Deepfakes are no longer easy to dismiss as awkward digital tricks. A recent experiment with AI video startup Synthesia produced a hyperrealistic avatar that looked and sounded like the person it was modeled on, including realistic intonation.
That result points to a larger problem for the age of AI: personal data is becoming more useful, more tradable, and harder to control. Faces, voices, expressions, posts, and old profiles can all become part of the raw material that powers artificial intelligence.
Deepfake Quality Is Moving Fast
The experiment took place earlier this month in a studio in East London, where Synthesia digitally cloned the writer using its AI video technology. The finished deepfake was convincing enough that someone who did not know the real person well could easily be fooled.
That matters because the technology has moved a long way from earlier AI avatars that were easier to spot. Synthesia has developed remarkably humanlike avatars after only one year of working with the latest generation of generative AI.
The immediate concern is not only that fake videos can look real. It is that the boundary between authentic content and synthetic content may soon become much harder for ordinary viewers to judge.
This risk is especially sharp because of the record number of elections happening around the world this year. When people cannot confidently tell what is real, the information environment becomes easier to manipulate.
The Trust Problem Cuts Both Ways
Deepfakes create an obvious danger: false content can be presented as real. But the source article points to another danger that may be just as damaging. If people become too skeptical, they may stop believing authentic material too.
Researchers have called this the "liar's dividend". The idea is that bad actors can benefit from a general collapse in trust by claiming real evidence is fake or AI-generated.
In practice, that means synthetic media can weaken trust even when no fake is directly involved. A person shown genuine incriminating information could try to dismiss it by saying it was made with AI.
The result is a difficult public information problem:
- Realistic deepfakes make fabricated content more persuasive.
- Public awareness of deepfakes makes real content easier to deny.
- Lower trust creates room for bad actors to exploit uncertainty.
The technology is exciting because it can create polished digital video with a humanlike presence. It is daunting because the same qualities that make the output impressive also make it harder to verify.
What Happens After AI Companies Collect Data?
The deepfake experiment raises a practical question: once someone gives an AI company their data, what happens next?
Synthesia says it does not sell data it collects from actors and customers. It does release some data for academic research purposes. The company uses avatars for three years, then asks actors whether they want to renew their contracts.
If actors renew, they return to the studio to make a new avatar. If they do not renew, Synthesia says it deletes their data. In the writer's case, Synthesia said it would delete the avatar after the experiment.
But the source article makes clear that not every company is equally transparent. Eileen Guo reported last year that companies such as Meta license actors' data, including faces and expressions, in ways that allow broad future use.
In those arrangements, actors may receive a small up-front fee. Their likeness can then be used to train AI models in perpetuity without their knowledge.
This is where the value of AI data becomes plain. A face is not just a face when it can help train a system. A voice, expression, or performance can become a long-term asset for a company.
Consent Gets Harder After Death
Even clear contracts have limits. Carl Öhman, an assistant professor at Uppsala University and author of The Afterlife of Data, says contracts for data do not apply if you die.
That creates a problem for the way AI systems treat consent. The current model often assumes the data subject and the company will both continue to exist indefinitely. Real life does not work that way.
Öhman says Facebook is projected to host, within the next couple of decades, a couple of billion dead profiles. Those profiles are not commercially useful in the same way living users are, because dead people do not click on ads, but the data still occupies server space.
That data may still have value. It could be used to train new AI models or to make inferences about descendants of deceased users.
Old social media platforms are another concern. When companies go out of business or lose popularity, their assets can be sold, and those assets may include user data. Öhman notes that MySpace data has been bought and sold multiple times since MySpace crashed, and says something similar may happen to Synthesia, X, or TikTok.
Why High-Quality Human Data Matters
AI language models are trained by scraping the web, and that can include personal data. A couple of years ago, the writer tested whether GPT-3, the predecessor of the language model powering ChatGPT, contained personal information. The system struggled with the writer's information, but personal information about MIT Technology Review's editor in chief, Mat Honan, could be retrieved.
The demand for high-quality, human-written data is growing because it is crucial to training the next generation of powerful AI models. The source article says the industry is on the verge of running out of free online training data.
That helps explain why AI companies are racing to make deals with news organizations and publishers. Their archives are valuable because they contain the kind of human-written material AI developers want.
Some people may not be worried about what happens to their personal data. But Öhman argues that exclusive access to high-quality data can help large corporations cement monopoly positions, which affects everyone.
The lesson from one convincing avatar is bigger than a single deepfake. In the age of AI, data is not just something people leave behind online. It is a resource that can be collected, licensed, sold, trained on, and kept alive in systems long after the original context has disappeared.