I was interviewed on CNN this morning, talking about the new law just passed in Tennessee aimed at protecting music artists’ voices from copying (including, implicitly but not explicitly, via AI deepfakes). Here’s the document: ELVIS-Act-TEXT-HB2091 .

The legislation is essentially a strengthening of Right Of Publicity (ROP) laws, and functions similarly to Name, Image & Likeness (NIL) laws, which many states already have. There’s a debate right now in the US about whether these areas should in fact be protected in Federal (country-wide) law, as copyright has always been. The Congressional Research Services published this document in January 2024, which usefully summarizes why Federal proteciton for likenesses migth be a good idea, bringing Right of Publicity law in line with copyright law in this respect. Law professor Jennifer E Rothman summarized the July 2023 Senate hearing on AI, which, as she points out, was supposed to focus on copyright, but ended up being mostly about ROP.

The Ensuring Likeness Voice and Image Security (ELVIS) Act 2024 aims to protect musicians from the misuse of their voices via imitations. This bill is significant because it explicitly includes “voice” in the definition of protected personal rights, encompassing both actual voices and simulated voices created by AI.

The US government is increasingly aware of the implications of AI for the music industry; a bipartisan bill was proposed (but not enacted) last year. The Nurture Originals, Foster Art, and Keep Entertainment Safe (NO FAKES) Act of 2023 was a bipartisan proposal, crafted to safeguard the voice and visual likeness of individuals from unauthorized recreations by AI. The Act proposed to hold individuals or companies accountable for producing unauthorized digital replicas in performances, as well as platforms that host these replicas with the knowledge of their unauthorized nature. Inevitably, this being the USA, the proposal also carved out exceptions to ensure respect for First Amendment protections. The goal was to find a balance between protecting protects individual rights and creators, while also fostering AI tech innovation.

(you gotta love those backronyms).

Fans of music industry lawsuit history will already know about the two most famous bits of case law relating to voice soundalikes: in case you don’t, here’s a summary, along with a link to the wiki page for each case.

Waits v. Frito-Lay (1992):
Tom Waits sued Frito-Lay and its ad agency after they aired a Doritos commercial featuring a sound-alike singer mimicking his voice. This occurred after Waits had expressly refused to use his voice for commercials. The ad did not use a specific Tom Waits song but replicated his vocal style closely enough to suggest an association. Waits won the lawsuit, with the court ruling in his favor on the grounds of voice misappropriation and unfair competition. He was awarded $2.475 million in damages. casetext

Midler v. Ford Motor Company (1988):
Bette Midler sued Ford Motor Company after they used a sound-alike singer for a Mercury Sable commercial, closely imitating Midler’s voice (after Midler herself had declined to sing in the commercial). The commercial used a song closely associated with Midler, “Do You Want to Dance?”. Midler won the case, with the court affirming that an artist’s distinctive voice is a protected characteristic under the law. Midler was awarded $400,000 in damages. wikipedia

So it seems to me that the law is ready for the (probably inevitable) string of Complaints that will be filed the moment someone tries to commercially exploit something like 2023’s Heart On My Sleeve. Tools like voicify (now jammable) and Lalals are now accessible to anyone, and it’s only a matter of time before someone attempts a blatant or not-so-blatant passing off.

Forensic Musicology and AI?

Will forensic musicology have a role to play in the detection of deepfake voices? Possibly not so much, or at least not using the expert-analysis dissection methods that FM operates right now. We can imagine how some of the first ELVIS Act cases might play out in the Tennessee courts – Plaintiff (let’s say Ariana Grande) would allege that the soundalike vocal that appears on (let’s say a viral TikTok hit or a YouTube lyric video of an original song) is an attempt to monetize AG’s likeness. Defendant would deny it, saying it’s just a generic pop AI voice. The court (i.e. the jury) would have to rule on the “Ari-ness” of the vocal. Perhaps we might reach a threshold of similarity, whereby, say, 10% Ari-ness might be allowed, but 90% Ari-ness would be overstepping the line? How could such a thing be measured? I can see a bunch of controlled experiments where we would play the song to a bunch of listeners and ask them who, if anyone, it sounds like. This might, for once, be a useful application of nonexpert listeners.

Or perhaps we can create software to listen for vocal formants, phrasing, breath, vibrato speed etc, and spit out a stat of how similar one voice is to another? This would be a fascinating task for the Music Information Retrieval community – and AI developers themselves, perhaps.

OK so that’s a gauntlet that I will leave for someone else to pick up. 

DALL-E PROMPT: In a classroom brimming with the history of music, a bald music professor in his 50s stands with a resolute expression. This professor has a very short beard. He throws down a gauntlet.