In 2023, Sam Altman predicted that society would learn to question synthetic videos and develop ways to verify their origins. Watermarking and content-provenance tools are advancing. But a more fundamental problem remains: verifying how something was made is not the same as verifying whether it is true.
- The first prediction: society would learn to question synthetic media
- Watermarks and digital credentials: the infrastructure is taking shape
- The trust gap: a verified origin does not guarantee a true story
- The harder problem Altman raised: personalised AI persuasion
- What this means for India
- The verdict: Altman identified the right direction, but the problem is larger than detection
- AI, Revisited: The series
In June 2023, Sam Altman sat down with Satyan Gajwani, Vice Chairman of Times Internet, for an Economic Times conversation about the future of artificial intelligence.
At one point, the discussion turned to a deceptively simple question: What happens when we can no longer trust what we see?
The audience was shown an AI-generated video of Altman speaking. It was recognisably artificial, with subtle imperfections in the lip movements. But the demonstration carried an obvious implication: those imperfections would not necessarily remain as the technology improved.
More Read
Altman acknowledged the concern. He said there was “a lot of fear” about the impact synthetic videos could have on elections, society and our ability to trust media.
His proposed response was not to abandon video as evidence. Instead, he anticipated a shift in how people establish authenticity. He argued that people would learn not to trust videos unless they could trust their provenance, and pointed to watermarking, detection tools and the possibility of cryptographically signing important communications.
Then he raised a second concern that deserves as much attention as the first: AI systems capable of personalised, interactive persuasion.
The threat, in his view, would not be limited to a fabricated video reaching millions of people. It could also involve an AI having a compelling, personalised conversation with one individual and influencing them in ways they might not recognise.
More Read
The question now is how much progress has been made on the safeguards—and whether they address the full problem Altman identified.
The first prediction: society would learn to question synthetic media
Altman’s first argument was that people would adapt. Just as audiences learned that photographs could be manipulated, they would learn not to accept videos at face value.
There is a useful distinction here. Learning to question media is a behavioural change. Establishing whether a particular video is authentic is a technical and institutional challenge. The first can encourage caution; the second requires evidence.
A viewer may become more sceptical of a video without becoming any better at determining whether it is genuine. Indiscriminate scepticism can create its own problem: authentic recordings may be dismissed as fake, while convincing fabrications continue to circulate.
The issue is therefore not simply whether people learn to question videos. It is whether they can access reliable evidence about a video’s origin, history and context.
India’s Election Commission addressed synthetic media as a practical governance issue during the 2024 general election. In its directions to political parties, the Commission said parties should not circulate deepfake audio or video that violates existing rules and required them to take down such material when it came to their notice, within a maximum of three hours.
The direction shows that the challenge is not merely theoretical. It also shows that the response requires more than individual media literacy: institutions, political organisations and digital platforms have roles to play.
Read the Election Commission’s directions to political parties (PDF).
The assessment: Society is responding to synthetic media, but that is not evidence that it has fully adapted. Awareness, verification and effective enforcement are different measures of progress.
Watermarks and digital credentials: the infrastructure is taking shape
Altman suggested that watermarking and detection systems could help distinguish authentic media from generated content. That idea has moved beyond the hypothetical.
Google DeepMind’s SynthID embeds digital watermarks into AI-generated content, including images, audio, text and video produced through supported Google systems. The watermark is designed to be imperceptible to people but detectable by the relevant technology. Google says it is designed to remain detectable through some common modifications, such as cropping, filters or compression, depending on the content type.
Google has also introduced a SynthID Detector to help users check whether content carries a SynthID watermark. This is a concrete step toward making some AI-origin signals more accessible to people outside AI research teams.
But a watermark is a signal, not a universal certificate of truth. SynthID is designed to identify content carrying its watermark. It does not follow that every AI-generated image, video or audio recording will carry that watermark, or that the absence of a detected watermark proves a recording is authentic.
Google itself has described watermarking as a building block rather than a complete solution to identifying AI-generated content. Coverage depends on whether the content was created or processed by a system that applies the watermark, and detection methods have their own limitations.
Content Credentials offer a second approach
Watermarking is not the only method being developed. The C2PA standard, which underpins Content Credentials, allows participating tools and devices to attach cryptographically verifiable information about a piece of digital media’s origin and editing history.
Think of it as a record of how a file came to exist: what created it, what tools were used and, where the information is available, how it was subsequently modified.
This is valuable because the question is not always whether an image was generated by AI. A journalist, investigator or viewer may also want to know whether a photograph came from a camera, whether it was edited, or whether its stated origin can be substantiated.
Content Credentials can help answer some of those questions. However, provenance information can be incomplete, and credentials may be absent from media that has not passed through a supporting workflow. Metadata can also be lost when files are processed or shared through systems that do not preserve it.
Most importantly, C2PA makes the limitation explicit: provenance information alone cannot establish whether the content is true, accurate or factual. It can provide evidence about origin and history, not a guarantee that the events depicted happened as presented.
Read the C2PA explainer on provenance and truth.
This distinction is central to understanding what the emerging infrastructure can—and cannot—do.
The trust gap: a verified origin does not guarantee a true story
Imagine receiving a video that appears to show a public official making an extraordinary announcement.
A verification system might establish that the video was generated using a particular AI tool. That would be useful information. Alternatively, a digital credential might show that a video originated from a particular device and record subsequent edits. That could also be useful.
Neither result, by itself, answers every question that matters:
- Does the video accurately represent the event being discussed?
- Is the speaker’s identity correctly represented?
- Has the footage been presented with misleading context?
- Does the accompanying claim accurately describe what happened?
- Is the source credible, and can its account be independently corroborated?
Even a genuine recording can be misleading when it is taken out of context. Conversely, AI-generated content can be clearly labelled and used legitimately—for satire, illustration or creative work.
Authenticity and truth are related, but they are not interchangeable.
The implication is that content provenance should be treated as one layer of verification, alongside source assessment, contextual reporting and independent corroboration.
For news organisations, a new technical capability does not eliminate the need for editorial judgement. It can make the process of establishing what can be trusted more structured, but not automatic. For audiences, a verification label should inform a decision rather than replace it.
The harder problem Altman raised: personalised AI persuasion
The most revealing part of Altman’s 2023 answer came after the discussion of fake videos.
He suggested that society might be paying insufficient attention to AI systems that could persuade individuals through personalised, interactive conversations.
This is a different problem from a deepfake video circulating online. A fabricated video is a piece of content that can be examined, shared and investigated. An interactive AI system can adapt its responses to an individual’s questions, reactions and apparent interests. Its influence may emerge through a sequence of exchanges rather than one easily identifiable piece of media.
The concern is not that every personalised AI interaction is manipulative. Personalisation can make education, customer support and information more useful.
The concern is that the same capacity to tailor communication could be used to influence people without making the source, purpose or persuasive intent sufficiently clear. A system might present misleading claims, exploit a person’s uncertainty or repeatedly reinforce a false premise.
In such cases, verifying whether an individual sentence or image was generated by AI would not necessarily address the wider interaction.
This is where the question of trust expands beyond media authentication. It becomes a question of who is communicating, what they are trying to achieve, what information is shaping the interaction, and whether the person on the other side can recognise and challenge that influence.
Altman identified the possibility in 2023. The interview itself does not establish how widespread such effects would become, and the existence of increasingly capable conversational systems is not proof that they inevitably produce harmful persuasion.
But the concern points to a limitation of an approach centred exclusively on detecting synthetic content. A watermark can help identify a source of generated media. It cannot, on its own, tell us whether a conversation is manipulative.
What this means for India
India’s scale, linguistic diversity and reliance on digital communication make trustworthy media consequential. The challenge is not limited to deepfakes involving prominent political figures. False or manipulated content can also affect businesses, consumers, public services and individuals.
For businesses, the risks include impersonation and fraudulent communications. For newsrooms, they include publishing material whose origin or context has not been adequately established. For the public, they include both believing fabricated content and losing confidence in authentic evidence.
The practical response needs to operate at several levels:
- For platforms and technology providers: Build mechanisms that make the origin and editing history of supported content easier to inspect. Explain their limitations, and avoid presenting probabilistic detection as definitive proof.
- For news organisations: Establish procedures for verifying unusual or consequential audio, images and video. Use provenance tools where available, but combine them with independent sourcing and contextual checks.
- For institutions and businesses: Create processes for verifying high-risk communications, especially those involving money, sensitive information, public announcements or instructions attributed to senior leaders.
- For audiences: Treat extraordinary media claims as prompts for verification. Look for the original source, check independent reporting and distinguish a verified file history from evidence that the claim itself is true.
These measures will not remove every risk. They can, however, make trust less dependent on visual plausibility alone.
The verdict: Altman identified the right direction, but the problem is larger than detection
Three years after the 2023 conversation, the response to synthetic media is becoming more technically sophisticated.
Watermarking systems are being deployed across different types of AI-generated content. Content Credentials provide a standardised way to record and verify aspects of digital media’s origin and history. Verification tools are becoming more accessible, and public institutions have had to address synthetic media in real-world settings.
These are meaningful developments. But they do not establish that society has solved the authenticity problem—or that audiences can reliably distinguish genuine media from fabricated material in every context.
Altman’s expectation that people would adapt remains plausible, but adaptation is not the same as resolution. The emerging tools can help establish where supported content came from and how it was modified. They cannot independently establish the truth of every claim, guarantee that every file carries verifiable credentials or prevent all forms of manipulation.
And his warning about personalised persuasion points to a problem that content authentication alone cannot solve.
The next phase of trustworthy AI will require more than better detectors. It will depend on provenance standards, transparent systems, responsible platforms, stronger verification practices and people who understand what technical signals actually mean.
The deeper lesson is simple: the future of trust will not depend on whether synthetic media looks real. It will depend on whether we can establish why it should be believed.
AI, Revisited: The series
AI, Revisited returns to predictions and warnings made during the early public adoption of generative AI, then tests them against evidence and developments that followed. The aim is not to declare every prediction right or wrong. It is to separate what has materialised, what remains uncertain and what the original conversation may have missed. Explore more of our AI coverage on FutureIsNow.
Previously: Episode 01 examined Sam Altman’s prediction that AI would change jobs before replacing entire occupations.
Next: Can AI genuinely serve India’s linguistic diversity—or will the benefits remain uneven?
