There is a new kind of recording in circulation that does not fit comfortably inside the old boxes. It is not a bootleg in the traditional sense, because there may be no lost studio session underneath it. It is not a cover, because the selling point is often that the singer sounds like the person who is gone. It is not a remix, because the words, melody and performance can be generated after the artist's death. And it is not quite a forgery in the ordinary sense when everyone involved openly labels it as artificial.
What it can be, however, is emotionally convincing.
That distinction is becoming important fast. The Guardian reported this week on an expanding ecosystem of AI-generated memorial content, including songs, images and conversational "digital ghosts" built around deceased people. Academic researchers have been working on the same territory under less lurid names such as digital afterlife technology, postmortem data and AI-mediated memorialization. The systems range from voice clones and avatars to chatbots trained on a dead person's messages. Music belongs in the same family, but it carries an additional weapon: memory already knows how to use a voice.
The first mistake is to ask a single question: "Is it authentic?" There are at least four different kinds of authenticity tangled together here, and AI can fail some of them while succeeding at others.
Authorship is not resemblance
The first kind is authorship authenticity. Did the deceased artist actually write, perform, approve or knowingly participate in the work? If the answer is no, then no amount of vocal resemblance changes that fact. A generated song does not become an unreleased track because a model reproduces a singer's rasp, cadence, ad-libs or melodic habits. A synthetic voice is not evidence of a hidden session.
The second kind is provenance authenticity. Can the listener tell where the recording came from, who made it, what tools were used and whether any estate, collaborator or rights holder authorized it? Provenance is an evidence problem rather than an aesthetic one. A clearly labeled fan experiment and an upload designed to masquerade as a newly discovered master may sound identical while being radically different acts.
Those two categories are where many arguments should end. A work either came from the artist or it did not. Its source chain is documented or it is not. Machines do not get to negotiate those facts by being persuasive.
But two other categories explain why the subject refuses to stay simple.
A fake performance can still be a good model of a style
The third category is stylistic plausibility. Does the generated performance behave like something the artist might have made? That is not the same as authorship. It is a modeling question. A system can learn recurring choices in phrasing, melodic contour, vocal texture, production, subject matter and song structure. A human producer can do the same thing by ear. The difference is that modern generative systems can compress those patterns and produce variations at industrial speed.
Sometimes the result is uncanny because it captures the surface. Sometimes it feels wrong because the surface is all it captures. The voice resembles the person but the writing lacks their odd turns of thought. The production copies a period but not the creative tension that made the original work interesting. The model knows what usually happens next and therefore produces exactly the thing the artist might have rejected for being too obvious.
That gap matters. Style is not a fingerprint. It is a distribution of tendencies. A plausible continuation can sound right without being something the person would have chosen.
The fourth authenticity belongs to the listener
The fourth category is emotional authenticity, and this is where the clean philosophical arguments tend to break down. A listener can know with complete certainty that a song is synthetic and still have a genuine reaction to it.
Research on music-evoked autobiographical memory gives a mechanism for why. Music is a potent cue for personal memories. In one controlled study of younger and older adults, familiar music associated with autobiographical memories produced stronger and longer-lasting sadness and negative affect than familiar music that did not evoke those memories. The important point is not that sad music mechanically makes people sad. It is that music can retrieve episodes from a person's own life, and those retrieved memories carry emotional weight into the present.
Now add a familiar dead voice.
The generated recording does not need to possess the dead person's consciousness. It only needs to activate a network the listener already built while that person was alive. A timbre can point toward an album. An ad-lib can point toward a car ride, a breakup, a jail cell, a bedroom, a funeral, a summer, a version of the listener who no longer exists. The model supplies a cue. The human supplies the history.
That is why "but it isn't really them" can be factually correct and emotionally beside the point. The listener's grief is not a claim about file provenance. It is an event happening in a nervous system.
This is not resurrection
Calling these systems resurrection technology makes them sound more mystical and more capable than they are. A voice clone does not recover a person. It does not restore the private experiences that shaped their judgment. It does not know which lyrics they would have refused, which jokes they would have cut, which producer they would have trusted, what they would have learned over the next decade or how aging would have changed the art.
A posthumous model freezes selected evidence from a life and extrapolates from it. Even a highly convincing system is therefore a branch generated by someone else's choices: which recordings were included, which model was used, which prompt was written, which outputs were discarded, which version was uploaded.
The phrase "digital ghost" is useful precisely because a ghost is not the person. It is a trace shaped by memory, expectation and the conditions of the haunting.
Researchers Giovanni Spitale and Federico Germani argue that AI afterlife systems should be designed around the ethical problems created by postmortem simulation rather than treated as ordinary consumer personalization. The consent problem is obvious: dead people cannot update their preferences when a new technology appears. An estate may hold legal authority without possessing perfect knowledge of what the person would have wanted. A fan may have sincere motives without having any authority at all.
Industry rules already reveal how unsettled the boundary is. Descript, for example, has publicly argued against unapproved cloning of deceased voices and says its own voice-cloning system is designed around cloning a user's own voice. Other tools and open models make imitation easier. The technical barrier is falling faster than social agreement can form.
The rights problem is not one problem either
Copyright, publicity rights, trademark, contract, platform policy and fraud are different systems. A generated song can avoid copying a protected recording while still creating an identity dispute. It can be labeled as AI and still use a person's likeness without permission. It can be authorized by an estate and still disturb fans. It can be an obvious parody and still circulate later without its original context.
New York expanded its postmortem publicity protections to cover digital replicas, while preserving exemptions for categories including expressive works and news. Other jurisdictions draw different lines. Litigation around AI music is also increasingly separating the rights in recordings and compositions from claims about identity and style. The result is not one grand legal answer. It is a stack of overlapping tests.
For platforms, provenance may ultimately matter as much as prohibition. A useful interface should tell a listener when the vocal performance is synthetic, identify who uploaded it, preserve the label when the audio is reposted and distinguish estate-authorized projects from unaffiliated simulations. That does not settle whether a work is ethical. It at least stops the system from manufacturing false evidence.
The dangerous feature is also the valuable one
It would be easy to treat the entire category as grotesque. Some of it is. Viral memorial sludge can reduce a complicated human life to engagement bait within hours of a death. Synthetic celebrity encounters can make grieving families watch strangers puppet a face they loved. A generated voice can also be used to manufacture statements the person never made.
But the opposite simplification is just as weak. Human beings have always used recordings, photographs, home video and old messages to maintain relationships with the dead. People replay a voicemail because the voice matters. They return to an album because memory has attached itself to it. Technology has always mediated mourning.
The new capability is generative rather than archival. The artifact can answer back. The song can be new. The dead voice can say a sentence that was never recorded. That is not merely more of the same. It changes the direction of the channel from retrieval to simulation.
The best design principle may therefore be brutally simple: never confuse comfort with evidence. A generated artifact can be allowed to carry emotional meaning while remaining visibly separate from the historical record. The system should preserve the sentence, "This was made after they died, by someone else, with a model."
Method, limits and falsification
This analysis separates authorship, provenance, stylistic plausibility and emotional response because collapsing them into one word, "authenticity," hides the mechanism. The evidence supports the claim that music can cue autobiographical memory and that posthumous simulation technologies are proliferating. It does not establish that a particular AI song will help or harm a particular grieving person, nor that synthetic memorials are inherently therapeutic or inherently damaging.
The framework would need revision if future studies show that clearly labeled synthetic voices reliably produce materially different memory and grief responses from archival recordings, or if long-term use of interactive posthumous systems has effects unlike passive music. Those are empirical questions. So are dependency, avoidance, benefit and harm. The responsible posture is measurement rather than moral theater.
Posthumous AI music should not be treated as recovered work from the dead, but dismissing the listener's response as fake is equally confused. The file can be synthetic while the memory it activates is old, personal and real. The practical line is provenance: label the simulation, preserve who made it, do not counterfeit authorship, and let people decide what emotional meaning they are willing to give the result.
Source trail
The Guardian, Sept. 15, 2026 — AI ghosts and digital mourning
Spitale & Germani, Ethics and Information Technology, 2026 — The making of digital ghosts
Mehl, Reschke-Hernandez & Hanson — Music-evoked autobiographical memories and affect
Descript — Thoughts on cloning voices of the deceased
Reuters, Feb. 5, 2026 — New York digital-replica and postmortem publicity legislation