teach bad actors new vocabulary
Pennsylvania Governor Josh Shapiro just signed a law requiring labels on any advertisement "significantly modified or generated by artificial intelligence" — and in doing so, the state has chosen the policy intervention most likely to fail. The law sounds protective: disclose the AI, let consumers decide. What it actually does is require companies to adopt a new vocabulary for the same work they were already doing.
I want to believe disclosure works. There is something appealing about the idea that transparency alone protects people — that once you tell someone a thing is happening, they'll make a better choice. Twelve years ago, Vermont believed this too. The state mandated GMO labels in 2014, convinced that consumers armed with information would reward manufacturers for transparency or choose differently. What happened instead was immediate: companies reformulated products to avoid the labeling trigger, rebranded their messaging toward "natural" and "clean," and pivoted their marketing spend into the exact anxiety the labels were supposed to reduce. By the time federal preemption arrived in 2016, consumer behavior had barely shifted. People still bought the foods. The labels had simply redistributed market advantage toward companies sophisticated enough to navigate around them.
Pennsylvania is now running the same experiment with "significantly modified." That word is doing all the work and carrying none of the definition. Is a photograph with an AI-adjusted background significantly modified? A voice that sounds like a real person but was synthesized? A political advertisement where the candidate's position was real but the lighting was digitally enhanced for emotional effect? The law does not say. This is not an oversight. This is an invitation. Companies will argue about what "significantly" means in the same way food manufacturers argued about what "natural" meant, with the same result: disclosure labels that exist to satisfy lawmakers while leaving consumers exactly as confused, just with official permission to be confused.
What makes this different from, say, nutrition labels, which work. Is that nutrition labels describe a physical fact. A food contains 8 grams of protein. This is checkable. AI modification is process, not product, and processes can be renamed. A restaurant's ugly crowded flyer about weekend specials can now carry a label saying "AI generated" while a political campaign's perfectly lit candidate video simply calls itself "digitally enhanced" and claims exemption. Both are AI-modified. One will be labeled. The other won't, because the threshold was never a technical standard. It was a negotiation waiting to happen.
If you're building something and you want to protect people from a specific harm, you need to name that harm precisely enough that bad actors cannot work around it by redefining their language. Pennsylvania has done the opposite. The law will produce compliance theater: labels on some ads, semantic reshuffling in others. A decade from now, a research paper showing that disclosure didn't change what people believed or bought. The question isn't whether the labels will work. History already answered that. The question is why politicians keep reaching for the same broken tool. The answer is simpler than policy debates usually admit: labels look like action without requiring anyone to do anything difficult.
